[{"data":1,"prerenderedAt":2603},["ShallowReactive",2],{"site-chrome-en":3,"site-chrome-th":-1,"blog-posts-en":287,"article-blog-en-\u002Fblog\u002Fdecision-intelligence-playbook-for-e-commerce":654,"blog-page-en":2495,"blog-categories-en":2547,"og-copy-en":2560,"og-posts-en-none":2602},{"id":4,"extension":5,"footer":6,"header":116,"indexFacets":202,"meta":243,"notFound":244,"stem":250,"ui":251,"__hash__":286},"chrome_en\u002Fen\u002Fchrome.yml","yml",{"columns":7,"legal":83,"quote":84,"askAi":87,"socials":103},[8,29,47,65],{"label":9,"children":10},"Product",[11,14,17,20,23,26],{"label":12,"to":13},"Ads optimization","\u002Fblog\u002Fads-optimization",{"label":15,"to":16},"Unit economics","\u002Fblog\u002Fprofit-margin",{"label":18,"to":19},"Agentic KAM","https:\u002F\u002Flearn.dataglasslabs.com\u002Fdocs\u002Fai-and-api\u002Fassistants",{"label":21,"to":22},"One-click deploy","https:\u002F\u002Flearn.dataglasslabs.com\u002Fdocs\u002Fgrowth\u002Factions\u002Faction-flow",{"label":24,"to":25},"Integrations","https:\u002F\u002Flearn.dataglasslabs.com\u002Fdocs\u002Foverview\u002Fget-started\u002Fsupported-platforms",{"label":27,"to":28},"Pricing","\u002Fpricing",{"label":30,"children":31},"Resources",[32,35,38,41,44],{"label":33,"to":34},"Blog","\u002Fblog",{"label":36,"to":37},"Customer stories","\u002Fcustomers",{"label":39,"to":40},"Research","\u002Fresearch",{"label":42,"to":43},"Glossary","\u002Fglossary",{"label":45,"to":46},"Docs","https:\u002F\u002Flearn.dataglasslabs.com\u002Fdocs",{"label":48,"children":49},"Company",[50,53,56,59,62],{"label":51,"to":52},"About","\u002Fcompany",{"label":54,"to":55},"Contact","\u002Fcontact",{"label":57,"to":58},"Book a demo","\u002Fbook-a-demo",{"label":60,"to":61},"Privacy","\u002Flegal\u002Fprivacy",{"label":63,"to":64},"Terms","\u002Flegal\u002Fterms",{"label":66,"children":67},"DataGlass for",[68,71,74,77,80],{"label":69,"to":70},"Agencies","\u002Ffor\u002Fagencies",{"label":72,"to":73},"Brands","\u002Ffor\u002Fbrands",{"label":75,"to":76},"Planners and optimizers","\u002Ffor\u002Fplanners",{"label":78,"to":79},"Managers and analysts","\u002Ffor\u002Fmanagers",{"label":81,"to":82},"Sellers and shop owners","\u002Ffor\u002Fsellers","© 2026 DataGlass Labs. All rights reserved.",{"text":85,"caption":86},"Measure what you keep, and then grow what you measure.","The DataGlass team",{"label":88,"prompt":89,"assistants":90},"Ask AI about DataGlass","I want to understand DataGlass, a decision engine for Shopee, Lazada and TikTok Shop sellers that optimizes ads, pricing and promotions for true profit. Use web search to read their website at https:\u002F\u002Fwww.dataglasslabs.com (with an agent-readable summary at https:\u002F\u002Fwww.dataglasslabs.com\u002Fllms.txt), then explain what DataGlass does, who it is for and how it differs from an analytics dashboard.",[91,95,99],{"label":92,"icon":93,"url":94},"ChatGPT","chatgpt","https:\u002F\u002Fchatgpt.com\u002F?prompt={prompt}",{"label":96,"icon":97,"url":98},"Claude","claude","https:\u002F\u002Fclaude.ai\u002Fnew?q={prompt}",{"label":100,"icon":101,"url":102},"Perplexity","perplexity","https:\u002F\u002Fwww.perplexity.ai\u002Fsearch?q={prompt}",[104,108,112],{"label":105,"icon":106,"to":107},"LinkedIn","i-lucide-linkedin","https:\u002F\u002Fwww.linkedin.com\u002Fcompany\u002Fdataglass-labs",{"label":109,"icon":110,"to":111},"Facebook","i-lucide-facebook","https:\u002F\u002Fwww.facebook.com\u002Fdataglasslabsth",{"label":113,"icon":114,"to":115},"Instagram","i-lucide-instagram","https:\u002F\u002Fwww.instagram.com\u002Fdataglasslabs\u002F",{"links":117,"cta":196,"secondaryCta":199},[118,152,191,193,194,195],{"label":9,"to":119,"menu":120},"\u002F",{"layout":121,"heading":122,"cards":123,"groups":135,"footer":142},"cards","What DataGlass does",[124,128,132],{"label":125,"description":126,"to":13,"ui":127},"True profit ads optimization","Budgets, pacing and Target ROAS tuned for profit.","budget-tomorrow",{"label":129,"description":130,"to":16,"ui":131},"Unit economics breakdown","What you actually earn from every sale.","unit-economics",{"label":18,"description":133,"to":19,"motif":134},"An AI agent that knows your shops.",true,[136],{"label":137,"size":138,"links":139},"More","secondary",[140,141],{"label":21,"to":22},{"label":24,"to":25},[143,148],{"lead":144,"label":145,"icon":146,"to":147},"DataGlass v.2.6.","See what's new","i-lucide-box","https:\u002F\u002Flearn.dataglasslabs.com\u002Freleases",{"lead":149,"label":150,"icon":151,"to":37},"Read our","customer stories","i-lucide-book-open",{"label":30,"to":34,"menu":153},{"layout":154,"groups":155},"directory",[156,164,170,182],{"label":66,"size":157,"links":158},"primary",[159,160,161,162,163],{"label":69,"to":70},{"label":72,"to":73},{"label":75,"to":76},{"label":78,"to":79},{"label":81,"to":82},{"label":165,"size":157,"links":166},"Insights",[167,168,169],{"label":33,"to":34},{"label":36,"to":37},{"label":39,"to":40},{"label":171,"size":157,"links":172},"Learn",[173,174,177,180],{"label":45,"to":46},{"label":175,"to":176},"Use cases","https:\u002F\u002Flearn.dataglasslabs.com\u002Fuse-cases",{"label":178,"to":179},"Training","https:\u002F\u002Flearn.dataglasslabs.com\u002Ftraining",{"label":181,"to":147},"Changelog",{"label":183,"size":138,"links":184},"Support",[185,186,189],{"label":42,"to":43},{"label":187,"to":188},"Brand kit","\u002Fbrand",{"label":190,"to":55},"Contact us",{"label":192,"to":37},"Customers",{"label":27,"to":28},{"label":48,"to":52},{"label":54,"to":55},{"label":197,"to":198},"Start for free","https:\u002F\u002Fapp.dataglasslabs.com\u002Fsign-up",{"label":200,"to":201},"Sign in","https:\u002F\u002Fapp.dataglasslabs.com\u002Fsign-in",{"clear":203,"groups":204},"Clear filters",[205,218],{"key":206,"label":207,"options":208},"platforms","Platform",[209,212,215],{"value":210,"label":211},"shopee","Shopee",{"value":213,"label":214},"lazada","Lazada",{"value":216,"label":217},"tiktok","TikTok Shop",{"key":219,"label":220,"options":221},"format","Format",[222,225,228,231,234,237,240],{"value":223,"label":224},"guide","Guide",{"value":226,"label":227},"article","Article",{"value":229,"label":230},"comparison","Comparison",{"value":232,"label":233},"calculator","Calculator",{"value":235,"label":236},"paper","Paper",{"value":238,"label":239},"survey","Survey",{"value":241,"label":242},"dataset","Dataset",{},{"title":245,"description":246,"cta":247,"errorTitle":249},"Page not found","The page may have moved, or it never existed.",{"label":248,"to":119},"Back to home","Something went wrong","en\u002Fchrome",{"previous":252,"next":253,"pause":254,"play":255,"goTo":256,"expand":257,"close":258,"readMore":259,"learnMore":260,"seeAll":261,"menu":262,"back":263,"share":264,"shareOn":265,"copyLink":266,"copied":267,"onThisPage":268,"minRead":269,"sources":270,"viewMore":271,"searchPosts":272,"allPosts":273,"all":274,"filters":275,"search":276,"searchClose":277,"searchPlaceholder":278,"searchSuggested":279,"searchResults":280,"searchNoResults":281,"searchLoading":282,"relatedTitle":283,"mainNav":284,"footerNav":285},"Previous","Next","Pause","Play","Go to item","Expand","Close","Read more","Learn more","See all","Menu","Back","Share","Share on","Copy link","Link copied","On this page","{n} min read","Sources","View more","Search posts","All posts","All","Filters","Search","Close search","Search DataGlass","Suggested","Results","No results for","Loading","More to discover","Main","Footer","HZQSEulcpFGY0nWnUTNYhG2n06nO97XnUBc_pEVOFxo",[288,298,304,310,315,320,324,330,335,341,347,353,358,363,368,374,380,386,392,399,405,411,417,422,427,433,438,443,449,455,460,466,472,478,483,489,495,501,506,510,515,520,526,533,539,545,551,557,563,569,575,581,587,593,599,605,611,617,622,627,632,637,642,648],{"title":289,"description":290,"category":291,"date":292,"readMinutes":293,"featured":294,"cover":295,"format":223,"platforms":296,"draft":294,"slug":297},"E-commerce ads optimization for Shopee, Lazada, and TikTok Shop, on profit","E-commerce ads optimization on profit: True ROAS, margin-based Target ROAS and nightly budget allocation for Shopee, Lazada and TikTok Shop ads, applied for you.","ads","2026-10-23",4,false,null,[210,213,216],"ads-optimization",{"title":299,"description":300,"category":291,"date":292,"readMinutes":301,"featured":294,"cover":295,"format":232,"platforms":302,"draft":294,"slug":303},"What ROAS is break-even? Calculate break-even ROAS and your maximum ad spend per order","Break-even ROAS is 1 ÷ your contribution-margin ratio. It does not depend on how much you spend on ads. This calculator derives it from a dated Shopee Thailand fee rate — or one you type in — then tells you the number sellers actually need: the most baht one order can pay for ads before it stops making money.",9,[210],"break-even-roas-calculator",{"title":305,"description":306,"category":307,"date":292,"readMinutes":308,"featured":294,"cover":295,"format":229,"platforms":295,"draft":294,"slug":309},"DataGlass vs a marketing agency for marketplace profit","When an agency is the right hire, where the agency model has structural blind spots for marketplace profit, and how DataGlass complements an agency engagement.","company",3,"dataglass-vs-agency",{"title":311,"description":312,"category":307,"date":292,"readMinutes":313,"featured":294,"cover":295,"format":229,"platforms":295,"draft":294,"slug":314},"DataGlass vs an ERP for marketplace decision-making","When an ERP is the right tool, where it falls short for marketplace ad\u002Fpricing decisions, and how DataGlass complements (not replaces) the ERP for SEA sellers.",2,"dataglass-vs-erp",{"title":316,"description":317,"category":307,"date":292,"readMinutes":313,"featured":294,"cover":295,"format":229,"platforms":318,"draft":294,"slug":319},"DataGlass vs Shopee Seller Center for profit-aware operations","Shopee Seller Center is essential for running a Shopee shop, but it does not show profit-aware metrics or cross-platform views. How DataGlass complements it.",[210],"dataglass-vs-shopee-seller-center",{"title":321,"description":322,"category":307,"date":292,"readMinutes":313,"featured":294,"cover":295,"format":229,"platforms":295,"draft":294,"slug":323},"DataGlass vs spreadsheets for marketplace sellers","When spreadsheets are enough, where they break for marketplace sellers, and how DataGlass connects ads, COGS, fees, and inventory into one decision-ready view.","dataglass-vs-spreadsheet",{"title":325,"description":326,"category":327,"date":292,"readMinutes":313,"featured":294,"cover":295,"format":223,"platforms":328,"draft":294,"slug":329},"Prevent stockouts and overstock across Shopee, Lazada, and TikTok Shop","Prevent Shopee, Lazada, and TikTok Shop stockouts with demand-aware forecasting, campaign-day planning, and cross-platform inventory unification.","operations",[210,213,216],"inventory-stockout",{"title":331,"description":332,"category":327,"date":292,"readMinutes":313,"featured":294,"cover":295,"format":226,"platforms":333,"draft":294,"slug":334},"Lazada integration for marketplace profit optimization","DataGlass integrates with Lazada to model true profit per SKU, detect ad waste, evaluate voucher impact, and forecast stockout risk.",[213],"lazada-integration",{"title":336,"description":337,"category":338,"date":292,"readMinutes":301,"featured":294,"cover":295,"format":232,"platforms":339,"draft":294,"slug":340},"Shopee raised its fees on 4 August 2026 — how much more is your shop paying?","The new rate table is everywhere. Your number is not. Enter last month’s GMV by category and this page computes the baht difference between the schedule effective 1 June 2026 and the one effective 4 August 2026 — including the extra hit for shops that were on the discounted-commission programme.","profit",[210],"marketplace-fee-impact-calculator",{"title":342,"description":343,"category":338,"date":292,"readMinutes":344,"featured":294,"cover":295,"format":223,"platforms":345,"draft":294,"slug":346},"Optimize e-commerce profit across Shopee, Lazada, and TikTok Shop","How to optimize e-commerce profit across Shopee, Lazada and TikTok Shop: one objective, five decision variables, and the constraint that binds each.",7,[210,213,216],"optimize-ecommerce-profit",{"title":348,"description":349,"category":350,"date":292,"readMinutes":313,"featured":294,"cover":295,"format":223,"platforms":351,"draft":294,"slug":352},"Price and promote without destroying margin","Optimize Shopee, Lazada, and TikTok Shop discounts, bundle deals, and campaign-day pricing on true margin. Project promotion impact before publishing.","pricing",[210,213,216],"pricing-promotions",{"title":354,"description":355,"category":338,"date":292,"readMinutes":293,"featured":294,"cover":295,"format":232,"platforms":356,"draft":294,"slug":357},"E-commerce profit calculator for Shopee sellers","Calculate true Shopee profit per SKU with a per-order breakdown covering commission fees, ads, COGS, and fulfillment.",[210],"profit-margin",{"title":359,"description":360,"category":327,"date":292,"readMinutes":313,"featured":294,"cover":295,"format":226,"platforms":361,"draft":294,"slug":362},"Shopee integration for profit, ads, stock, and promotion decisions","DataGlass connects to Shopee to model true profit per SKU, detect wasted ad spend, and surface daily decisions. Live integration, no manual exports.",[210],"shopee-integration",{"title":364,"description":365,"category":327,"date":292,"readMinutes":313,"featured":294,"cover":295,"format":226,"platforms":366,"draft":294,"slug":367},"TikTok Shop integration for ads, inventory, and profit decisions","DataGlass integrates with TikTok Shop for True ROAS on video and live ads, margin-aware discount analysis, and stock prioritization before live events.",[216],"tiktok-shop-integration",{"title":369,"description":370,"category":291,"date":371,"readMinutes":344,"featured":294,"cover":295,"format":223,"platforms":372,"draft":294,"slug":373},"Shopee ROAS protection and TikTok ROI protection: when the credit actually protects profit","Shopee ROAS protection credits campaigns back to 80% of target, TikTok to 90%. Why that protects profit only above 1.25× break-even, with a worked example.","2026-10-08",[210,216],"roas-protection",{"title":375,"description":376,"category":291,"date":371,"readMinutes":377,"featured":294,"cover":295,"format":223,"platforms":378,"draft":294,"slug":379},"Shopee Ads optimization: set Target ROAS, budget and timing from profit","How to optimize Shopee Ads for profit: break-even ROAS per SKU, which lever binds, and CPC by hour from 9.11M Thai clicks. Worked THB example.",10,[210],"shopee-ads-optimization",{"title":381,"description":382,"category":338,"date":371,"readMinutes":383,"featured":294,"cover":295,"format":223,"platforms":384,"draft":294,"slug":385},"Shopee seller fees in the Philippines 2026: every fee, what it is charged on, and your break-even ROAS","Shopee Philippines seller fees 2026: commission, 2.24% transaction fee, 5.6% platform shipping, seller growth support and ₱5 order fee, worked on one order.",5,[210],"shopee-seller-fees",{"title":387,"description":388,"category":291,"date":371,"readMinutes":389,"featured":294,"cover":295,"format":223,"platforms":390,"draft":294,"slug":391},"TikTok Shop ads optimization: set GMV Max ROI from margin and affiliate mix","How to optimize TikTok Shop GMV Max for profit: break-even ROI after TikTok fees, the affiliate commission trap, and the 3% commission rule. Worked THB example.",8,[216],"tiktok-shop-ads-optimization",{"title":393,"description":394,"category":291,"date":395,"readMinutes":396,"featured":294,"cover":295,"format":226,"platforms":397,"draft":294,"slug":398},"How Shopee ads bidding works: what your Target ROAS and budget actually do to a bid","How Shopee ads bidding works, derived: GMV Max gives you two boxes — a Target ROAS and a daily budget — and every guide treats them as two levers. They are not. Both collapse into a single threshold on value-to-cost before any bid is placed, only one is binding at a time, and the other is doing nothing at all. Here is the derivation, the three regimes a campaign can be in, and the one-line test that tells you which of your two levers is switched off.","2026-09-01",13,[210],"how-shopee-ads-bidding-works",{"title":400,"description":401,"category":291,"date":395,"readMinutes":402,"featured":294,"cover":295,"format":223,"platforms":403,"draft":294,"slug":404},"Break-even ROAS is the zero line: the Target ROAS that actually maximizes Shopee profit","Setting your Target ROAS to your break-even ROAS does not protect profit — it targets exactly zero, for every possible volume. The target that maximizes profit is break-even marked up three times: by the demand your ads did not create, by the price your own expansion pays, and by the gap between your average click and your last one. On mid-market parameters those compound to 3.3×, moving the correct target on a 20%-margin product from 5.00 to 16.67.",12,[210],"profit-maximizing-target-roas",{"title":406,"description":407,"category":408,"date":395,"readMinutes":396,"featured":294,"cover":295,"format":241,"platforms":409,"draft":294,"slug":410},"Shopee ad benchmarks 2026: what a click and a thousand impressions actually cost","Across 333.9 million Shopee impressions and 9.11 million clicks on 80+ Thai shops, the price of a click varies 2.01× by hour of day while the rate at which people click varies 1.14×. CPC ranges THB 3.36 to THB 6.75, CPM ranges THB 89.16 to THB 200.15, and every behavioural input is nearly flat. The full table, the methodology, and the two ways a benchmark like this gets misused.","data-science",[210],"shopee-ads-benchmarks-cpc-cpm",{"title":412,"description":413,"category":307,"date":414,"readMinutes":415,"featured":294,"cover":295,"format":229,"platforms":295,"draft":294,"slug":416},"DataGlass is not an OMS — what the difference is, and when to run both","An OMS makes the order happen. DataGlass decides what to do next. The daily-question boundary, an explicit list of what DataGlass does not do, and the one seam: price.","2026-08-25",6,"dataglass-vs-oms",{"title":418,"description":419,"category":291,"date":414,"readMinutes":344,"featured":294,"cover":295,"format":223,"platforms":420,"draft":294,"slug":421},"How much should you spend on Shopee Ads? Derive it from margin, not from feel","Break-even ROAS is 1 ÷ contribution margin ratio. Worked THB example: a fashion SKU allows THB 71 per attributed order and roughly THB 994 a day. Budget is an output.",[210],"shopee-ads-budget",{"title":423,"description":424,"category":338,"date":414,"readMinutes":344,"featured":294,"cover":295,"format":223,"platforms":425,"draft":294,"slug":426},"Losing money on Shopee — the leaks a fee calculator cannot see","Shopee deducts 7.49–17.12% plus a 3.21% payment fee — the half a calculator sees. Six month-level leaks turn a 28.6% \"profit\" into 1.6%. Sourced THB waterfall.",[210],"shopee-losing-money",{"title":428,"description":429,"category":430,"date":414,"readMinutes":344,"featured":294,"cover":295,"format":229,"platforms":431,"draft":294,"slug":432},"Shopee vs TikTok Shop vs Lazada — what survives after fees and the ads it takes to sell","TikTok Shop's 11.77% fashion commission becomes a 23.2% take rate once the Commerce Growth and Transaction Fees stack. The crossover with Shopee sits near THB 3,720.","market",[210,213,216],"where-to-sell-shopee-tiktok-lazada",{"title":434,"description":435,"category":291,"date":436,"readMinutes":301,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":437},"CTR and Target ROAS under CPM vs CPC: what changes when marketplaces bill for impressions","Under cost-per-click billing, a better creative is worth nothing to your unit economics — it buys the same clicks from fewer impressions, and your cost per click does not move. Under cost-per-impression billing the same improvement cuts the price of every click by g\u002F(1+g). Five lemmas on the transition, and why click-through rate is the input most marketplace sellers are least prepared to have priced.","2026-08-18","ctr-target-roas-cpm-vs-cpc",{"title":439,"description":440,"category":291,"date":441,"readMinutes":377,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":442},"What is ROAS — and why a high ROAS doesn't always mean more sales or more profit","ROAS (Return on Ad Spend) is sales ÷ ad spend: spend THB 1, get THB 20 of sales back, and your ROAS is 20. But a high ROAS only means a high return per ad baht — not the highest total sales, and not real profit once you subtract COGS, platform fees, payment fees, vouchers, shipping, and returns. Here is what ROAS is, what it hides, and the number sellers should actually watch.","2026-06-01","what-is-roas",{"title":444,"description":445,"category":291,"date":446,"readMinutes":389,"featured":294,"cover":295,"format":223,"platforms":447,"draft":294,"slug":448},"What is Target ROAS? How to set it in Shopee GMV Max without burning profit","What is Target ROAS? It is the revenue-to-ad-spend ratio Shopee GMV Max optimises toward — and on the Shopee fee schedule effective 4 August 2026, a THB 1,000 product with THB 600 of COGS is left with 7.7% margin before ads, so a Target ROAS of 5 recovers THB 0.39 of contribution per THB 1.00 of ad spend while the campaign reports that it hit its target. What Target ROAS means, how to set it in GMV Max, and the two formulas — break-even ROAS and target-margin ROAS — that protect your margin.","2026-05-30",[210],"how-to-set-target-roas-gmv-max",{"title":450,"description":451,"category":291,"date":452,"readMinutes":415,"featured":294,"cover":295,"format":226,"platforms":453,"draft":294,"slug":454},"Platform ROAS vs true ROAS: why your ads make sales but no profit","A ROAS of 5 feels like a win — until you notice the platform computed it on gross attributed sales, before fees, vouchers, shipping, and COGS, and while still counting orders that were later returned or cancelled. The dashboard number is a revenue ratio wearing a profit costume. A primer on what platform ROAS hides and how true ROAS is calculated across Shopee, Lazada, and TikTok Shop.","2026-05-29",[210,213,216],"platform-roas-vs-true-roas",{"title":456,"description":457,"category":408,"date":452,"readMinutes":415,"featured":294,"cover":295,"format":226,"platforms":458,"draft":294,"slug":459},"What is e-commerce optimization?","Optimization is not a marketing word — it is an operations-research discipline: use data and math to compute the best decision against an explicit objective. Point it at a Shopee, Lazada, or TikTok Shop store and the objective is true profit, not sales. A primer on why those two numbers stopped moving together, and what optimization actually answers.",[210,213,216],"what-is-ecommerce-optimization",{"title":461,"description":462,"category":338,"date":463,"readMinutes":301,"featured":294,"cover":295,"format":223,"platforms":464,"draft":294,"slug":465},"How to cut Shopee seller costs: the supply-side margin audit","The median gap between COGS refreshes on a top-50 Shopee catalogue is 87 days, and over that window supplier and packaging drift alone compresses contribution margin by 1.5–3.0 percentage points. The eight supply-side cost lines that compound against Shopee margin between audits, ranked by recoverable magnitude, with the review cadence that holds each one.","2026-05-08",[210],"how-to-increase-shopee-profit-without-more-sales",{"title":467,"description":468,"category":291,"date":463,"readMinutes":469,"featured":294,"cover":295,"format":223,"platforms":470,"draft":294,"slug":471},"How to sell more on Shopee without burning margin","The standard Shopee growth advice — discount harder, scale ad budget, opt into every campaign — works against the operator at scale. A research note on the lever order that actually compounds Shopee revenue, and the metric that tells you whether the lift was earned or bought.",11,[210],"how-to-sell-more-on-shopee",{"title":473,"description":474,"category":338,"date":475,"readMinutes":415,"featured":294,"cover":295,"format":223,"platforms":476,"draft":294,"slug":477},"Increase profit on Shopee without selling more","Shopee ROAS hides commission, fees, vouchers and shipping. Optimize Shopee profit with the per-SKU break-even framework that replaces it — 280 Thai accounts.","2026-05-07",[210],"increase-shopee-profit",{"title":479,"description":480,"category":430,"date":475,"readMinutes":415,"featured":294,"cover":295,"format":223,"platforms":481,"draft":294,"slug":482},"Increase sales on Shopee in 2026","Shopee CPC is climbing — buying demand got pricier while repeat demand stayed flat. The three levers that compound revenue, plus a 90-day operator cadence.",[210],"increase-shopee-sales",{"title":484,"description":485,"category":408,"date":475,"readMinutes":486,"featured":294,"cover":295,"format":235,"platforms":487,"draft":294,"slug":488},"Mechanism-Aware Hierarchical Causal Elasticity Modeling for Platform E-Commerce","A methodological working paper proposing MA-HCEM: a hierarchical Bayesian, double-machine-learning, mechanism-aware estimator for SKU-level price elasticity on Shopee, Lazada, and TikTok Shop, with Bundle Deal and voucher mechanisms treated as structured price discrimination.",27,[210,213,216],"mechanism-aware-hierarchical-causal-elasticity-modeling-platform-ecommerce",{"title":490,"description":491,"category":350,"date":475,"readMinutes":492,"featured":294,"cover":295,"format":226,"platforms":493,"draft":294,"slug":494},"Price elasticity modeling for marketplace sellers — why Shopee, Lazada, and TikTok Shop pricing decisions need it","A practical guide from DataGlass Labs Research on the most under-used lever in Southeast Asian e-commerce: knowing how customers actually respond to price changes.",14,[210,213,216],"price-elasticity-modeling-shopee-lazada-tiktok-shop",{"title":496,"description":497,"category":408,"date":498,"readMinutes":499,"featured":294,"cover":295,"format":235,"platforms":295,"draft":294,"slug":500},"From Gut Feel to Posterior Inference: A Research Article on the DataGlass Decision-Intelligence System for E-Commerce Ad Budget Allocation","A rigorous public communication of the DataGlass decision-intelligence system for daily ad-budget allocation on platform-controlled marketplaces — the pre-algorithmic baseline, the analytical reasons it fails (the reported-vs-true ROAS gap and the rolling-mean bias), the Bayesian + bandits-with-knapsacks methodology, and the empirical 18–24% portfolio-profit lift with reduced reallocation frequency.","2026-05-04",20,"bayesian-budget-allocation-for-marketplace-ads",{"title":502,"description":503,"category":408,"date":498,"readMinutes":504,"featured":294,"cover":295,"format":238,"platforms":295,"draft":294,"slug":505},"Decision Intelligence for E-commerce: How Retailers Optimise Pricing, Demand Forecasting, Inventory, Promotions & Personalization","Pricing, forecasting, inventory, promotions, and personalization — a deep technical survey of the techniques large retailers use, the variants that matter, and how to deploy them",42,"decision-intelligence-playbook-for-e-commerce",{"title":507,"description":508,"category":408,"date":498,"readMinutes":308,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":509},"DataGlass Research Methodology","A single canonical reference for the sample frame, time window, attribution rules, and exclusions that calibrate every \"in our data\" claim across the blog and research surfaces.","methodology",{"title":511,"description":512,"category":408,"date":498,"readMinutes":513,"featured":294,"cover":295,"format":238,"platforms":295,"draft":294,"slug":514},"Prediction and Risk Optimization Under Uncertainty: A Cross-Domain Meta-Review","Mature decision systems in finance, operations, insurance, energy, healthcare, causal inference, and e-commerce share the same four primitives — a calibrated probabilistic model, a coherent risk-aware objective, an operational constraint set, and a principled exploration mechanism. Eleven worked cases ground the framework, with the DataGlass marketplace ad-budget system as the connecting tissue.",49,"prediction-and-risk-optimization-meta-review",{"title":516,"description":517,"category":327,"date":498,"readMinutes":492,"featured":294,"cover":295,"format":226,"platforms":518,"draft":294,"slug":519},"Shopee optimization in 2026 is a feedback loop, not a checklist","The popular \"10 ways to optimize your Shopee shop\" advice optimises the wrong objective. The lever that actually moves margin is the tightness of the diagnose → decide → operate loop — and most checklists skip the diagnose stage entirely.",[210],"shopee-optimization-in-2026",{"title":521,"description":522,"category":430,"date":498,"readMinutes":523,"featured":134,"cover":295,"format":241,"platforms":524,"draft":294,"slug":525},"Shopee Sellers in 2026: Southeast Asia E-commerce Market Research, GMV & Seller Economics","Market growth has resumed; seller economics have become more exacting. Southeast Asia's platform e-commerce reached US$157.6B in 2025 (up 22.8% YoY), top-three platforms now control about 98.8% of platform GMV, and content commerce accounts for ~32% of platform GMV. The 2026 question for sellers is no longer \"how big is the market?\" — it is \"who controls the decision loop?\"",22,[210],"southeast-asia-seller-market-research-2026",{"title":527,"description":528,"category":327,"date":529,"readMinutes":530,"featured":294,"cover":295,"format":223,"platforms":531,"draft":294,"slug":532},"Shopee fee structure and commission calculation in 2026: the six layers, the four-term formula","Most 'Shopee fees' explainers stop at commission and transaction fee. The four layers underneath — voucher co-funding, Free Shipping Program subsidy, Mall premium, payment-channel pass-through — are where the gap between dashboard revenue and bank deposit actually lives. With a sourced layer-by-layer breakdown, the four-term commission-per-order formula, two worked THB examples, a sensitivity table, and the reconciliation procedure.","2026-05-03",18,[210],"shopee-fee-structure-2026",{"title":534,"description":535,"category":291,"date":536,"readMinutes":469,"featured":294,"cover":295,"format":223,"platforms":537,"draft":294,"slug":538},"Cross-platform ad budget allocation for SEA marketplace sellers","Most multi-platform sellers split ad budget across Shopee, Lazada, and TikTok Shop by historical revenue share. The math says that's wrong. Optimal allocation equalises marginal ROAS, not historical share — and the gap between the two on a typical account is 4–7 percentage points of net contribution margin per quarter.","2026-04-29",[210,213,216],"cross-platform-ad-budget-allocation",{"title":540,"description":541,"category":291,"date":542,"readMinutes":344,"featured":294,"cover":295,"format":223,"platforms":543,"draft":294,"slug":544},"Why scaling Shopee ad budget grows revenue and shrinks profit","A Shopee account moving its monthly ad budget from THB 40,000 to THB 100,000 commonly grows revenue 60% and loses a third of its net profit — and not one decision in that sequence looks wrong on the dashboard. The arithmetic of the scaling trap, why platform-reported ROAS is blind to it by construction, and the per-SKU break-even bar that stops it.","2026-04-22",[210],"how-to-increase-profit-on-shopee",{"title":546,"description":547,"category":338,"date":548,"readMinutes":396,"featured":294,"cover":295,"format":223,"platforms":549,"draft":294,"slug":550},"How to increase profit on Lazada in 2026","The LazMall badge lifts conversion. It also raises commission, mandates free-shipping subsidies, and pulls the price ceiling down through the platform's own competitive-parity rules. Whether the badge pays is a per-SKU question. A research note on the LazMall economics, why Sponsored Discovery leaks more margin than Sponsored Search, and the audit that recovers 4–6 percentage points of margin in 30 days.","2026-04-15",[213],"how-to-increase-profit-on-lazada",{"title":552,"description":553,"category":291,"date":554,"readMinutes":396,"featured":294,"cover":295,"format":223,"platforms":555,"draft":294,"slug":556},"How to reduce Shopee ad waste without killing sales","On a typical Shopee account, 20–30% of ad spend runs at a structural loss the platform dashboard ranks as winning campaigns. Pausing \"underperformers\" misses the leak. A research note on the two structural defaults that cause hidden ad waste — and the audit that surfaces it without losing revenue.","2026-04-08",[210],"how-to-reduce-shopee-ad-waste",{"title":558,"description":559,"category":338,"date":560,"readMinutes":469,"featured":294,"cover":295,"format":223,"platforms":561,"draft":294,"slug":562},"How to calculate Shopee seller margin","The math, the inputs, and the program-specific traps. Why Shopee's in-dashboard income view almost always overstates margin — and how to reconstruct true contribution margin per SKU using the order-line data Shopee already exposes via Open Platform.","2026-04-05",[210],"how-to-calculate-shopee-seller-margin",{"title":564,"description":565,"category":338,"date":566,"readMinutes":469,"featured":294,"cover":295,"format":223,"platforms":567,"draft":294,"slug":568},"How to calculate Lazada seller margin","The math, the inputs, and the program-specific traps. Why Lazada's in-platform P&L summary almost always overstates margin — and how to reconstruct true contribution margin per SKU using the order-line data Lazada exposes via the Open Platform.","2026-04-01",[213],"how-to-calculate-lazada-seller-margin",{"title":570,"description":571,"category":291,"date":572,"readMinutes":492,"featured":294,"cover":295,"format":223,"platforms":573,"draft":294,"slug":574},"How to calculate true Shopee ROAS for profit","A methodology note. Shopee's in-platform ROAS is gross-revenue based and structurally biased toward overspend at scale. True ROAS is the same formula with one input substituted — and that substitution flips winners into losses on roughly half the typical Shopee catalog. With charts, three SKU profiles, sensitivity analysis, and the operating procedure that applies the substitution at production cadence.","2026-03-25",[210],"how-to-calculate-true-shopee-roas",{"title":576,"description":577,"category":291,"date":578,"readMinutes":469,"featured":294,"cover":295,"format":223,"platforms":579,"draft":294,"slug":580},"How to reduce Lazada ad waste","Lazada runs two structurally different ad products — Sponsored Search and Sponsored Discovery — with break-even economics that diverge by ~1.4–1.8× in true-ROAS terms. A single account-wide ROAS target misjudges both. A research note on the audit that surfaces the structural waste, the pruning order that recovers margin without losing revenue, and the per-placement framework that survives Pay Day and 11.11.","2026-03-18",[213],"how-to-reduce-lazada-ad-waste",{"title":582,"description":583,"category":338,"date":584,"readMinutes":396,"featured":294,"cover":295,"format":223,"platforms":585,"draft":294,"slug":586},"How to find low-margin SKUs on Shopee","On a typical Shopee account, the top-10 SKUs by revenue and the top-10 by contribution profit overlap by roughly 50%. Half of every shop's bestsellers are not the most profitable products. A research note on the audit that surfaces the gap, the patterns hiding inside it, and the per-SKU operating decisions that recover margin.","2026-03-12",[210],"how-to-find-low-margin-skus-on-shopee",{"title":588,"description":589,"category":350,"date":590,"readMinutes":492,"featured":294,"cover":295,"format":223,"platforms":591,"draft":294,"slug":592},"How to plan Lazada campaigns around profit","Pay Day, Mega Sale, 11.11, and the long tail of category windows make participation feel mandatory. The platform documents the eligibility tiers; the seller absorbs them. A research note on what each campaign actually costs, when participation pays, and the margin-first rubric that survives all three.","2026-03-04",[213],"how-to-plan-lazada-campaigns-around-profit",{"title":594,"description":595,"category":338,"date":596,"readMinutes":402,"featured":294,"cover":295,"format":223,"platforms":597,"draft":294,"slug":598},"How to increase profit on TikTok Shop in 2026","TikTok Shop is the only SEA marketplace with a stacked second commission — affiliate commission (10–25% via the Open Affiliate Plan) layered on top of platform commission. A 6.0 platform ROAS routinely becomes ~1.4 true ROAS once the full four-line cost stack is subtracted. A research note on the affiliate-stack arithmetic, live-stream pricing discipline, and the per-SKU framework that recovers margin without retreating from the platform.","2026-02-26",[216],"how-to-increase-profit-on-tiktok-shop",{"title":600,"description":601,"category":430,"date":602,"readMinutes":603,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":604},"The SEA marketplace in 2026: a fragmented arena with one survival rule","Eighteen distinct platform-market cells, six countries, three platforms, and roughly a million sellers competing on a saturated product surface. The aggregate is growing; the per-cell margin distribution is compressing. A research note on the structural shifts and the operating rule that survives them.","2026-02-22",16,"sea-marketplace-2026",{"title":606,"description":607,"category":408,"date":608,"readMinutes":469,"featured":294,"cover":295,"format":226,"platforms":609,"draft":294,"slug":610},"Data ingestion for Shopee sellers: why zero-setup analytics matters","Most Shopee sellers don't have a strategy problem first. They have a data plumbing problem — orders, ads, COGS, fees, vouchers, inventory, pricing, and returns live in seven different surfaces, and by the time the seller has stitched them together the campaign is over. A research note on the data-source matrix, the canonical-entity model, and the zero-setup architecture that recovers ~10 hours per week.","2026-02-18",[210],"data-ingestion-for-shopee-sellers",{"title":612,"description":613,"category":430,"date":614,"readMinutes":344,"featured":294,"cover":295,"format":226,"platforms":615,"draft":294,"slug":616},"The race to zero margin — and how Shopee, Lazada, and TikTok Shop got there","Read the platform documentation end-to-end and the conclusion is uncomfortable: the discounts buyers see, the free shipping that drives conversion, and the affiliate spend that drives reach are all paid by the seller. The race-to-zero is the equilibrium output of that design.","2026-02-08",[210,213,216],"race-to-zero-margin",{"title":618,"description":619,"category":408,"date":620,"readMinutes":492,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":621},"ML demand forecasting for e-commerce sellers","Machine learning in e-commerce gets discussed in vague terms; for marketplace sellers the operating question is concrete — how many units of this SKU will sell in the next N days, with what confidence, and what decision flows from the answer? A research note on the practical model architecture, the stockout-distortion problem, sensitivity analysis, and the operating decisions forecasts feed.","2026-02-04","ml-demand-forecasting-for-ecommerce",{"title":623,"description":624,"category":430,"date":625,"readMinutes":415,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":626},"Complexity is the new tax on small sellers","A multi-shop seller in 2026 logs into seven platforms, reconciles four fee schedules, exports six CSVs, and re-enters COGS by hand. Big brands absorb the cost with a data team. Small sellers absorb it with their evenings — and it's the single biggest reason multi-shop operators stall before they reach scale.","2026-01-25","complexity-is-the-tax",{"title":628,"description":629,"category":327,"date":630,"readMinutes":402,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":631},"Stockout math for e-commerce sellers","A stockout is not one cost; it is five compounding costs. Lost contribution profit on the missed unit, plus wasted ad spend during the stockout window, plus algorithmic ranking demotion, plus repeat-buyer trust erosion, plus distorted forecasting that increases the likelihood of the next stockout. A research note on the multi-line stockout cost function, the per-SKU reorder-point math that accounts for it, and the campaign-aware adjustment that survives Pay Day and 11.11.","2026-01-18","inventory-stockout-math-ecommerce",{"title":633,"description":634,"category":350,"date":635,"readMinutes":469,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":636},"Dynamic pricing for marketplace sellers","Discounting is easy. Profitable pricing is hard. A 30% volume lift on a 10% price cut routinely lowers total contribution profit — the math says volume must lift by ~33% just to break even, and most SKUs underperform that bar. A research note on the price-elasticity arithmetic, the inventory × demand four-quadrant framework, and the per-SKU pricing decision that survives campaign-window pressure.","2026-01-04","dynamic-pricing-ecommerce-marketplaces",{"title":638,"description":639,"category":408,"date":640,"readMinutes":402,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":641},"E-commerce Decision Engine: How Marketplace Sellers Turn Data Into Profit Recommendations","A dashboard tells you what happened. A decision engine tells you what to do next, ranks the options by projected profit lift, and surfaces the math behind every recommendation. A research note on the five-layer architecture that separates the two, why marketplace commerce now requires the latter, and where the operating model breaks.","2025-12-12","how-ecommerce-decision-engines-work",{"title":643,"description":644,"category":327,"date":645,"readMinutes":389,"featured":294,"cover":295,"format":226,"platforms":646,"draft":294,"slug":647},"Multi-shop analytics for Shopee, Lazada, and TikTok Shop sellers","Modern marketplace sellers rarely operate on one channel. They have multiple dashboards but not one operating view. They know sales by channel — but not profit by channel.","2025-11-26",[210,213,216],"multi-shop-analytics-lazada-shopee-tiktok",{"title":649,"description":650,"category":430,"date":651,"readMinutes":652,"featured":294,"cover":295,"format":226,"platforms":295,"draft":294,"slug":653},"The new competitive world of commerce","A research note on what changed in marketplace commerce between 2022 and 2026. Platforms became active algorithmic counterparties; sellers who treat them as passive listing markets are competing against optimization systems they cannot see. The fix is to run your own optimization layer on top.","2025-11-12",15,"new-competitive-world-of-commerce",{"entry":655,"document":778,"toc":2453},{"id":656,"title":502,"category":408,"cover":295,"cta":295,"date":498,"description":503,"draft":294,"draftReason":295,"extension":657,"featured":294,"format":238,"listen":295,"math":294,"meta":658,"navigation":295,"ogImage":295,"path":659,"platforms":295,"readMinutes":504,"related":295,"robots":295,"schemaOrg":295,"seo":660,"sitemap":663,"sources":664,"stem":775,"tldr":295,"topic":776,"updatedAt":295,"__hash__":777},"blog_en\u002Fblog\u002Fdecision-intelligence-playbook-for-e-commerce.md","md",{},"\u002Fblog\u002Fdecision-intelligence-playbook-for-e-commerce",{"title":661,"description":662},"Decision Intelligence Playbook for E-commerce | DataGlass Research","A technical survey of decision intelligence for platform retailers — dynamic pricing, demand forecasting, inventory, promotions, and personalization.",{"loc":659},[665,667,669,671,673,675,677,679,681,683,685,687,689,691,693,695,697,699,701,703,705,707,709,711,713,715,717,719,721,723,725,727,729,731,733,735,737,739,741,743,745,747,749,751,753,755,757,759,761,763,765,767,769,771,773],{"title":666},"Amazon.com Inc. (2024). 2023 Annual Report and Form 10-K.",{"title":668},"Athey, S., & Imbens, G. W. (2019). Machine learning methods that economists should know about. Annual Review of Economics, 11, 685–725.",{"title":670},"Agrawal, S., & Goyal, N. (2012). Analysis of Thompson sampling for the multi-armed bandit problem. COLT 2012.",{"title":672},"Auer, P., Cesa-Bianchi, N., & Fischer, P. (2002). Finite-time analysis of the multiarmed bandit problem. Machine Learning, 47, 235–256.",{"title":674},"Bertsimas, D., & Kallus, N. (2020). From predictive to prescriptive analytics. Management Science, 66(3), 1025–1044.",{"title":676},"Berry, S., Levinsohn, J., & Pakes, A. (1995). Automobile prices in market equilibrium. Econometrica, 63(4), 841–890.",{"title":678},"Bezos, J. (2017). Letter to Shareholders. Amazon.com Inc.",{"title":680},"Chen, N., & Mišić, V. V. (2022). Decision-focused learning of revenue-management policies. Management Science, 68(8), 5921–5947.",{"title":682},"Chen, X., Owen, A. B., Pixton, C., & Simchi-Levi, D. (2015). Statistical learning of dynamic pricing strategies. Operations Research, 63(2), 326–339.",{"title":684},"Chen, M., Beutel, A., Covington, P., et al. (2019). Top-K off-policy correction for a REINFORCE recommender system. WSDM 2019.",{"title":686},"Cohen, M. C., Leung, N.-H. Z., Panchamgam, K., Perakis, G., & Smith, A. (2021). The impact of linear optimization on promotion planning. Operations Research, 69(1), 105–124.",{"title":688},"Covington, P., Adams, J., & Sargin, E. (2016). Deep neural networks for YouTube recommendations. RecSys 2016.",{"title":690},"Donoho, D. (2017). 50 years of data science. Journal of Computational and Graphical Statistics, 26(4), 745–766.",{"title":692},"Dudík, M., Erhan, D., Langford, J., & Li, L. (2014). Doubly robust policy evaluation and optimization. Statistical Science, 29(4), 485–511.",{"title":694},"Elmachtoub, A. N., & Grigas, P. (2022). Smart ‘predict, then optimize'. Management Science, 68(1), 9–26.",{"title":696},"Fisher, M., Gallino, S., & Li, J. (2018). Competition-based dynamic pricing in online retailing. Management Science, 64(6), 2496–2514.",{"title":698},"Gallego, G., & van Ryzin, G. (1994). Optimal dynamic pricing of inventories with stochastic demand. Management Science, 40(8), 999–1020.",{"title":700},"He, X., Deng, K., Wang, X., et al. (2020). LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. SIGIR 2020.",{"title":702},"Hyndman, R. J., Ahmed, R. A., Athanasopoulos, G., & Shang, H. L. (2011). Optimal combination forecasts for hierarchical time series. Computational Statistics & Data Analysis, 55(9), 2579–2589.",{"title":704},"Kallus, N. (2018). Balanced policy evaluation and learning. NeurIPS 2018.",{"title":706},"Kang, W.-C., & McAuley, J. (2018). Self-attentive sequential recommendation. ICDM 2018.",{"title":708},"Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender systems. IEEE Computer, 42(8), 30–37.",{"title":710},"Künzel, S. R., Sekhon, J. S., Bickel, P. J., & Yu, B. (2019). Meta-learners for estimating heterogeneous treatment effects using machine learning. PNAS, 116(10), 4156–4165.",{"title":712},"Lai, T. L., & Robbins, H. (1985). Asymptotically efficient adaptive allocation rules. Advances in Applied Mathematics, 6(1), 4–22.",{"title":714},"Li, L., Chu, W., Langford, J., & Schapire, R. E. (2010). A contextual-bandit approach to personalized news article recommendation. WWW 2010.",{"title":716},"Lim, B., Arık, S. O., Loeff, N., & Pfister, T. (2021). Temporal Fusion Transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4), 1748–1764.",{"title":718},"McKinsey & Company. (2023). The Multiplier Effect: How dynamic pricing is reshaping retail. McKinsey Insights.",{"title":720},"McKinsey & Company. (2024). The State of Small and Medium-Sized Businesses Report.",{"title":722},"Misra, K., Schwartz, E. M., & Abernethy, J. (2019). Dynamic online pricing with incomplete information using multi-armed bandit experiments. Marketing Science, 38(2), 226–252.",{"title":724},"Nie, X., & Wager, S. (2021). Quasi-oracle estimation of heterogeneous treatment effects. Biometrika, 108(2), 299–319.",{"title":726},"Oreshkin, B. N., Carpov, D., Chapados, N., & Bengio, Y. (2020). N-BEATS: Neural basis expansion analysis for interpretable time series forecasting. ICLR 2020.",{"title":728},"Phillips, R. L. (2005). Pricing and Revenue Optimization. Stanford University Press.",{"title":730},"Rendle, S., Freudenthaler, C., Gantner, Z., & Schmidt-Thieme, L. (2009). BPR: Bayesian personalized ranking from implicit feedback. UAI 2009.",{"title":732},"Robbins, H. (1952). Some aspects of the sequential design of experiments. Bulletin of the American Mathematical Society, 58(5), 527–535.",{"title":734},"Russo, D., Van Roy, B., Kazerouni, A., Osband, I., & Wen, Z. (2018). A tutorial on Thompson sampling. Foundations and Trends in ML, 11(1), 1–96.",{"title":736},"Salinas, D., Flunkert, V., Gasthaus, J., & Januschowski, T. (2020). DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting, 36(3), 1181–1191.",{"title":738},"Scott, S. L. (2010). A modern Bayesian look at the multi-armed bandit. Applied Stochastic Models in Business and Industry, 26(6), 639–658.",{"title":740},"Shopify. (2024). Commerce Trends 2024 Report.",{"title":742},"Smith, S. A., & Achabal, D. D. (1998). Clearance pricing and inventory policies for retail chains. Management Science, 44(3), 285–300.",{"title":744},"Snyder, L. V., & Shen, Z.-J. M. (2019). Fundamentals of Supply Chain Theory (2nd ed.). Wiley.",{"title":746},"Statista Research Department. (2024). Global retail e-commerce sales 2014–2027.",{"title":748},"Talluri, K. T., & van Ryzin, G. J. (2004). The Theory and Practice of Revenue Management. Springer.",{"title":750},"Taylor, S. J., & Letham, B. (2018). Forecasting at scale. The American Statistician, 72(1), 37–45.",{"title":752},"U.S. Census Bureau. (2024). Quarterly Retail E-Commerce Sales, 4th Quarter 2023.",{"title":754},"Jungle Scout. (2024). The State of the Amazon Seller 2024.",{"title":756},"Wager, S., & Athey, S. (2018). Estimation and inference of heterogeneous treatment effects using random forests. Journal of the American Statistical Association, 113(523), 1228–1242.",{"title":758},"Wedel, M., & Kannan, P. K. (2016). Marketing analytics for data-rich environments. Journal of Marketing, 80(6), 97–121.",{"title":760},"Ying, R., He, R., Chen, K., et al. (2018). Graph convolutional neural networks for web-scale recommender systems (PinSage). KDD 2018.",{"title":762},"Yi, X., et al. (2019). Sampling-bias-corrected neural modeling for large corpus item recommendations. RecSys 2019.",{"title":764},"Zhou, G., Zhu, X., Song, C., et al. (2018). Deep Interest Network for click-through rate prediction. KDD 2018.",{"title":766},"Scarf, H. (1958). A min-max solution of an inventory problem. Studies in the Mathematical Theory of Inventory and Production.",{"title":768},"Clark, A. J., & Scarf, H. (1960). Optimal policies for a multi-echelon inventory problem. Management Science, 6(4), 475–490.",{"title":770},"Mohajerin Esfahani, P., & Kuhn, D. (2018). Data-driven distributionally robust optimization using the Wasserstein metric. Mathematical Programming, 171(1–2), 115–166.",{"title":772},"Chen, T., Sun, Y., Shi, Y., & Hong, L. (2017). On sampling strategies for neural network-based collaborative filtering. KDD 2017.",{"title":774},"Sun, F., Liu, J., Wu, J., et al. (2019). BERT4Rec: Sequential recommendation with bidirectional encoder representations from Transformer. CIKM 2019.","blog\u002Fdecision-intelligence-playbook-for-e-commerce","decision-science","EW6c6Hzu_8J3c64jVCqqk37Bm4nrksFV6HHWndC45jw",{"frontmatter":779,"meta":837,"nodes":952},{"title":502,"description":503,"seo":780,"category":408,"format":238,"topic":776,"date":498,"readMinutes":504,"sources":781},{"title":661,"description":662},[782,783,784,785,786,787,788,789,790,791,792,793,794,795,796,797,798,799,800,801,802,803,804,805,806,807,808,809,810,811,812,813,814,815,816,817,818,819,820,821,822,823,824,825,826,827,828,829,830,831,832,833,834,835,836],{"title":666},{"title":668},{"title":670},{"title":672},{"title":674},{"title":676},{"title":678},{"title":680},{"title":682},{"title":684},{"title":686},{"title":688},{"title":690},{"title":692},{"title":694},{"title":696},{"title":698},{"title":700},{"title":702},{"title":704},{"title":706},{"title":708},{"title":710},{"title":712},{"title":714},{"title":716},{"title":718},{"title":720},{"title":722},{"title":724},{"title":726},{"title":728},{"title":730},{"title":732},{"title":734},{"title":736},{"title":738},{"title":740},{"title":742},{"title":744},{"title":746},{"title":748},{"title":750},{"title":752},{"title":754},{"title":756},{"title":758},{"title":760},{"title":762},{"title":764},{"title":766},{"title":768},{"title":770},{"title":772},{"title":774},{"toc":838},{"title":839,"searchDepth":313,"depth":313,"links":840},"",[841,844,847,850,863,882,892,905,918,937,940,943,946,949],{"id":842,"depth":313,"text":843},"executive-summary","Executive summary",{"id":845,"depth":313,"text":846},"_1-the-state-of-e-commerce-and-where-margin-is-being-made-and-lost","1.  The state of e-commerce, and where margin is being made and lost",{"id":848,"depth":313,"text":849},"_2-quantifying-the-smb-sellers-pain","2.  Quantifying the SMB seller's pain",{"id":851,"depth":313,"text":852,"children":853},"_3-pricing-from-heuristics-to-learning-systems","3.  Pricing — from heuristics to learning systems",[854,857,860],{"id":855,"depth":308,"text":856},"_3-pricing-from-heuristics-to-learning-systems-_31-demand-modeling","3.1  Demand modeling",{"id":858,"depth":308,"text":859},"_3-pricing-from-heuristics-to-learning-systems-_32-optimization-given-a-demand-model","3.2  Optimization given a demand model",{"id":861,"depth":308,"text":862},"_3-pricing-from-heuristics-to-learning-systems-_33-online-learning-exploration-policies","3.3  Online learning — exploration policies",{"id":864,"depth":313,"text":865,"children":866},"_4-forecasting-making-demand-legible","4.  Forecasting — making demand legible",[867,870,873,876,879],{"id":868,"depth":308,"text":869},"_4-forecasting-making-demand-legible-_41-classical-statistical-forecasters","4.1  Classical statistical forecasters",{"id":871,"depth":308,"text":872},"_4-forecasting-making-demand-legible-_42-hierarchical-and-reconciled-forecasts","4.2  Hierarchical and reconciled forecasts",{"id":874,"depth":308,"text":875},"_4-forecasting-making-demand-legible-_43-machine-learning-forecasters","4.3  Machine-learning forecasters",{"id":877,"depth":308,"text":878},"_4-forecasting-making-demand-legible-_44-deep-learning-forecasters","4.4  Deep-learning forecasters",{"id":880,"depth":308,"text":881},"_4-forecasting-making-demand-legible-_45-how-to-choose","4.5  How to choose",{"id":883,"depth":313,"text":884,"children":885},"_5-inventory-matching-supply-to-demand-under-uncertainty","5.  Inventory — matching supply to demand under uncertainty",[886,889],{"id":887,"depth":308,"text":888},"_5-inventory-matching-supply-to-demand-under-uncertainty-_51-multi-period-and-continuous-review-variants","5.1  Multi-period and continuous-review variants",{"id":890,"depth":308,"text":891},"_5-inventory-matching-supply-to-demand-under-uncertainty-_52-markdowns-and-clearance","5.2  Markdowns and clearance",{"id":893,"depth":313,"text":894,"children":895},"_6-promotions-and-discounting-where-causality-matters","6.  Promotions and discounting — where causality matters",[896,899,902],{"id":897,"depth":308,"text":898},"_6-promotions-and-discounting-where-causality-matters-_61-meta-learners-for-cate","6.1  Meta-learners for CATE",{"id":900,"depth":308,"text":901},"_6-promotions-and-discounting-where-causality-matters-_62-from-cate-to-assignment-the-optimization-layer","6.2  From CATE to assignment — the optimization layer",{"id":903,"depth":308,"text":904},"_6-promotions-and-discounting-where-causality-matters-_63-measurement","6.3  Measurement",{"id":906,"depth":313,"text":907,"children":908},"_7-personalization-and-recommendation","7.  Personalization and recommendation",[909,912,915],{"id":910,"depth":308,"text":911},"_7-personalization-and-recommendation-_71-retrieval","7.1  Retrieval",{"id":913,"depth":308,"text":914},"_7-personalization-and-recommendation-_72-ranking-and-re-ranking","7.2  Ranking and re-ranking",{"id":916,"depth":308,"text":917},"_7-personalization-and-recommendation-_73-learning-from-logs-bandits-and-counterfactual-evaluation","7.3  Learning from logs — bandits and counterfactual evaluation",{"id":919,"depth":313,"text":920,"children":921},"_8-putting-it-together-a-reference-decision-intelligence-stack","8.  Putting it together — a reference decision-intelligence stack",[922,925,928,931,934],{"id":923,"depth":308,"text":924},"_8-putting-it-together-a-reference-decision-intelligence-stack-_81-data-layer","8.1  Data layer",{"id":926,"depth":308,"text":927},"_8-putting-it-together-a-reference-decision-intelligence-stack-_82-feature-store","8.2  Feature store",{"id":929,"depth":308,"text":930},"_8-putting-it-together-a-reference-decision-intelligence-stack-_83-model-layer","8.3  Model layer",{"id":932,"depth":308,"text":933},"_8-putting-it-together-a-reference-decision-intelligence-stack-_84-optimization-and-decision-layer","8.4  Optimization and decision layer",{"id":935,"depth":308,"text":936},"_8-putting-it-together-a-reference-decision-intelligence-stack-_85-experimentation-and-off-policy-evaluation","8.5  Experimentation and off-policy evaluation",{"id":938,"depth":313,"text":939},"_9-how-to-select-techniques-by-seller-stage","9.  How to select techniques by seller stage",{"id":941,"depth":313,"text":942},"_10-risks-governance-and-considerations","10.  Risks, governance, and considerations",{"id":944,"depth":313,"text":945},"_11-conclusion","11.  Conclusion",{"id":947,"depth":313,"text":948},"faq","FAQ",{"id":950,"depth":313,"text":951},"references","References",[953,956,965,976,980,982,992,999,1005,1007,1021,1045,1048,1051,1057,1156,1173,1175,1186,1189,1192,1197,1204,1208,1214,1218,1228,1232,1239,1243,1254,1259,1361,1366,1368,1371,1380,1386,1390,1393,1397,1403,1407,1414,1418,1425,1429,1439,1441,1452,1456,1459,1463,1470,1474,1488,1493,1664,1669,1673,1680,1684,1687,1691,1701,1704,1706,1712,1714,1718,1721,1725,1728,1732,1739,1741,1748,1750,1754,1757,1761,1764,1766,1770,1777,1781,1788,1792,1799,1803,1806,1808,1813,1896,1898,1901,1906,1909,1941,1943,1947,1953,1957,1960,1964,1971,1975,1978,1982,1992,1996,1999,2001,2007,2009,2012,2017,2020,2022,2026,2029,2033,2036,2040,2047,2051,2057,2061,2064,2068,2075,2079,2082,2084,2087,2090,2096,2098,2101,2103,2110,2112,2116,2119,2123,2134,2138,2145,2149,2160,2162,2166,2169,2173,2183,2187,2198,2200,2209,2211,2214,2219,2248,2250,2253,2255,2258,2260,2263,2265,2268,2270,2276,2278,2281,2286,2369,2371,2375,2378,2382,2385,2389,2392,2396,2399,2403,2406,2408,2411,2414,2416,2419,2424,2432,2436,2440,2447,2451],[954,955,843],"h2",{"id":842},[957,958,959,960,964],"p",{},"Retail e-commerce now absorbs roughly one in six retail dollars worldwide, and the operating leverage is concentrated at the top of the distribution: a small number of platform-scale retailers run end-to-end decision-intelligence systems that touch every economically significant lever, while most small and medium-sized (SMB) sellers continue to rely on spreadsheets, marketplace dashboards, and rule-based heuristics ",[961,962],"blog-cite",{"n":963},"44",":blog-cite:blog-cite:blog-cite. This report is a deep technical survey of the methods used by the leading operators, the variants that matter, and the architectural and operational choices that determine whether those methods translate into margin in practice.",[957,966,967,968,971,972,975],{},"Our central claim is simple. The pain points reported by SMB sellers—pricing pressure, inventory mismatch, rising acquisition cost, returns, lack of analytics tools—are not, in 2026, primarily a research problem. The methods needed to address them are documented in textbooks ",[961,969],{"n":970},"32",":blog-cite:blog-cite, in canonical journal articles ",[961,973],{"n":974},"17",":blog-cite:blog-cite:blog-cite:blog-cite:blog-cite, and in widely adopted open-source software (Stan, PyMC, GluonTS, EconML, CausalML, LightGBM, Transformers). The bottleneck has shifted from algorithmic novelty to integration: connectors, feature stores, optimization, experimentation, off-policy evaluation, and a human-in-the-loop UI assembled into a single workflow that a merchandiser will actually use.",[977,978],"blog-pullquote",{"text":979},"The bottleneck for most sellers is no longer algorithmic novelty. It is the integration of data, models, and optimization into a single workflow.",[954,981,846],{"id":845},[957,983,984,985,987,988,991],{},"Three structural forces have reshaped seller economics over the past five years. First, online penetration is high but uneven: U.S. e-commerce penetration was 15.4% of total retail in Q4 2023 ",[961,986],{"n":963},", above 27% in China and roughly 26% in the United Kingdom ",[961,989],{"n":990},"41",". Within categories, penetration ranges from below 5% in food and beverage to above 50% in consumer electronics and apparel. The marketplace structure—where a small number of platforms intermediate between many buyers and many sellers—has become the dominant organizational form, increasing price comparability and shortening the half-life of any given price advantage.",[957,993,994,995,998],{},"Second, returns and last-mile fulfillment have re-priced the unit economics. Industry estimates place online return rates at 25–40% in apparel ",[961,996],{"n":997},"38",", and free-shipping thresholds plus carrier surcharges have compressed the gap between gross and contribution margin. The seller's problem is no longer ‘maximize revenue subject to inventory'; it is ‘maximize contribution after returns, ad cost, and platform fees, subject to working-capital and service-level constraints.' This reframing matters mathematically because the relevant objective function changes: profit per unit becomes price minus cost minus expected return cost minus expected ad cost minus expected platform fee, and the optimization variables that determine each term differ.",[957,1000,1001,1002,1004],{},"Third, paid-media customer-acquisition cost has risen faster than basket size. Median cost-per-click on Meta and Google search has risen by an estimated 60–110% since 2020 ",[961,1003],{"n":997},":blog-cite, while average order value has grown roughly with inflation. The competitive equilibrium has shifted toward sellers who can extract more revenue per customer—through pricing, recommendation, and retention—rather than acquire more customers at constant value. These three forces explain why the levers we focus on (pricing, inventory, promotion, personalization) are precisely the ones with the highest residual marginal value at scale, and why they are the levers that platform-scale retailers have invested most heavily in automating.",[954,1006,849],{"id":848},[957,1008,1009,1010,1013,1014,1016,1017,1020],{},"Figure 1 summarizes the operational pain points reported by SMB sellers in 2023–24, synthesized from the Jungle Scout State of the Amazon Seller survey (n = 2,164) ",[961,1011],{"n":1012},"45",", the Shopify Commerce Trends 2024 report ",[961,1015],{"n":997},", and the McKinsey State of Small and Medium-Sized Businesses survey ",[961,1018],{"n":1019},"28",".",[1022,1023],"blog-chart",{"kind":1024,"title":1025,"labels":1026,"values":1035,"unit":1043,"caption":1044},"bar","Figure 1. Top operational pain points reported by SMB e-commerce sellers (2023–24).",[1027,1028,1029,1030,1031,1032,1033,1034],"Finding profitable products","Pricing and margin pressure","Inventory and stockouts","Rising ad-acquisition cost","Cash flow and working capital","Returns and reverse logistics","Platform fee changes","Lack of analytics tools",[1036,1037,1038,1039,513,1040,1041,1042],62,58,55,53,41,38,34,"%","Share of SMB sellers reporting the issue.",[957,1046,1047],{},"Three patterns dominate. After a seller has chosen what to sell, the binding constraints are commercial—price, promotion, inventory, ad cost—not supply. Roughly one in three sellers volunteers lack of analytics tools as a constraint, an unusually high rate for an unprompted answer, indicating a strong revealed preference for tooling. And the cluster of margin-, inventory-, and acquisition-related issues is not coincidental: each is a symptom of a missing decision-support apparatus, and each is addressed by a distinct branch of the technical literature surveyed in the rest of this report.",[957,1049,1050],{},"Table 1 contrasts the typical SMB tooling stack with what platform-scale retailers use. The pattern is consistent across surveys: SMB tooling is dominated by spreadsheets, marketplace dashboards, and rule-based repricers, while platform-scale tooling is dominated by purpose-built optimization, machine learning, and continuous experimentation infrastructure. The capability gap (Figure 2) is largest in causal attribution and experimentation—precisely the layers that compound, because they are the layers that allow every other model to improve over time.",[957,1052,1053],{},[1054,1055,1056],"em",{},"Table 1. The capability gap, decision domain by decision domain.",[1058,1059,1060,1076],"table",{},[1061,1062,1063],"thead",{},[1064,1065,1066,1070,1073],"tr",{},[1067,1068,1069],"th",{},"Decision domain",[1067,1071,1072],{},"Typical SMB tool",[1067,1074,1075],{},"Typical large-retailer tool",[1077,1078,1079,1090,1101,1112,1123,1134,1145],"tbody",{},[1064,1080,1081,1084,1087],{},[1082,1083,27],"td",{},[1082,1085,1086],{},"Manual; rule-based ‘competitor minus $0.01' repricer",[1082,1088,1089],{},"Demand model + bandit \u002F RL system, A\u002FB tested daily",[1064,1091,1092,1095,1098],{},[1082,1093,1094],{},"Promotion planning",[1082,1096,1097],{},"Calendar in spreadsheet; gut-feel discount depth",[1082,1099,1100],{},"MIP optimizer + uplift model, vendor-budget-aware",[1064,1102,1103,1106,1109],{},[1082,1104,1105],{},"Demand forecast",[1082,1107,1108],{},"Trailing-30-day average",[1082,1110,1111],{},"Hierarchical Bayesian \u002F deep-learning forecaster with covariates",[1064,1113,1114,1117,1120],{},[1082,1115,1116],{},"Inventory",[1082,1118,1119],{},"Reorder point set by hand",[1082,1121,1122],{},"Multi-echelon optimization with newsvendor critical fractiles",[1064,1124,1125,1128,1131],{},[1082,1126,1127],{},"Personalization",[1082,1129,1130],{},"Static collections \u002F bestsellers",[1082,1132,1133],{},"Per-user neural ranker retrained nightly",[1064,1135,1136,1139,1142],{},[1082,1137,1138],{},"Attribution",[1082,1140,1141],{},"Last-click in platform UI",[1082,1143,1144],{},"Causal media-mix model + uplift testing + geo experiments",[1064,1146,1147,1150,1153],{},[1082,1148,1149],{},"Experimentation",[1082,1151,1152],{},"Ad-hoc trials",[1082,1154,1155],{},"Continuous A\u002FB platform with sequential testing",[1022,1157],{"kind":1024,"title":1158,"labels":1159,"values":1165,"compare":1166,"series":1168,"max":1171,"caption":1172},"Figure 2. Capability maturity (1–10) across decision domains. The shaded band is the decision-intelligence gap between platform-scale retailers and a typical SMB seller.",[1160,1161,1162,1163,1127,1164,1149],"Demand forecasting","Dynamic pricing","Promo optimization","Inventory optimization","Causal attribution",[301,301,301,389,301,389,301],[308,313,313,293,308,1167,313],1,[1169,1170],"Platform-scale retailer","Typical SMB seller","10","Capability maturity, scored out of 10.",[954,1174,852],{"id":851},[957,1176,1177,1178,1181,1182,1185],{},"Pricing is the single highest-leverage lever in retail because every dollar of price flows directly to contribution. McKinsey estimates that even a basic dynamic-pricing program adds 1–5% margin across categories ",[961,1179],{"n":1180},"27",", and well-instrumented field experiments have shown gross-profit lifts of 80% or more relative to rule-based heuristics ",[961,1183],{"n":1184},"29",". A modern pricing system has three components, treated below in turn: a demand model that maps prices to expected sales; an optimizer that maximizes contribution subject to business rules; and an exploration policy that updates the demand model as new data arrives. The literature treats these as separable, but they interact: the optimizer's structure determines what the demand model needs to predict, and the exploration policy determines what the demand model can be identified from.",[1187,1188,856],"h3",{"id":855},[957,1190,1191],{},"A demand model is a function Q(p, x) predicting expected unit sales at price p and covariates x (seasonality, competitor prices, marketing spend, inventory position). The choice of functional form matters because it determines what the optimizer can do, how much data is needed for stable estimation, and what kind of identification problem must be solved. Six families of demand model are in routine industrial use.",[1193,1194,1196],"h4",{"id":1195},"_3-pricing-from-heuristics-to-learning-systems-_31-demand-modeling-linear-and-log-linear-demand","Linear and log-linear demand",[957,1198,1199,1200,1203],{},"The linear specification Q = a − bp + γ′x + ε is the simplest workable model. It is fast to estimate by ordinary least squares, transparent, and admits a closed-form revenue maximum at p* = (a + γ′x) \u002F (2b) when marginal cost is zero. The log-linear specification log Q = α − β log p + γ′x is more useful in retail because the slope coefficient β is the price elasticity of demand—directly interpretable, dimensionless, and stable across price levels in many categories. Both forms are vulnerable to price endogeneity: prices in observational data are not random; they reflect the seller's beliefs about demand. Naive regression therefore conflates the demand curve with the supply response. The standard remedies are instrumental variables (cost shocks, competitor cost shocks ",[961,1201],{"n":1202},"6","), within-product fixed effects with exogenous price changes (e.g., calendar-driven promotional waves), or a deliberate randomization policy that supplies its own identification, which is why the bandit machinery in §3.3 is more than a curiosity: it is the cleanest way to estimate elasticity at all.",[1193,1205,1207],{"id":1206},"_3-pricing-from-heuristics-to-learning-systems-_31-demand-modeling-constant-elasticity-demand","Constant-elasticity demand",[957,1209,1210,1211,1213],{},"The constant-elasticity form Q(p) = A · p^ε, with ε \u003C 0, admits closed-form revenue R(p) = A · p^(1+ε) and a profit maximum that satisfies the classical Lerner condition (p* − c) \u002F p* = 1 \u002F |ε|. The form is appealing for two reasons. First, the optimal markup over marginal cost is a simple function of one parameter, which makes the model auditable. Second, in panel data with cost variation, ε is identified from the slope of log-quantity on log-price, holding fixed effects constant. The form's main weakness is that it assumes elasticity is constant across price levels, which is rarely true near psychological thresholds (e.g., $9.99, $19.99) or near competitor caps. Production implementations use a piecewise-constant-elasticity model with breakpoints at salient prices ",[961,1212],{"n":970},", or fall back to logit and ML alternatives below.",[1193,1215,1217],{"id":1216},"_3-pricing-from-heuristics-to-learning-systems-_31-demand-modeling-discrete-choice-logit-nested-logit-mixed-logit","Discrete-choice (logit, nested logit, mixed logit)",[957,1219,1220,1221,1224,1225,1227],{},"When customers choose among substitutes—the typical online-shopping decision—a discrete-choice model is the structurally appropriate object. The multinomial logit (MNL) gives the share of product j in category C as s_j(p) = exp(α_j − β_j p_j) \u002F (1 + Σ_ exp(α_k − β_k p_k)). The MNL is the workhorse of marketing and retail assortment optimization ",[961,1222],{"n":1223},"42",":blog-cite, but inherits the Independence of Irrelevant Alternatives (IIA) property: the relative odds of two products are unaffected by the presence of a third, which is empirically false for close substitutes (red vs. blue versions of the same shirt). The nested logit relaxes IIA across nests (color\u002Fsize within product) by introducing a within-nest correlation parameter. The mixed (random-coefficients) logit allows the price coefficient β to vary across customers as a draw from a population distribution, capturing taste heterogeneity that no aggregate model can. Estimating the mixed logit on aggregated market data is the BLP problem ",[961,1226],{"n":1202},", which combines a contraction mapping for shares with GMM for parameters and is the structural foundation of modern industrial-organization studies of pricing. For sellers with customer-level data, hierarchical Bayesian estimation is more practical and is the bridge to the next family.",[1193,1229,1231],{"id":1230},"_3-pricing-from-heuristics-to-learning-systems-_31-demand-modeling-hierarchical-bayesian-demand","Hierarchical Bayesian demand",[957,1233,1234,1235,1238],{},"A wide catalog with short per-SKU history is the typical SMB regime. Hierarchical (multi-level) Bayesian models exploit the structure that, within a category, SKU-level parameters are exchangeable draws from a category-level prior. Concretely, log Q_ = α_j − β_j log p_ + γ_j′ x_ + ε_ with (α_j, β_j, γ_j) ∼ N(μ_c, Σ_c) where c indexes the category. The model is fitted with MCMC (Stan, PyMC) or variational inference. The benefit is partial pooling: SKUs with thin data are shrunk toward the category mean, while SKUs with abundant data dominate their own posterior. Empirically this delivers calibrated posteriors with as little as 6–8 weeks of data per SKU when the catalog has a few hundred comparable items ",[961,1236],{"n":1237},"2",", which is the regime where every SMB seller actually operates. The downside is computational: fitting a thousand-SKU panel with full MCMC is non-trivial, but variational inference and stochastic gradient samplers have largely closed the gap in practice.",[1193,1240,1242],{"id":1241},"_3-pricing-from-heuristics-to-learning-systems-_31-demand-modeling-black-box-ml-demand-models","Black-box ML demand models",[957,1244,1245,1246,1249,1250,1253],{},"Gradient-boosted trees (LightGBM, XGBoost) and deep nets can fit Q(p, x) non-parametrically, capturing seasonality, holiday effects, weather, and feature interactions that a parametric form would miss. The technical risk is that maximum-likelihood estimation does not respect the downstream optimization: a model with low MSE on the demand surface can still produce poorly conditioned profit landscapes. Two responses to this have crystallized in the literature. Decision-focused learning (the Smart-Predict-Then-Optimize, SPO+, framework) trains the predictor under a loss function that scores end-task decisions rather than predictions ",[961,1247],{"n":1248},"15",":blog-cite. And predictive-to-prescriptive analytics ",[961,1251],{"n":1252},"5"," uses the predictor to weight historical scenarios in a sample-average optimization, sidestepping calibration issues at the cost of richer scenario data. For pricing specifically, ML demand models work best when paired with bandit-driven exploration so the model is identified, not just fitted.",[957,1255,1256],{},[1054,1257,1258],{},"Figure 3. Constant-elasticity demand and revenue. Both panels assume A = 10⁵; the location of the revenue maximum depends only on |ε| and marginal cost.",[1058,1260,1261,1277],{},[1061,1262,1263],{},[1064,1264,1265,1268,1271,1274],{},[1067,1266,1267],{},"Price",[1067,1269,1270],{},"ε = −0.8",[1067,1272,1273],{},"ε = −1.5",[1067,1275,1276],{},"ε = −2.5",[1077,1278,1279,1293,1306,1319,1333,1347],{},[1064,1280,1281,1284,1287,1290],{},[1082,1282,1283],{},"8",[1082,1285,1286],{},"18,946 (151,568)",[1082,1288,1289],{},"4,419 (35,352)",[1082,1291,1292],{},"552 (4,416)",[1064,1294,1295,1297,1300,1303],{},[1082,1296,1171],{},[1082,1298,1299],{},"15,849 (158,490)",[1082,1301,1302],{},"3,162 (31,620)",[1082,1304,1305],{},"316 (3,160)",[1064,1307,1308,1310,1313,1316],{},[1082,1309,1248],{},[1082,1311,1312],{},"11,458 (171,870)",[1082,1314,1315],{},"1,721 (25,815)",[1082,1317,1318],{},"115 (1,725)",[1064,1320,1321,1324,1327,1330],{},[1082,1322,1323],{},"20",[1082,1325,1326],{},"9,103 (182,060)",[1082,1328,1329],{},"1,118 (22,360)",[1082,1331,1332],{},"56 (1,120)",[1064,1334,1335,1338,1341,1344],{},[1082,1336,1337],{},"30",[1082,1339,1340],{},"6,581 (197,430)",[1082,1342,1343],{},"609 (18,270)",[1082,1345,1346],{},"20 (600)",[1064,1348,1349,1352,1355,1358],{},[1082,1350,1351],{},"40",[1082,1353,1354],{},"5,228 (209,120)",[1082,1356,1357],{},"395 (15,800)",[1082,1359,1360],{},"10 (400)",[957,1362,1363],{},[1054,1364,1365],{},"Demand q = 100,000 · p^ε at each price; revenue p · q in brackets.",[1187,1367,859],{"id":858},[957,1369,1370],{},"Once demand is estimated, prices are chosen by solving a constrained optimization problem. The simplest case—single product, constant elasticity, marginal cost c, no constraints—gives the Lerner condition by setting the derivative of profit (p − c)·A·p^ε to zero:",[1372,1373,1375],"pre",{"language":1374},"text",[1376,1377,1379],"code",{"class":1378},"language-text","(p* − c) \u002F p*  =  1 \u002F |ε|     ⇔     p* = c · |ε| \u002F (|ε| − 1)",[957,1381,1382,1383,1385],{},"The same logic generalizes. Under linear demand the closed form is p* = (a + bc) \u002F (2b), and under MNL with constant marginal costs the optimal price for product j satisfies p_j* = c_j + 1 \u002F (β_j (1 − s_j*)), a fixed-point equation in the equilibrium share ",[961,1384],{"n":1223},". These closed forms are what make the constant-elasticity and logit families durable: even when a black-box ML model fits the data better, sellers often run the optimizer on a parametric approximation in the neighborhood of the current operating point because the parametric form admits a defensible, interpretable optimum.",[1193,1387,1389],{"id":1388},"_3-pricing-from-heuristics-to-learning-systems-_32-optimization-given-a-demand-model-constrained-single-sku-pricing","Constrained single-SKU pricing",[957,1391,1392],{},"Production pricing systems do not solve unconstrained problems. Margin floors (p ≥ c · (1 + m_min)), Manufacturer-Authorized-Price (MAP) rules, competitor caps (p ≤ p_comp + δ), and price-ending requirements (p ∈ ) narrow the feasible set. The standard approach is to evaluate the unconstrained Lerner-implied price, then project onto the feasible set; for non-convex constraints (price endings) this is a small enumeration. KKT conditions tell us when binding constraints distort the optimum, which is operationally useful: a category whose Lerner-implied price is consistently above the MAP cap is one where the seller is leaving margin on the table because of vendor policy, not demand.",[1193,1394,1396],{"id":1395},"_3-pricing-from-heuristics-to-learning-systems-_32-optimization-given-a-demand-model-joint-multi-sku-pricing-under-cannibalization","Joint multi-SKU pricing under cannibalization",[957,1398,1399,1400,1402],{},"When products are substitutes, the optimal price for one is a function of the prices of all close substitutes. Under linear demand Q_j = a_j − Σ_k B_ p_k the optimization is a quadratic program with closed form p* = (B + Bᵀ)⁻¹ (a + Bc). Under logit demand the structure is non-trivial but well behaved: the seller's problem on a single nest reduces to choosing a single markup multiplier across the nest ",[961,1401],{"n":1223},", which dramatically simplifies estimation. Modern implementations cluster SKUs into demand-similar groups and solve the joint program at the group level, then back out individual prices. The mathematical pay-off of joint optimization is largest in categories with high cross-elasticity (apparel sizes, color variants); in categories with low cross-elasticity (specialty hardware), separable single-SKU optimization captures most of the benefit.",[1193,1404,1406],{"id":1405},"_3-pricing-from-heuristics-to-learning-systems-_32-optimization-given-a-demand-model-mixed-integer-programs-for-promotion-aware-pricing","Mixed-integer programs for promotion-aware pricing",[957,1408,1409,1410,1413],{},"Many production pricing problems are intrinsically discrete: the seller chooses, for each SKU and each week, whether to promote, by how much, and through which mechanism (TPR, multibuy, coupon, bundle). The natural formulation is a mixed-integer program (MIP) over binary promotion variables and continuous depth variables, with vendor-funded budget constraints, category caps, and pantry-loading penalties to discourage same-customer reuse. Cohen et al. ",[961,1411],{"n":1412},"11"," formalize this and report empirical lifts of 3–5% in retail trials when MIP-based promotion planning replaces calendar-based heuristics. The LP relaxation is typically very tight in this problem class, which is why LP solutions—rounded with a small post-processing step—are the practical workhorse rather than full branch-and-bound.",[1193,1415,1417],{"id":1416},"_3-pricing-from-heuristics-to-learning-systems-_32-optimization-given-a-demand-model-robust-and-distributionally-robust-pricing","Robust and distributionally robust pricing",[957,1419,1420,1421,1424],{},"All of the above takes the demand model as given. In practice elasticity estimates have wide credible intervals, and over-fitting the optimizer to a point estimate can produce large losses when the world deviates. Robust optimization hedges by maximizing the worst-case profit over a credible set of demand functions; distributionally robust optimization (DRO) does the same over a Wasserstein or moment-based ambiguity set around the empirical demand distribution ",[961,1422],{"n":1423},"53",". The practical effect is that DRO solutions are biased toward prices in the interior of the historical price support, where data is densest. For a seller deploying pricing for the first time, this is exactly the right inductive bias.",[1193,1426,1428],{"id":1427},"_3-pricing-from-heuristics-to-learning-systems-_32-optimization-given-a-demand-model-dynamic-pricing-with-finite-inventory","Dynamic pricing with finite inventory",[957,1430,1431,1432,1434,1435,1438],{},"Single-period pricing ignores the trade-off between selling now and saving units for later. The Gallego–van Ryzin formulation ",[961,1433],{"n":974}," treats price as a control of a Poisson sales process over a finite horizon with given starting inventory, and derives an optimal price that decreases as remaining inventory grows and increases as remaining time shrinks. The same machinery underlies airline yield management and retail markdown systems ",[961,1436],{"n":1437},"39",". In e-commerce, the most common application is end-of-season clearance: the optimizer schedules a sequence of markdowns that draws inventory toward zero by season end while maximizing expected residual revenue.",[1187,1440,862],{"id":861},[957,1442,1443,1444,1447,1448,1451],{},"The optimizer of §3.2 requires an estimated demand function. Where does the estimate come from? In principle, from historical data; in practice, that data has been generated by the seller's own past pricing rules, which makes it a censored, non-experimental sample. The cleanest solution is to treat pricing as a sequential learning problem: a multi-armed bandit in which arms are candidate prices, rewards are realized contributions, and the seller's job is to balance exploitation of the currently best-estimated arm against exploration of arms that may turn out to be better. The bandit literature dates to Robbins (1952) ",[961,1445],{"n":1446},"34",", and the modern theory begins with the Lai–Robbins lower bound ",[961,1449],{"n":1450},"24",", which establishes that any policy with logarithmic regret in the number of trials is asymptotically optimal.",[1193,1453,1455],{"id":1454},"_3-pricing-from-heuristics-to-learning-systems-_33-online-learning-exploration-policies-greedy","ε-greedy",[957,1457,1458],{},"At each step, the seller plays the empirically best arm with probability 1 − ε and a uniformly random arm with probability ε. Implementation is trivial; the regret bound is O(εT) + O((K log T) \u002F Δ), which is sublinear only if ε decays appropriately. ε-greedy is therefore the right baseline, not the right policy. Its main use in production is as a smoke test for the rest of the pipeline: if ε-greedy with ε = 0.1 cannot find a better-than-status-quo price within a few thousand customer arrivals, the data plumbing or the demand structure is the problem, not the algorithm.",[1193,1460,1462],{"id":1461},"_3-pricing-from-heuristics-to-learning-systems-_33-online-learning-exploration-policies-upper-confidence-bound-ucb","Upper Confidence Bound (UCB)",[957,1464,1465,1466,1469],{},"UCB1 ",[961,1467],{"n":1468},"4"," plays the arm that maximizes p̂_k + √(2 log t \u002F n_k), where p̂_k is the empirical mean reward and n_k is the number of pulls. The intuition is optimism in the face of uncertainty: the seller picks the arm whose plausibly highest reward is the largest. UCB has provably O(log T) regret, and—important for retail compliance—is fully deterministic given the data, which makes audit trails straightforward. The reliability cost is that UCB often over-explores high-variance arms early, which can be expensive when one arm has a large negative cost (e.g., a much-too-low price). KL-UCB and Bayes-UCB are tighter variants that perform better in practice.",[1193,1471,1473],{"id":1472},"_3-pricing-from-heuristics-to-learning-systems-_33-online-learning-exploration-policies-thompson-sampling","Thompson Sampling",[957,1475,1476,1477,1480,1481,1484,1485,1487],{},"Thompson Sampling ",[961,1478],{"n":1479},"37",":blog-cite maintains a posterior over each arm's reward parameter and at each step draws a sample, choosing the arm whose sampled reward is highest. For Bernoulli rewards (purchase \u002F no purchase), the Beta(α_k, β_k) posterior is conjugate and the algorithm is two lines of code. Agrawal & Goyal ",[961,1482],{"n":1483},"3"," establish O(√(KT log T)) regret bounds; in practice Thompson Sampling outperforms UCB on most benchmarks because its randomized exploration is naturally tempered by the posterior's shape: an arm with a tight, high posterior gets drawn often; an arm with a wide posterior gets explored. For pricing specifically, Misra et al. ",[961,1486],{"n":1184}," report a field-experiment gross-profit lift of 86% versus a price-discrimination heuristic, and recover more than 80% of the oracle policy's profit within roughly two months. Figure 4 shows the same qualitative behavior in a deliberately simple simulated marketplace.",[957,1489,1490],{},[1054,1491,1492],{},"Figure 4. Cumulative revenue under three pricing policies on a simulated marketplace (n = 80 replications). Thompson Sampling closes most of the gap to the oracle policy within roughly 200 customers; ε-greedy lags by about half a posterior; a fixed price leaves a quantifiable gap that grows linearly in T.",[1058,1494,1495,1513],{},[1061,1496,1497],{},[1064,1498,1499,1502,1505,1507,1510],{},[1067,1500,1501],{},"Arrivals",[1067,1503,1504],{},"Thompson sampling",[1067,1506,1455],{},[1067,1508,1509],{},"Fixed price",[1067,1511,1512],{},"Oracle",[1077,1514,1515,1528,1545,1562,1579,1596,1613,1630,1647],{},[1064,1516,1517,1520,1522,1524,1526],{},[1082,1518,1519],{},"0",[1082,1521,1519],{},[1082,1523,1519],{},[1082,1525,1519],{},[1082,1527,1519],{},[1064,1529,1530,1533,1536,1539,1542],{},[1082,1531,1532],{},"50",[1082,1534,1535],{},"470",[1082,1537,1538],{},"410",[1082,1540,1541],{},"380",[1082,1543,1544],{},"520",[1064,1546,1547,1550,1553,1556,1559],{},[1082,1548,1549],{},"100",[1082,1551,1552],{},"940",[1082,1554,1555],{},"820",[1082,1557,1558],{},"670",[1082,1560,1561],{},"1030",[1064,1563,1564,1567,1570,1573,1576],{},[1082,1565,1566],{},"150",[1082,1568,1569],{},"1420",[1082,1571,1572],{},"1260",[1082,1574,1575],{},"1010",[1082,1577,1578],{},"1570",[1064,1580,1581,1584,1587,1590,1593],{},[1082,1582,1583],{},"200",[1082,1585,1586],{},"1890",[1082,1588,1589],{},"1710",[1082,1591,1592],{},"1360",[1082,1594,1595],{},"2070",[1064,1597,1598,1601,1604,1607,1610],{},[1082,1599,1600],{},"250",[1082,1602,1603],{},"2360",[1082,1605,1606],{},"2160",[1082,1608,1609],{},"1690",[1082,1611,1612],{},"2580",[1064,1614,1615,1618,1621,1624,1627],{},[1082,1616,1617],{},"300",[1082,1619,1620],{},"2850",[1082,1622,1623],{},"2620",[1082,1625,1626],{},"2040",[1082,1628,1629],{},"3110",[1064,1631,1632,1635,1638,1641,1644],{},[1082,1633,1634],{},"350",[1082,1636,1637],{},"3350",[1082,1639,1640],{},"3070",[1082,1642,1643],{},"2390",[1082,1645,1646],{},"3620",[1064,1648,1649,1652,1655,1658,1661],{},[1082,1650,1651],{},"400",[1082,1653,1654],{},"3860",[1082,1656,1657],{},"3480",[1082,1659,1660],{},"2740",[1082,1662,1663],{},"4120",[957,1665,1666],{},[1054,1667,1668],{},"Cumulative revenue, simulated.",[1193,1670,1672],{"id":1671},"_3-pricing-from-heuristics-to-learning-systems-_33-online-learning-exploration-policies-contextual-bandits","Contextual bandits",[957,1674,1675,1676,1679],{},"If the optimal price depends on the customer (returning vs. new), the channel (organic vs. paid), or the basket composition, the seller is in a contextual bandit. LinUCB ",[961,1677],{"n":1678},"25"," posits a linear reward model r = θ_aᵀ x + ε for each arm a, maintains a ridge-regression estimate of θ_a, and selects the arm with the highest UCB on its predicted reward. Disjoint LinUCB allows arm-specific θ_a; hybrid LinUCB shares parameters across arms. Neural-Thompson and neural-UCB extend the same idea with deep feature representations. Contextual bandits are also the dominant pattern for ranking and personalization at large platforms (see §7), where each arm is a candidate item rather than a candidate price.",[1193,1681,1683],{"id":1682},"_3-pricing-from-heuristics-to-learning-systems-_33-online-learning-exploration-policies-reinforcement-learning","Reinforcement learning",[957,1685,1686],{},"When the seller's actions affect future state—inventory levels, customer lifetime value, brand perception—the bandit framing is incomplete and a Markov decision process is more accurate. The state s_t encodes inventory, demand history, and competitor state; the action a_t is a price; the reward r_t = (p_t − c_t) Q_t − h_t I_t includes contribution net of holding cost. Policies are learned by Q-learning, policy gradients, or actor–critic methods. RL is most useful for inventory-coupled pricing (markdown sequences, ride-share surge) and least useful for routine catalog pricing where the bandit framing is sufficient. The operational risk is that off-the-shelf RL agents are notoriously sample-hungry, so most retail RL deployments learn in a high-fidelity simulator and use the bandit framing in production.",[1193,1688,1690],{"id":1689},"_3-pricing-from-heuristics-to-learning-systems-_33-online-learning-exploration-policies-off-policy-evaluation-and-counterfactual-learning","Off-policy evaluation and counterfactual learning",[957,1692,1693,1694,1697,1698,1700],{},"A persistent obstacle to deploying any new policy is that the seller cannot afford to A\u002FB-test every candidate. Off-policy evaluation (OPE) lets the seller estimate, from logged data generated by an old policy, the expected reward of a new policy. Inverse-propensity scoring weights each logged event by the ratio of new-policy to old-policy action probabilities; doubly-robust estimators ",[961,1695],{"n":1696},"14"," combine IPS with an outcome model and remain consistent if either is correct; balanced policy evaluation ",[961,1699],{"n":1323}," reduces variance by reweighting toward covariate balance. For seller teams, OPE is what turns ‘which model is best?' from a question that requires months of experimentation into a question answerable on logged data overnight.",[957,1702,1703],{},"The exploration policies above sit on top of the demand models of §3.1 and feed the optimizers of §3.2. The integration choice—what model is updated at what cadence with what data—is more important than the choice of algorithm: a daily Thompson-Sampling run on a hierarchical demand model with 24-hour observation lag is, in our experience, more reliable than a real-time contextual bandit on a black-box predictor.",[954,1705,865],{"id":864},[957,1707,1708,1709,1711],{},"Demand forecasting feeds inventory, pricing, promotion, and capacity decisions, so the marginal value of accuracy compounds across the stack. Forecasting accuracy is also one of the few metrics whose lift is empirically additive: a 10% reduction in MAPE typically translates into a 1–3% reduction in inventory carrying cost at constant service level ",[961,1710],{"n":1351},", because the safety stock that protects against forecast error scales with the forecast standard deviation. Selection across forecasting methods comes down to three questions: how dense is the per-SKU history, how stationary is the underlying process, and does the downstream task need a point estimate or the full predictive distribution?",[1187,1713,869],{"id":868},[1193,1715,1717],{"id":1716},"_4-forecasting-making-demand-legible-_41-classical-statistical-forecasters-exponential-smoothing-and-the-ets-family","Exponential smoothing and the ETS family",[957,1719,1720],{},"Exponential smoothing decomposes a series into level, trend, and seasonal components, each updated by an exponentially weighted moving average of past observations. The state-space ETS class formalizes this with explicit Error, Trend, Seasonality components in additive or multiplicative combinations, providing a likelihood for parameter estimation and a closed form for prediction intervals. ETS is robust, fast, and routinely beats more complex methods on dense, well-behaved series; it remains the right baseline for any single-SKU forecast and is the default in commercial planning software.",[1193,1722,1724],{"id":1723},"_4-forecasting-making-demand-legible-_41-classical-statistical-forecasters-arima-and-sarimax","ARIMA and SARIMAX",[957,1726,1727],{},"ARIMA(p, d, q) models a series as a function of its own lags and lagged shocks after differencing. The seasonal extension SARIMAX adds seasonal differencing and exogenous regressors. ARIMA is the right tool when the series exhibits clean autoregressive structure and few external drivers; in retail it is more useful at the category aggregate than at the SKU level, because individual SKUs are too sparse and too promotion-driven for the ARIMA family's stationarity assumptions to hold.",[1193,1729,1731],{"id":1730},"_4-forecasting-making-demand-legible-_41-classical-statistical-forecasters-bayesian-structural-time-series-bsts-and-prophet","Bayesian Structural Time Series (BSTS) and Prophet",[957,1733,1734,1735,1738],{},"BSTS expresses a series as y_t = trend_t + season_t + regression_t + ε_t, with each component evolving as a Gaussian state-space process and parameters fitted by MCMC or Kalman smoothing. The major operational benefit is calibrated uncertainty in every component, which lets a planner attribute forecast revisions to specific causes. Prophet ",[961,1736],{"n":1737},"43"," is a deliberately simpler relative—it fits a piecewise-linear trend with Fourier seasonality and a holiday-effect regressor, with priors that make it robust to messy data—and has become a popular default for analyst-facing forecasting.",[1187,1740,872],{"id":871},[957,1742,1743,1744,1747],{},"Retail forecasts live in a natural hierarchy: SKU within category within store within region. Forecasting each level independently produces inconsistent numbers (the SKU forecasts will not sum to the category forecast). Reconciliation methods enforce consistency. Bottom-up sums SKU forecasts; top-down disaggregates the aggregate forecast by historical proportions; the optimal MinT reconciliation ",[961,1745],{"n":1746},"19"," is a generalized-least-squares projection that minimizes total trace of the reconciled forecast covariance subject to coherence constraints. MinT-reconciled forecasts deliver lower MAPE at every level than any single forecast, which is why hierarchical reconciliation is now the default for any organization that plans at multiple aggregations.",[1187,1749,875],{"id":874},[1193,1751,1753],{"id":1752},"_4-forecasting-making-demand-legible-_43-machine-learning-forecasters-gradient-boosted-trees-with-engineered-features","Gradient-boosted trees with engineered features",[957,1755,1756],{},"LightGBM and XGBoost dominate Kaggle retail-forecasting competitions and underpin many production systems. The recipe is to engineer lag features (sales 1, 7, 14, 28 days back), rolling statistics (mean, max, min over moving windows), calendar features (day-of-week, month, holidays), and price\u002Fpromo features, then train a global model on the entire panel with ID embeddings. Tree boosting handles non-linearities and interactions natively, scales to millions of series, and supports quantile regression directly via pinball loss. The main weakness is that the model has no internal notion of time and can extrapolate unevenly outside the training distribution; the standard remedy is to retrain on a rolling window and to monitor extrapolation by feature coverage.",[1193,1758,1760],{"id":1759},"_4-forecasting-making-demand-legible-_43-machine-learning-forecasters-quantile-regression-and-conformal-prediction","Quantile regression and conformal prediction",[957,1762,1763],{},"Inventory does not need a point forecast; it needs a quantile (typically the 90–98th percentile of demand over the lead-time-plus-review interval). Quantile regression—either as a separate model per quantile or as a single multi-quantile network—targets these directly. Conformal prediction wraps any base predictor with a non-parametric calibration step that produces prediction intervals with finite-sample coverage guarantees, and is increasingly the right choice for inventory-grade forecasts where the calibration must be defensible.",[1187,1765,878],{"id":877},[1193,1767,1769],{"id":1768},"_4-forecasting-making-demand-legible-_44-deep-learning-forecasters-deepar","DeepAR",[957,1771,1772,1773,1776],{},"DeepAR ",[961,1774],{"n":1775},"36"," is an autoregressive recurrent network with parameters shared across all series in the panel. At each step the network outputs a distribution over the next value (negative-binomial for counts, Gaussian for continuous), conditioned on the series' lagged values, an embedding vector that identifies the series, and exogenous covariates. The shared parameters let the model transfer information from data-rich SKUs to data-poor ones, which is precisely the regime where SMB sellers operate. Probabilistic outputs are produced by Monte Carlo rollout. Production deployments at AWS, JD, and others have reported MAPE reductions of 10–25% over ETS baselines on retail panels.",[1193,1778,1780],{"id":1779},"_4-forecasting-making-demand-legible-_44-deep-learning-forecasters-temporal-fusion-transformer-tft","Temporal Fusion Transformer (TFT)",[957,1782,1783,1784,1787],{},"TFT ",[961,1785],{"n":1786},"26"," is an attention-based architecture that handles the three covariate types found in retail forecasting: static metadata (category, brand), time-varying known-future inputs (planned promotions, holidays), and time-varying observed inputs (price, weather). Variable-selection networks gate which inputs contribute at each step, and an interpretable multi-head attention block surfaces which past timesteps drive each forecast. TFT typically beats DeepAR on covariate-rich problems and is the right choice when explainability of the forecast is part of the production requirement.",[1193,1789,1791],{"id":1790},"_4-forecasting-making-demand-legible-_44-deep-learning-forecasters-n-beats-and-n-hits","N-BEATS and N-HiTS",[957,1793,1794,1795,1798],{},"N-BEATS ",[961,1796],{"n":1797},"31"," is a stack of fully-connected residual blocks that decomposes the series into interpretable basis functions (trend, seasonality) without recurrence or convolution. The architecture is simple, trains fast, and competitive on benchmarks like M4. N-HiTS adds multi-rate sampling for long-horizon forecasting. Both are useful when the panel is too small to support a transformer but too large to forecast series-by-series.",[1193,1800,1802],{"id":1801},"_4-forecasting-making-demand-legible-_44-deep-learning-forecasters-forecasting-foundation-models","Forecasting foundation models",[957,1804,1805],{},"A new class of pre-trained, zero-shot forecasters has emerged in 2023–24: Chronos, TimesFM, Lag-Llama. They are trained on broad corpora of time series and produce reasonable forecasts on unseen series with no fine-tuning. For SMB sellers, the immediate value is in cold-start: a brand-new SKU with no history can be forecast on day one by analog matching against the foundation model's prior, then refined as data accumulates. The maturity of these models is still uneven and they should be treated as a strong prior, not a final answer—but they have already become the default cold-start tool inside the larger demand-planning vendors.",[1187,1807,881],{"id":880},[957,1809,1810],{},[1054,1811,1812],{},"Table 2. Forecasting method by problem shape and downstream consumer.",[1058,1814,1815,1828],{},[1061,1816,1817],{},[1064,1818,1819,1822,1825],{},[1067,1820,1821],{},"Situation",[1067,1823,1824],{},"Recommended forecaster",[1067,1826,1827],{},"Why",[1077,1829,1830,1841,1852,1863,1874,1885],{},[1064,1831,1832,1835,1838],{},[1082,1833,1834],{},"Long, dense, single SKU",[1082,1836,1837],{},"ETS or SARIMAX",[1082,1839,1840],{},"Strong baseline; clean uncertainty",[1064,1842,1843,1846,1849],{},[1082,1844,1845],{},"Wide catalog, short per-SKU history",[1082,1847,1848],{},"DeepAR \u002F TFT, hierarchical Bayesian",[1082,1850,1851],{},"Information sharing across SKUs",[1064,1853,1854,1857,1860],{},[1082,1855,1856],{},"Heavy promotion-driven demand",[1082,1858,1859],{},"GBM with covariates, or TFT",[1082,1861,1862],{},"Handles non-linear price\u002Fpromo interactions",[1064,1864,1865,1868,1871],{},[1082,1866,1867],{},"Cold-start \u002F new SKU",[1082,1869,1870],{},"Foundation model + analog matching",[1082,1872,1873],{},"Useful prior with no in-series data",[1064,1875,1876,1879,1882],{},[1082,1877,1878],{},"Inventory-grade quantiles",[1082,1880,1881],{},"Quantile GBM, BSTS, conformal wrapper",[1082,1883,1884],{},"Calibrated tails matter for safety stock",[1064,1886,1887,1890,1893],{},[1082,1888,1889],{},"Multi-level planning",[1082,1891,1892],{},"Reconcile via MinT",[1082,1894,1895],{},"Forces coherence across aggregations",[954,1897,884],{"id":883},[957,1899,1900],{},"Inventory translates a demand distribution into orders. The core idea is the newsvendor: in a single period with stochastic demand D ∼ F, per-unit overage cost c_o (= cost minus salvage), and per-unit underage cost c_u (= price minus cost), the expected-profit-maximizing order quantity Q* satisfies the critical-fractile equation:",[1372,1902,1903],{"language":1374},[1376,1904,1905],{"class":1378},"F(Q*) = c_u \u002F (c_u + c_o)",[957,1907,1908],{},"The derivation is short. Expected profit is E[π(Q)] = c_u · E[min(D, Q)] − c_o · E[(Q − D)⁺]; differentiating with respect to Q gives c_u (1 − F(Q)) − c_o F(Q) = 0, which rearranges to the critical fractile. The strength of the framing is that it converts a strategic question (how cautious should I be?) into a parameter (the ratio c_u \u002F (c_u + c_o)) that has a defensible value as soon as the unit economics are written down.",[1022,1910],{"kind":1911,"title":1912,"labels":1913,"values":1922,"highlight":1202,"xLabel":1938,"yLabel":1939,"caption":1940},"line","Figure 5. Newsvendor expected profit as a function of order quantity. The optimum corresponds to the critical fractile c_u \u002F (c_u + c_o); profit is locally flat near Q*, which is operationally valuable because it means small mis-estimates of demand variance are not catastrophic.",[1532,1914,1549,1915,1566,1916,1917,1918,1600,1919,1617,1920,1634,1921,1651],"75","125","175","207","225","275","325","375",[1923,1924,1925,1926,1927,1928,1929,1930,1931,1932,1933,1934,1935,1936,1937],600,930,1210,1480,1680,1820,1880,1860,1760,1590,1370,1130,890,630,400,"Order quantity","Expected profit","Profit peaks at an order quantity of 207.",[1187,1942,888],{"id":887},[1193,1944,1946],{"id":1945},"_5-inventory-matching-supply-to-demand-under-uncertainty-_51-multi-period-and-continuous-review-variants-s-s-policies","(s, S) policies",[957,1948,1949,1950,1952],{},"When ordering incurs a fixed cost K in addition to per-unit cost, the optimal policy is of the (s, S) form: order up to S whenever inventory drops below s, and otherwise do nothing. The optimality is established under stationary demand and was generalized by Scarf and others to broader classes ",[961,1951],{"n":1351},". In practice, s is determined by service-level targets (typically the demand-over-lead-time α-quantile), and S − s trades off ordering cost against holding cost.",[1193,1954,1956],{"id":1955},"_5-inventory-matching-supply-to-demand-under-uncertainty-_51-multi-period-and-continuous-review-variants-base-stock-policies","Base-stock policies",[957,1958,1959],{},"When the fixed ordering cost is negligible (typical for digital purchase orders to a single vendor), the optimal policy collapses to a base-stock rule: continuously raise inventory to a fixed target. The base-stock target is the critical fractile of the lead-time demand distribution, which is the multi-period generalization of the newsvendor.",[1193,1961,1963],{"id":1962},"_5-inventory-matching-supply-to-demand-under-uncertainty-_51-multi-period-and-continuous-review-variants-multi-echelon-optimization","Multi-echelon optimization",[957,1965,1966,1967,1970],{},"Real retailers hold inventory across DCs, regional warehouses, and stores. Local optimization is suboptimal because it double-counts safety stock. Clark and Scarf ",[961,1968],{"n":1969},"52"," established the optimality of echelon base-stock policies for serial supply chains, and the modern multi-echelon literature extends this to assembly and distribution systems. The structural advantage of platform-scale retailers is precisely that they coordinate inventory across the network rather than at each node, which is why same-day delivery is feasible without holding store-level worst-case stock for every SKU.",[1193,1972,1974],{"id":1973},"_5-inventory-matching-supply-to-demand-under-uncertainty-_51-multi-period-and-continuous-review-variants-joint-replenishment","Joint replenishment",[957,1976,1977],{},"When several SKUs share a vendor and a fixed ordering cost (case-pack purchase from a single supplier), the joint replenishment problem chooses a common ordering frequency and SKU-specific multipliers. Closed-form solutions exist for special cases; in general the problem is solved by a Lagrangian or by enumeration over a small set of candidate frequencies. SMB sellers buying from a single overseas supplier are an obvious application.",[1193,1979,1981],{"id":1980},"_5-inventory-matching-supply-to-demand-under-uncertainty-_51-multi-period-and-continuous-review-variants-distributionally-robust-inventory","Distributionally robust inventory",[957,1983,1984,1985,1988,1989,1991],{},"Newsvendor solutions are sensitive to demand-distribution misspecification. Scarf's min-max ",[961,1986],{"n":1987},"51"," is the classical worst-case-distribution solution given only the mean and variance of demand. Modern Wasserstein-based DRO ",[961,1990],{"n":1423}," replaces this with a tractable convex program over an ambiguity set centered on the empirical distribution, with the ambiguity radius tuned by cross-validation. DRO is attractive when forecast errors are heavy-tailed or when the seller's history covers an unusual macro regime.",[1193,1993,1995],{"id":1994},"_5-inventory-matching-supply-to-demand-under-uncertainty-_51-multi-period-and-continuous-review-variants-rl-based-inventory","RL-based inventory",[957,1997,1998],{},"When demand is non-stationary, lead times are stochastic, and the network is multi-echelon, deep RL can learn policies that beat analytic baselines. Production deployments include sub-systems of large fulfillment networks. The pattern is similar to RL pricing: train in a high-fidelity simulator that has been calibrated against real demand, deploy a constrained policy that always respects safety-stock floors, and instrument continuous off-policy evaluation against the analytic baseline.",[1187,2000,891],{"id":890},[957,2002,2003,2004,2006],{},"End-of-season inventory liquidation is a joint pricing-inventory problem: the seller must draw inventory toward zero by season end while maximizing residual revenue. Smith and Achabal ",[961,2005],{"n":1437}," formalize this as an optimal-control problem in which price decreases over time as the urgency of selling rises, and derive structural properties of the optimal markdown trajectory. Modern implementations layer demand-forecasting and bandit-based exploration on top of the same backbone, and treat the markdown schedule as a sequence of small price experiments that update the seller's elasticity estimate as the season progresses.",[954,2008,894],{"id":893},[957,2010,2011],{},"Promotion decisions are intrinsically causal. The relevant question is not ‘did customers who received a coupon buy more?' but ‘would they have bought without it?'. Conventional propensity-weighted regression conflates the two, and as a result chronically over-credits promotions to customers who would have purchased anyway. The relevant statistical object is the conditional average treatment effect (CATE):",[1372,2013,2014],{"language":1374},[1376,2015,2016],{"class":1378},"τ(x, w) = E[Y(w) − Y(0) | X = x]",[957,2018,2019],{},"Estimating τ is hard because, by the fundamental problem of causal inference, we never observe both Y(w) and Y(0) for the same unit. Identification requires either randomized assignment of w (an experiment) or an unconfoundedness assumption that (Y(0), Y(w)) ⊥ w | X. Most production uplift modeling rests on the second route, so the choice of features X is part of the modeling decision, not a preprocessing step.",[1187,2021,898],{"id":897},[1193,2023,2025],{"id":2024},"_6-promotions-and-discounting-where-causality-matters-_61-meta-learners-for-cate-s-learner","S-learner",[957,2027,2028],{},"The S-learner fits a single model μ̂(x, w) of the outcome on covariates and treatment, then estimates τ̂_S(x) = μ̂(x, 1) − μ̂(x, 0). The simplicity is its strength and its weakness: regularizing the base learner shrinks both the prognostic and treatment-effect signals, and the latter is usually much smaller, so the S-learner is biased toward zero treatment effect when the prognostic signal is large. It remains a sensible default for small samples and weak treatments.",[1193,2030,2032],{"id":2031},"_6-promotions-and-discounting-where-causality-matters-_61-meta-learners-for-cate-t-learner","T-learner",[957,2034,2035],{},"The T-learner fits two models, one per arm: μ̂_w(x) on the subsample with treatment w, then τ̂_T(x) = μ̂_1(x) − μ̂_0(x). The T-learner is flexible but inherits high variance in the smaller arm, which in retail is usually the treated arm because the seller does not give discounts to everyone. It is also at risk of differential overfitting between arms.",[1193,2037,2039],{"id":2038},"_6-promotions-and-discounting-where-causality-matters-_61-meta-learners-for-cate-x-learner","X-learner",[957,2041,2042,2043,2046],{},"Künzel et al. ",[961,2044],{"n":2045},"23"," propose the X-learner: fit T-learner-style outcome models, then impute counterfactual treatment effects D_i for each unit using the opposite arm's model, and finally regress D_i on covariates separately within each arm to produce arm-specific CATE estimates. A propensity-weighted combination of the two estimates yields the final CATE. The X-learner outperforms S- and T-learners when arm sizes are imbalanced, which is the typical retail regime.",[1193,2048,2050],{"id":2049},"_6-promotions-and-discounting-where-causality-matters-_61-meta-learners-for-cate-r-learner","R-learner",[957,2052,2053,2054,2056],{},"Nie & Wager ",[961,2055],{"n":1337}," formulate CATE estimation as residualized regression: fit nuisance models for the conditional outcome m̂(x) = E[Y|X = x] and propensity ê(x) = P(W = 1|X = x), then minimize Σ ((Y − m̂(X)) − (W − ê(X)) τ(X))² in a flexible class. The result has quasi-oracle properties: the CATE estimator behaves as if the nuisance functions were known, provided they are estimated at sufficient rates. The R-learner has become a strong default in modern uplift work because the residualization decouples nuisance estimation error from CATE estimation.",[1193,2058,2060],{"id":2059},"_6-promotions-and-discounting-where-causality-matters-_61-meta-learners-for-cate-doubly-robust-learners-aipw-dr-learner","Doubly robust learners (AIPW, DR-learner)",[957,2062,2063],{},"Augmented inverse-propensity weighting (AIPW) constructs a pseudo-outcome that is consistent for the CATE if either the outcome model or the propensity model is correctly specified. The DR-learner regresses this pseudo-outcome on covariates to get an explicit CATE function. Doubly robust estimators are the right starting point in observational settings where neither the outcome nor the propensity model is fully trusted.",[1193,2065,2067],{"id":2066},"_6-promotions-and-discounting-where-causality-matters-_61-meta-learners-for-cate-causal-forests-and-generalized-random-forests","Causal forests and generalized random forests",[957,2069,2070,2071,2074],{},"Wager & Athey ",[961,2072],{"n":2073},"46"," adapt random forests to CATE estimation by enforcing honesty (using disjoint subsamples for split selection and leaf estimation) and recasting splits to maximize heterogeneity in treatment effects rather than outcome variance. Causal forests yield asymptotically normal CATE estimates with valid confidence intervals, and they handle high-dimensional X without manual feature selection. The implementation in EconML is a practical default for tabular retail data.",[1193,2076,2078],{"id":2077},"_6-promotions-and-discounting-where-causality-matters-_61-meta-learners-for-cate-deep-cate-estimators","Deep CATE estimators",[957,2080,2081],{},"When covariates include dense embeddings (session features, product images, text), neural CATE estimators—TARNet, CFRNet, Dragonnet—are appropriate. These typically share a representation across arms and add arm-specific heads, with regularization to limit treatment-imbalance bias in the shared representation. They require larger samples than tree-based methods and are most useful in personalization-adjacent settings where the relevant features are not naturally tabular.",[1187,2083,901],{"id":900},[957,2085,2086],{},"Given a CATE estimator and a budget, the assignment problem is a 0\u002F1 knapsack:",[977,2088],{"text":2089},"max  Σ_i τ̂(x_i, w_i) · m       s.t.   Σ_i cost(w_i) ≤ B,   w_i ∈ {0, w_1, …, w_K}",[957,2091,2092,2093,2095],{},"with m the contribution margin per unit. At retail scale the problem becomes a mixed-integer program with vendor-funded budgets, pantry-loading constraints, category caps, and customer-frequency caps. The LP relaxation is typically very tight, so a column-generation approach with simple rounding is the production workhorse ",[961,2094],{"n":1412},". The structural insight is that the value of better targeting (a tighter τ̂) is proportional to budget pressure: when the budget is unconstrained, even mediocre targeting captures most of the lift; under tight budgets, the value of accurate uplift modeling rises sharply.",[1187,2097,904],{"id":903},[957,2099,2100],{},"Last-touch attribution remains the default in most marketplace UIs but systematically over-credits bottom-of-funnel channels, leading to under-investment in awareness. Three modern alternatives have stabilized in industry practice. Bayesian media-mix models regress geo-week revenue on geo-week spend across channels with informative priors on saturation and adstock, producing channel-level marginal-ROI curves. Geo-randomized experiments (GeoLift, synthetic controls) randomize spend at the geography level and identify incrementality from the gap between treated and synthetic-control units. Switchback designs alternate the treatment on\u002Foff in time within a single unit to handle two-sided-marketplace interference. The trend is unmistakably toward continuous experimentation rather than periodic measurement studies, which is why the experimentation layer in §8 is non-negotiable.",[954,2102,907],{"id":906},[957,2104,2105,2106,2109],{},"Recommender systems convert browsing into purchases by ranking items per user. Bezos ",[961,2107],{"n":2108},"7"," famously attributed substantial Amazon GMV to recommendations; downstream estimates place direct lift in the 20–35% range for mature implementations. The architectural arc has run from collaborative filtering through matrix factorization to two-tower neural retrieval and multi-task neural ranking. Three structural choices distinguish industrial recommenders: how candidate items are retrieved from a catalog of millions; how those candidates are ranked at low latency; and how the system continues to learn from logged data.",[1187,2111,911],{"id":910},[1193,2113,2115],{"id":2114},"_7-personalization-and-recommendation-_71-retrieval-collaborative-filtering","Collaborative filtering",[957,2117,2118],{},"User-based and item-based collaborative filtering compute similarities (cosine, Pearson) over the user-item interaction matrix, then recommend items similar to those the user has interacted with. CF is the original recommender approach and remains a strong baseline, especially with implicit-feedback corrections. Its weaknesses are scalability (similarity matrices are O(n²)) and cold-start (no recommendations for new users or new items).",[1193,2120,2122],{"id":2121},"_7-personalization-and-recommendation-_71-retrieval-matrix-factorization","Matrix factorization",[957,2124,2125,2126,2129,2130,2133],{},"Matrix factorization ",[961,2127],{"n":2128},"22"," decomposes the user-item rating (or implicit-feedback) matrix into low-rank user and item embeddings U and V, with predicted score ŝ_ = u_uᵀ v_i. For implicit feedback (clicks, purchases, no explicit rating), the standard objective is weighted least squares over observed and (down-weighted) unobserved interactions, solved by alternating least squares (ALS-WR) or stochastic gradient descent. Bayesian Personalized Ranking ",[961,2131],{"n":2132},"33"," replaces this with a pairwise ranking loss that directly optimizes AUC on observed-vs-unobserved pairs. Matrix factorization is still in production at many sellers, particularly as a strong cold-start fallback.",[1193,2135,2137],{"id":2136},"_7-personalization-and-recommendation-_71-retrieval-two-tower-neural-retrieval","Two-tower neural retrieval",[957,2139,2140,2141,2144],{},"Industrial-scale retrieval has converged on the two-tower architecture: a user encoder f(u) and an item encoder g(i) produce embeddings whose dot product is the predicted relevance ",[961,2142],{"n":2143},"49",". The model is trained with sampled softmax over the catalog, with sampling-bias correction to compensate for the fact that popular items appear disproportionately often as negatives. At serving time, item embeddings are pre-computed and indexed with an approximate-nearest-neighbor structure (FAISS, ScaNN), so retrieval is sub-millisecond on catalogs of millions. The two-tower pattern is the dominant industry recipe and has the operational advantage that the user and item towers can be retrained on different cadences.",[1193,2146,2148],{"id":2147},"_7-personalization-and-recommendation-_71-retrieval-graph-neural-networks","Graph neural networks",[957,2150,2151,2152,2155,2156,2159],{},"GNNs exploit the bipartite user-item graph and item-item co-purchase graph by propagating embeddings along edges. PinSage ",[961,2153],{"n":2154},"48"," is the canonical web-scale recipe; LightGCN ",[961,2157],{"n":2158},"18"," strips out non-essential transformation and activation layers and trains a much simpler weighted aggregation, often outperforming heavier architectures. GNNs are particularly useful for cold-start (new items inherit embeddings from their neighbors) and complementary-product recommendation (items frequently co-purchased rather than substituted).",[1187,2161,914],{"id":913},[1193,2163,2165],{"id":2164},"_7-personalization-and-recommendation-_72-ranking-and-re-ranking-pointwise-vs-pairwise-vs-listwise","Pointwise vs. pairwise vs. listwise",[957,2167,2168],{},"Once a few hundred candidate items are retrieved, the ranker reorders them. Pointwise rankers predict per-item scores (logistic regression on engagement); pairwise rankers (RankNet, LambdaRank) optimize ordered pairs; listwise rankers (LambdaMART, ListNet) optimize the full list using metrics like NDCG. Listwise approaches generally win on engagement metrics but are more expensive; LambdaMART remains a workhorse for tabular feature sets.",[1193,2170,2172],{"id":2171},"_7-personalization-and-recommendation-_72-ranking-and-re-ranking-multi-task-neural-rankers","Multi-task neural rankers",[957,2174,2175,2176,2179,2180,2182],{},"Industrial rankers must balance competing objectives—click, add-to-cart, purchase, return rate—because optimizing a single objective leads to clickbait. Multi-task architectures (shared bottom, MMoE, PLE) ",[961,2177],{"n":2178},"12"," maintain shared representations with task-specific heads, and weighted task losses reflect the seller's value function. The Deep Interest Network ",[961,2181],{"n":1532}," adds an attention mechanism over the user's past behaviors that gates which historical signals are relevant for the current candidate, improving CTR meaningfully on large e-commerce platforms.",[1193,2184,2186],{"id":2185},"_7-personalization-and-recommendation-_72-ranking-and-re-ranking-sequence-and-session-models","Sequence and session models",[957,2188,2189,2190,2193,2194,2197],{},"User intent is more legible from a recent sequence of clicks than from an aggregate profile. SASRec ",[961,2191],{"n":2192},"21"," applies self-attention to the user's interaction history; BERT4Rec ",[961,2195],{"n":2196},"55"," replaces the autoregressive objective with a masked-item prediction more analogous to BERT in NLP. Session-based recommenders win in domains with thin user histories and strong intra-session intent (apparel, electronics).",[1187,2199,917],{"id":916},[957,2201,2202,2203,2205,2206,2208],{},"Production recommenders cannot afford to A\u002FB-test every candidate model, and online metrics are noisy. The right architectural pattern is contextual-bandit ranking: each ranking decision logs the action probability, and the system continuously learns from the resulting reward. Top-K off-policy correction ",[961,2204],{"n":1171}," adapts policy-gradient training to the multi-action setting required for a list of recommendations. Counterfactual evaluation ",[961,2207],{"n":1323},":blog-cite lets teams compare candidate models on logged data, which is essential because the catalog and user distribution drift on weekly time scales. Mature recommenders treat exploration as a first-class concern—not as a randomization gimmick, but as the data-collection mechanism that keeps the model identified.",[954,2210,920],{"id":919},[957,2212,2213],{},"The methods above are necessary but not sufficient. The binding constraint for most sellers is the absence of a system that ingests their data, runs the relevant models, and surfaces actionable decisions. Figure 6 sketches a five-layer reference architecture that we treat as the minimum viable stack for a seller crossing into the ‘growing' or ‘scaled' regime of Table 3.",[957,2215,2216],{},[1054,2217,2218],{},"Figure 6. Reference architecture for a seller-side decision-intelligence stack.",[2220,2221,2222,2230,2236,2242],"ol",{},[2223,2224,2225,2229],"li",{},[2226,2227,2228],"strong",{},"Inputs:"," order and clickstream, inventory and warehouse, ad spend, market prices",[2223,2231,2232,2235],{},[2226,2233,2234],{},"Shared data layer:"," feature store, event log, customer and SKU embeddings",[2223,2237,2238,2241],{},[2226,2239,2240],{},"Models:"," demand forecaster, price elasticity and bandit, promo and uplift model, inventory optimizer",[2223,2243,2244,2247],{},[2226,2245,2246],{},"Decision:"," optimization engine (LP \u002F MIP, constraint solver), then human review",[1187,2249,924],{"id":923},[957,2251,2252],{},"Connectors to marketplaces (Amazon SP-API, Shopify GraphQL, Walmart Marketplace, eBay Trading API), the seller's warehouse-management system, ad platforms (Meta Marketing API, Google Ads, Pinterest), and competitive scrapers. The output is an event log with consistent SKU, customer, and time keys, partitioned and stored in a columnar format (Parquet on object storage, or a warehouse like Snowflake or BigQuery). Without this layer, every downstream layer regresses to spreadsheet quality. The hardest single problem at this layer is identity resolution across marketplaces, which determines whether ‘the customer' is a usable abstraction.",[1187,2254,927],{"id":926},[957,2256,2257],{},"A small but disciplined feature store containing SKU embeddings (text via a pre-trained encoder, image via a vision encoder, taxonomy via learned look-ups), customer recency-frequency-monetary features, price and competitor-price histories, and seasonal and holiday calendars. Each feature has an explicit freshness budget (intraday for pricing; daily for forecasting; weekly for cohort features) and an SLA for serving latency. The feature store is what makes models reproducible and what lets the same feature vector flow into pricing, forecasting, and ranking.",[1187,2259,930],{"id":929},[957,2261,2262],{},"A demand forecaster (hierarchical Bayesian or DeepAR\u002FTFT for the catalog), an elasticity estimator with bandit-driven exploration for pricing, an uplift model (R-learner or causal forest) for promotions, an inventory optimizer that consumes the forecaster's predictive distribution, and a two-tower ranker for personalization. Crucially, models are decoupled from decisions: the same forecaster feeds both pricing and inventory, the same SKU embeddings feed retrieval and uplift, and the model registry is the unit of versioning rather than the decision endpoint.",[1187,2264,933],{"id":932},[957,2266,2267],{},"A linear- or mixed-integer-programming solver translates model outputs into decisions, subject to business constraints (margin floors, MAP, brand consistency, vendor budgets). Decision proposals are not auto-executed: a thin review UI exposes the recommendations, the model evidence, and the constraint shadow prices to a human merchandiser. The shadow prices in particular are commercially valuable—they tell the seller how much margin is left on the table because of a binding constraint, which is exactly the conversation a merchandiser should be having with a vendor or a category manager.",[1187,2269,936],{"id":935},[957,2271,2272,2273,2275],{},"A switchback or geo-randomized experimentation framework wraps every decision so that the marginal contribution of each model is observable. Off-policy evaluation lets the team rank candidate models on logged data before promoting any of them to a live A\u002FB, which is essential when the seller's traffic is too thin to support many simultaneous experiments. This layer is what converts the architecture from a one-off consulting deliverable into a continuously improving asset ",[961,2274],{"n":1252},", and it is the layer where SMB sellers most consistently under-invest.",[954,2277,939],{"id":938},[957,2279,2280],{},"Capability should track the seller's data volume, decision frequency, and operational maturity. Table 3 maps a pragmatic technique progression that we have found durable across categories.",[957,2282,2283],{},[1054,2284,2285],{},"Table 3. Recommended technique progression by seller stage.",[1058,2287,2288,2307],{},[1061,2289,2290],{},[1064,2291,2292,2295,2297,2300,2302,2305],{},[1067,2293,2294],{},"Seller stage",[1067,2296,27],{},[1067,2298,2299],{},"Forecasting",[1067,2301,1116],{},[1067,2303,2304],{},"Promotion",[1067,2306,1127],{},[1077,2308,2309,2329,2349],{},[1064,2310,2311,2314,2317,2320,2323,2326],{},[1082,2312,2313],{},"Early (≤ $1M GMV)",[1082,2315,2316],{},"Lerner with hand-set elasticity; rule-based repricer with margin floor",[1082,2318,2319],{},"ETS \u002F Holt-Winters per SKU; manual seasonal overrides",[1082,2321,2322],{},"Newsvendor with empirical CDF; service-level rule of thumb",[1082,2324,2325],{},"Last-click + simple S-learner uplift",[1082,2327,2328],{},"Bestsellers, simple item-CF",[1064,2330,2331,2334,2337,2340,2343,2346],{},[1082,2332,2333],{},"Growing ($1M–$50M)",[1082,2335,2336],{},"Bandit on price points; LP for category",[1082,2338,2339],{},"Hierarchical Bayesian; GBM with covariates and conformal quantiles",[1082,2341,2342],{},"(s, S) per SKU with computed safety stock",[1082,2344,2345],{},"T- or X-learner uplift; LP-based budget assignment",[1082,2347,2348],{},"Two-tower retrieval; pointwise ranker; off-policy eval",[1064,2350,2351,2354,2357,2360,2363,2366],{},[1082,2352,2353],{},"Scaled ($50M+)",[1082,2355,2356],{},"Contextual bandits \u002F RL; MIP under cannibalization",[1082,2358,2359],{},"DeepAR \u002F TFT with covariates; MinT-reconciled",[1082,2361,2362],{},"Multi-echelon; distributionally robust hedging",[1082,2364,2365],{},"Causal forest \u002F R-learner; MIP with vendor budgets",[1082,2367,2368],{},"Multi-task neural ranker; bandit exploration; sequence model",[954,2370,942],{"id":941},[1193,2372,2374],{"id":2373},"_10-risks-governance-and-considerations-pricing-fairness-and-regulation","Pricing fairness and regulation",[957,2376,2377],{},"Dynamic and personalized pricing have a poor reputation for reasons that are partly justified: surge pricing on essentials, opaque discrimination across geographies, and the perception of being ‘gouged' all impose long-run brand cost. Regulators in the EU (the Omnibus Directive) and several U.S. states now require disclosure of personalized pricing, and the Federal Trade Commission has signaled active interest in price discrimination by digital intermediaries. Sellers should prefer context-level optimization (channel, time-of-day, basket composition) over identity-level optimization until the legal landscape stabilizes, and should design their pricing system so that the price actually shown to a customer is a deterministic function of disclosable inputs.",[1193,2379,2381],{"id":2380},"_10-risks-governance-and-considerations-cold-start","Cold start",[957,2383,2384],{},"New SKUs and new customers have no per-entity history. Hierarchical Bayes (§3.1, §4.2), embedding-based analog matching, and forecasting foundation models (§4.4) are the right architectural responses. The wrong response is to wait until enough data accumulates—by then the cold-start window has closed and the seller has ceded margin to whichever competitor was willing to make a guess.",[1193,2386,2388],{"id":2387},"_10-risks-governance-and-considerations-data-drift-and-silent-failure","Data drift and silent failure",[957,2390,2391],{},"Promo cycles, macro shocks, and platform algorithm changes invalidate models built on stationary assumptions. The mitigation is continuous evaluation: holdout reservoirs that the production model never sees, conformal prediction intervals that flag unusual coverage breaches, and KS-style monitoring of feature distributions against training-set baselines. A model that has not been re-evaluated in 90 days should be treated as suspect by default.",[1193,2393,2395],{"id":2394},"_10-risks-governance-and-considerations-algorithmic-monoculture","Algorithmic monoculture",[957,2397,2398],{},"When every seller deploys the same repricer or recommender, equilibrium prices and assortments collapse to a narrow strip and the platform's overall consumer surplus falls. The platforms that actively allow multiple algorithms to compete (and that publish realistic counterfactual data so vendors can train differentiated models) tend to retain healthier ecosystems. From the seller's side, monoculture is a reason to instrument differentiation against the platform default, not a reason to abandon optimization.",[1193,2400,2402],{"id":2401},"_10-risks-governance-and-considerations-human-in-the-loop-and-auditability","Human-in-the-loop and auditability",[957,2404,2405],{},"Auto-execution amplifies errors. The stable production pattern is decision proposals reviewed before they hit the catalog, with audit logs for every override and for every model decision (input features, model version, output, applied constraints). Auditability is also increasingly a regulatory requirement and is the difference between a defensible system and one that simply works most of the time.",[954,2407,945],{"id":944},[957,2409,2410],{},"Large retailers have, for a decade, used the techniques surveyed in this report to move pricing, inventory, promotions, and personalization from intuition to inference. The techniques themselves are no longer the moat. The moat is integration: connectors, feature stores, optimization, experimentation, off-policy evaluation, and a human-in-the-loop UI assembled into a single workflow. For an SMB seller, the highest-return investment is no longer hiring a single data scientist; it is adopting—or, where the right product does not yet exist, building—an integrated decision stack of which that data scientist would be one component. Closing this gap is the most concrete way to move e-commerce from a winner-take-most market toward one in which a much wider set of merchants can earn fair returns on the capital and labor they invest.",[977,2412],{"text":2413},"The techniques are no longer the moat. Integration is the moat.",[954,2415,948],{"id":947},[957,2417,2418],{},"Six questions sellers, operators, and engineers ask most often when deciding whether — and how — to adopt the techniques surveyed in this report.",[2420,2421,2423],"blog-question",{"question":2422},"What is decision intelligence in e-commerce?","Decision intelligence in e-commerce is the discipline of turning marketplace and operating data into ranked, auditable next-actions across pricing, demand forecasting, inventory, promotions, and personalization — and integrating those decisions into a single workflow rather than handing them off as charts. Mathematically it sits at the intersection of statistical learning, causal inference, optimization under uncertainty, and reinforcement learning; operationally it sits behind a feature store, an experimentation system, and a human-in-the-loop UI. The bottleneck for most sellers in 2026 is no longer choosing between algorithms — those are documented in textbooks and shipped in open-source software. It is the integration layer that turns those algorithms into a decision a merchandiser will actually approve.",[2420,2425,2427,2428,2431],{"question":2426},"How does dynamic pricing actually work for marketplace sellers?","Production dynamic pricing combines three pieces. First, an elasticity-aware demand model — typically a hierarchical Bayesian model or a structural choice model — that yields per-SKU expected demand at candidate prices. Second, an optimization layer that picks the price that maximises expected contribution after returns, ad cost, and platform fees, subject to margin floors, competitor-aware bounds, and category guardrails. Third, an experimentation system (A\u002FB or off-policy evaluation) that closes the learning loop so the elasticity estimates do not drift. Repricers without an elasticity model behave like rule engines and tend to chase the floor; the literature is clear that elasticity-aware pricing recovers materially more contribution at scale ",[961,2429],{"n":2430},"16",":blog-cite:blog-cite.",[2420,2433,2435],{"question":2434},"Which forecasting methods do platform-scale retailers actually use?","The honest answer: a combination, not a single winner. Hierarchical reconciliation is used to keep store\u002Fregion\u002FSKU forecasts coherent. Intermittent-demand methods (Croston, ADIDA, TSB) handle the long tail of slow-moving SKUs. Deep methods — DeepAR, N-BEATS, Temporal Fusion Transformers — earn their keep on dense, regular series and on cross-sectional learning across many similar SKUs. Quantile forecasts, not point forecasts, feed downstream inventory and promotion decisions because the tail risk is what costs money. Stacking and exogenous-feature ingestion (price, promotion, holiday, weather) are standard. The differentiator is rarely the model class — it is the data hygiene and the quantile-aware optimization on top.",[2420,2437,2439],{"question":2438},"Should a marketplace seller use the newsvendor model or a multi-echelon model?","The newsvendor model is the right starting point: a single SKU, a single period, a known demand distribution, an explicit critical fractile that trades off underage cost (lost sale, lost margin, lost rank) against overage cost (carrying cost, markdown, storage). It is also where most production systems still live, augmented with quantile demand forecasts and stochastic lead times. Multi-echelon models matter when inventory is held in more than one location and shipments cascade — typical of marketplace sellers running fulfilment-by-marketplace plus their own warehouse. In that case Clark–Scarf base-stock policies, or their robust-optimization variants under demand ambiguity, replace single-echelon newsvendor logic. Both are cheap to implement; the gain comes from honest service-level targets, not exotic policy structure.",[2420,2441,2443,2444,2431],{"question":2442},"Why do off-policy evaluation and uplift modeling matter more than another A\u002FB test?","Marketplace sellers run far more decisions than they can A\u002FB test. Every campaign, every voucher, every category-level price move would, in a pure-A\u002FB world, need its own randomised exposure — and platforms increasingly do not allow it at the granularity sellers need. Uplift models (causal forests, X-learner, R-learner) estimate the per-unit effect of an intervention from observational or partially-randomised data, so promotions can be targeted at customers and SKUs where the lift is highest rather than at the overall average. Off-policy evaluation (doubly robust estimation, IPS, weighted importance sampling) lets the seller score new policies against historical data before deployment. Together they extend the reach of experimentation to the long tail of decisions that A\u002FB testing alone cannot cover ",[961,2445],{"n":2446},"9",[2420,2448,2450],{"question":2449},"What does it actually take to deploy a decision-intelligence stack — and what's the integration moat?","A working deployment needs six components in production: data connectors that ingest order-line, ad-spend, inventory, fee, and return data with reliable joining keys; a feature store that serves the same features to training and inference at SKU and customer grain; a model layer that produces calibrated forecasts, elasticities, uplifts, and personalisation scores; an optimization layer that combines those signals into ranked decisions subject to operating constraints; an experimentation system covering A\u002FB, multi-armed bandits, and off-policy evaluation; and a human-in-the-loop UI that surfaces the math behind every recommendation so the merchandiser can approve, override, or escalate. The moat in 2026 is not which algorithms are used — those are commoditised. It is whether all six components are wired into a single workflow that a merchandiser uses every week.",[954,2452,951],{"id":950},[2454,2455,2456,2457,2462,2469,2473,2478,2483,2490,2491,2492,2493,2494],{"id":842,"depth":313,"text":843},{"id":845,"depth":313,"text":846},{"id":848,"depth":313,"text":849},{"id":851,"depth":313,"text":852,"children":2458},[2459,2460,2461],{"id":855,"depth":308,"text":856},{"id":858,"depth":308,"text":859},{"id":861,"depth":308,"text":862},{"id":864,"depth":313,"text":865,"children":2463},[2464,2465,2466,2467,2468],{"id":868,"depth":308,"text":869},{"id":871,"depth":308,"text":872},{"id":874,"depth":308,"text":875},{"id":877,"depth":308,"text":878},{"id":880,"depth":308,"text":881},{"id":883,"depth":313,"text":884,"children":2470},[2471,2472],{"id":887,"depth":308,"text":888},{"id":890,"depth":308,"text":891},{"id":893,"depth":313,"text":894,"children":2474},[2475,2476,2477],{"id":897,"depth":308,"text":898},{"id":900,"depth":308,"text":901},{"id":903,"depth":308,"text":904},{"id":906,"depth":313,"text":907,"children":2479},[2480,2481,2482],{"id":910,"depth":308,"text":911},{"id":913,"depth":308,"text":914},{"id":916,"depth":308,"text":917},{"id":919,"depth":313,"text":920,"children":2484},[2485,2486,2487,2488,2489],{"id":923,"depth":308,"text":924},{"id":926,"depth":308,"text":927},{"id":929,"depth":308,"text":930},{"id":932,"depth":308,"text":933},{"id":935,"depth":308,"text":936},{"id":938,"depth":313,"text":939},{"id":941,"depth":313,"text":942},{"id":944,"depth":313,"text":945},{"id":947,"depth":313,"text":948},{"id":950,"depth":313,"text":951},{"id":2496,"title":33,"blog":2497,"body":2535,"customers":295,"description":2540,"draft":294,"draftReason":295,"extension":657,"glossary":295,"industry":295,"meta":2541,"navigation":294,"ogImage":2542,"path":34,"research":295,"robots":295,"schemaOrg":295,"seo":2543,"showcase":294,"sitemap":2544,"stem":2545,"__hash__":2546},"pages_en\u002Fblog.md",{"emptyLabel":2498,"labels":2499,"indexLabels":2504,"listenLabels":2515,"indexCta":2528,"ctaButtons":2531},"No posts match yet. Try another category or a shorter search.",{"tldr":2500,"published":2501,"updated":2502,"related":2503},"In short","Published","Updated","Keep reading",{"categories":2505,"sort":2506,"newest":2507,"oldest":2508,"grid":2509,"list":2510,"loadMore":2511,"results":2512,"result":2513,"clearSearch":2514},"Categories","Sort","Newest first","Oldest first","Grid view","List view","Load more","{n} results for “{q}”","1 result for “{q}”","Clear search",{"listen":2516,"loading":2517,"play":2518,"pause":2519,"back":2520,"forward":2521,"speed":2522,"seek":2523,"player":2524,"close":2525,"follow":2526,"of":2527},"Listen to article","Loading…","Play audio of this article","Pause audio","Back 15 seconds","Forward 15 seconds","Playback speed","Seek","Audio player","Pause audio and close player","Follow the text","{a} of {b}",{"headline":2529,"subline":2530},"Turn every post into a decision","See the numbers behind these notes for your own shops, with a demo on your data.",{"primaryCta":2532,"secondaryCta":2533},{"label":57,"to":58},{"label":2534,"to":28},"See pricing",{"type":2536,"value":2537,"toc":2538},"minimark",[],{"title":839,"searchDepth":308,"depth":308,"links":2539},[],"Notes on profit, pricing and ads for sellers on Shopee, Lazada and TikTok Shop, with the numbers behind them.",{},{"component":33},{"title":33,"description":2540},{"loc":34},"blog","VuN4_Knpp3IpkJqH-kRoXag4vgoPV2p71ZFiR-VpgzQ",[2548,2550,2551,2553,2555,2557,2559],{"slug":338,"label":2549},"Profit",{"slug":350,"label":27},{"slug":291,"label":2552},"Ads",{"slug":327,"label":2554},"Operations",{"slug":430,"label":2556},"Market",{"slug":408,"label":2558},"Data Science",{"slug":307,"label":48},{"id":2561,"categories":2562,"domain":2569,"extension":5,"fallback":2570,"home":2573,"industry":2582,"meta":2587,"pricing":2588,"stem":2600,"__hash__":2601},"og_en\u002Fen\u002Fog.yml",[2563,2564,2565,2566,2567,2568],{"slug":338,"label":2549},{"slug":350,"label":27},{"slug":291,"label":2552},{"slug":327,"label":2554},{"slug":408,"label":2558},{"slug":307,"label":48},"dataglasslabs.com",{"title":2571,"description":2572},"DataGlass","The decision engine for Shopee, Lazada and TikTok Shop sellers.",{"chart":2574,"metrics":2575},"Net profit, last 30 days",[2576,2579],{"label":2577,"value":2578},"Net profit","฿1.24M",{"label":2580,"value":2581},"Profit ROAS","3.4x",{"axis":2583,"point":2584},"Ad spend",{"label":2585,"value":2586},"Next baht still pays","+฿0.42",{},{"plans":2589},[2590,2594,2597],{"name":2591,"price":2592,"period":2593},"Free","฿0","\u002Fmonth",{"name":2595,"price":2596,"period":2593},"Standard","฿1,499",{"name":2598,"price":2599,"period":2593},"Pro","฿2,499","en\u002Fog","63L8y4Lv8I2q3NvR26fMXU1sE9Wd7TZaz2l_tV6oZuo",[],1792784876189]