From Gut Feel to Posterior Inference: A Research Article on the DataGlass Decision-Intelligence System for E-Commerce Ad Budget Allocation

Working paper, May 2026. Companion technical note to the peer-reviewed manuscript [1] and the internal model proposal series [2]–[5].

Abstract

This article presents, at a level of detail appropriate for a public-facing research communication, the DataGlass decision-intelligence system for daily advertising-budget allocation on platform-controlled marketplaces such as Shopee, Lazada, TikTok Shop, and Amazon Sponsored Products. We first document the pre-algorithmic era of seller behaviour, in which the dominant practices are trial-and-error bid adjustments, rolling-mean extrapolation of dashboard metrics, and gut-feel reallocations driven by the reported return-on-ad-spend (ROAS) figure rather than the true contribution-margin-adjusted return. We show analytically that under realistic margin, return, and platform-fee assumptions, rolling-mean heuristics on reported ROAS systematically misallocate spend, overweight low-contribution campaigns, and induce excessive bid churn that interacts adversarially with platform learning systems. We then summarise the methodological foundations of the DataGlass system — a Bayesian response model, a constrained-portfolio optimiser solving an equal-marginal-profit condition, and a non-stationarity layer combining Thompson Sampling, CUSUM-based changepoint detection, and three-layer regime detection — and report empirical results from offline backtesting and live A/B deployment showing 18–24% portfolio-profit improvements with reduced reallocation frequency. Proprietary calibration details are deliberately omitted; the article is intended as a rigorous public communication of the system's research foundations rather than a reproduction guide.

1. Introduction

Online retail is now a US$6.3-trillion industry that absorbs roughly one in six retail dollars spent worldwide [6], [22]. Within it, the distribution of analytical capability is sharply bimodal: a small group of platform-scale retailers operates end-to-end decision-intelligence stacks, while the long tail of small and medium marketplace sellers depends on spreadsheets, dashboard widgets, and intuition [7]. The decision-intelligence gap — the difference between optimal and observed seller behaviour on commercial levers such as price, promotion, and ad spend — has been documented quantitatively in survey work and is widely understood to constitute a meaningful share of the contribution lost by SMB sellers each year [7], [10], [17].

Within this gap, advertising budget allocation is a particularly acute problem for three structural reasons. First, on modern marketplace platforms the seller does not bid: the platform does. The seller controls only daily budget and target ROAS, and the platform's auto-bidding algorithm decides keywords, audiences, placement, and pacing [1], [4]. Second, the platform is opaque: the auction mechanism, the matching algorithm, and the relevance score are not disclosed, so the seller observes only an input–output mapping whose internal structure must be inferred from data [21], [25]. Third, the platform actively penalises high-frequency intervention through learning-phase freezes, minimum-change windows, and pacing throttles, so the obvious response — bid harder, retune more often — is operationally counterproductive [1], [3], [5].

This article documents how DataGlass — a production decision-intelligence system that, as of Q1 2026, manages budgets for several hundred active sellers on a major Southeast Asian marketplace — addresses this constraint structure. The contribution of the article itself, as distinct from the underlying methodological work, is twofold:

1. We characterise the pre-algorithmic baseline — the family of manual heuristics that small sellers use today — with enough quantitative precision to make the failure modes legible. In particular, we formalise the gap between reported ROAS and true contribution-margin-adjusted ROAS and show that rolling-mean extrapolation under this gap is an asymptotically biased estimator of marginal profit per dollar.

2. We summarise the methodological core of DataGlass at the level required to evaluate its scientific positioning, while withholding the proprietary calibration that distinguishes the production system from a textbook implementation.

The remainder of the article is organised as follows. Section 2 surveys the historical and current practice of manual ad-budget management. Section 3 derives the analytical reasons that practice fails. Section 4 fixes notation and states the formal optimisation problem. Section 5 reviews the relevant academic literature. Section 6 describes the DataGlass system at a high level. Section 7 discusses positioning and generalisation. Sections 8 and 9 cover limitations and conclusion. Citations follow IEEE numerical style.

2. The Pre-Algorithmic Era: How Sellers Bid Without Optimisation

For most of the post-2015 marketplace era, seller-side ad-budget decisions have been made by humans operating on spreadsheets, dashboards, and intuition. Three families of heuristics dominate.

2.1 Trial-and-error reallocation

The most common pattern observed in the seller surveys synthesised by Jungle Scout [10], Shopify Commerce Trends [17], and the McKinsey SMB index [13] is a weekly or twice-weekly manual sweep: the seller opens the marketplace dashboard, sorts campaigns by reported ROAS, increases the budget on the top-quartile campaigns by a fixed percentage (typically 10–30%), decreases or pauses the bottom quartile, and revisits the next week. There is no model of the budget–response relationship, no quantification of uncertainty, and no separation of signal from noise. The implicit assumption is that yesterday's reported ROAS is a good predictor of tomorrow's marginal ROAS — an assumption that we show in Section 3 fails systematically and asymmetrically.

The operational consequences are well documented. Frequent budget edits trigger platform-side learning-phase resets, during which the auto-bidder collects fresh exploration data and the campaign's effective cost-per-click is elevated [1, §3]. The seller observes degraded performance, reacts with further edits, and a self-reinforcing churn cycle is established. Empirical estimates from internal DataGlass auditing of un-managed seller accounts suggest that 30–55% of all manual budget edits made during a learning phase are net-negative in expected contribution.

2.2 Rolling-mean extrapolation

A more numerate seller will run a rolling mean — typically 7-day or 14-day — of dashboard ROAS or revenue per campaign, and use it as a point estimate for next-day performance. Letting $R_t$ denote the reported ROAS on day $t$ and ${\bar{R}}_t^{(k)}$ the trailing-$k$ mean,

the seller then sets tomorrow's budget proportionally:

where ${\bar{R}}_{\bullet ,t}^{(k)}$ is the portfolio-wide mean. This is the poor man's portfolio optimiser: it reallocates toward campaigns whose recent reported return is above average. It is more disciplined than pure intuition, but it carries every pathology of a moving-average estimator on a non-stationary, censored, attribution-lagged signal — and, as we show below, it is biased in a direction that systematically rewards low-contribution campaigns.

2.3 Gut feel and intuition

The third practice, which often coexists with the first two, is intuition: the seller "knows" that a campaign is doing well because of a recent flurry of orders, or "feels" that a competitor is being aggressive on a particular SKU. There is a substantial behavioural-economics literature documenting that humans systematically over-react to recent reinforcement, under-weight base rates, and confuse variance with signal — see, among others, the survey by Kahneman [12] and the experimental advertising work of Lewis and Rao [14], who showed that the signal-to-noise ratio in observational ad measurement is so unfavourable that confident-sounding causal claims by SMB sellers are typically not supported by their own data.

2.4 The aggregate cost of the pre-algorithmic baseline

The economic implication is direct. SMB sellers spent an estimated US$80–110 billion on marketplace ads in 2024 [22]. Conservatively, if 15–25% of that spend is misallocated by the heuristics in sections 2.1–2.3, the deadweight loss is on the order of US$12–28 billion annually. This is the contestable surface that decision-intelligence systems address.

3. Why Manual Heuristics Fail: The True-ROAS Gap and the Rolling-Mean Bias

We now formalise the two principal failure modes of the pre-algorithmic baseline.

3.1 The reported-vs-true ROAS gap

The dashboard figure displayed to the seller is

This is not a measure of profit. The seller's actual contribution margin from a marginal advertising dollar must net out cost of goods sold, platform commission, payment processing, fulfilment subsidy, and post-sale return loss. Letting $m_g$ denote the gross margin on the basket, $f$ the all-in platform-fee fraction, $r$ the post-sale return rate, and $\rho$ a fulfilment-and-payment overhead fraction, the true profit-adjusted ROAS is approximately

For a typical Southeast Asian apparel seller with $m_g\approx 0.45$, $r\approx 0.18$, $f+\rho \approx 0.17$, a reported ROAS of 5.0× corresponds to a true ROAS of approximately

i.e., a contribution multiplier of 1.53× on advertising spend rather than the headline 5×. A campaign reporting 2.5× ROAS — which most sellers would treat as healthy — has a true contribution multiplier of approximately 0.77×, i.e., it is loss-making on every marginal dollar. The phenomenon is not subtle; it is structural; and it is not visible on any standard marketplace dashboard. See the longer treatment in the true-ROAS reconstruction note for a Shopee-specific worked example, and contribution margin for the canonical definition.

The same arithmetic generalises across categories. For consumer electronics, where $m_g$ is often 0.10–0.20, the break-even reported ROAS is in the 6–10× range. For high-margin beauty SKUs with low return rates, it is closer to 2×. The break-even reported ROAS therefore varies over an order of magnitude across categories, yet sellers routinely apply a single dashboard cutoff (commonly 3× or 4×) across heterogeneous portfolios.

3.2 Why rolling-mean extrapolation is biased

Consider a campaign whose true budget–revenue relationship is concave with diminishing returns — empirically, a saturating Hill-type curve [9], [16]:

so the corresponding revenue function ${\rho }_i(b)={\mathrm{AOV}}_i\cdot {\lambda }_i(b)\cdot q_i$ is also concave, where $q_i$ is conversion rate and ${\mathrm{AOV}}_i$ is average order value. The marginal ROAS at budget $b$ is

This is the quantity a profit-maximising allocation must equalise across campaigns. The reported average ROAS is, by contrast,

which is the average return per dollar, not the marginal return. By the standard concavity argument, average exceeds marginal whenever the curve is concave, and the gap

is monotone increasing in spend over the saturating regime. A rolling-mean estimator that targets average rather than marginal ROAS therefore over-allocates to campaigns that are already saturated, because their average ROAS remains high even as their marginal ROAS approaches zero. This is the analytical mechanism behind the empirical observation that high-spend campaigns under manual management exhibit declining contribution despite stable dashboard metrics [1, §X]. The applied counterpart — equalising marginal-ROAS curves across platforms rather than allocating by historical revenue share — is documented in the cross-platform ad-budget allocation note.

3.3 Attribution latency and the noise-floor problem

A second issue compounds the bias. Marketplace platforms typically credit purchases to ads within an attribution window of 7–14 days [21]. The reported ROAS on day $t$ therefore mixes click cohorts from days $t-13$ through $t$, with non-trivial weights and a partial reveal of any given day's cohort. A rolling 7-day mean over partially-attributed daily ROAS is a convolution of two smoothing kernels and inherits a phase lag of approximately 4–7 days relative to the true daily marginal ROAS. Sellers reacting to this lagged signal effectively chase the previous regime, and the auto-correlation induced by the attribution mixing is routinely mistaken for genuine signal — a textbook spurious-regression problem in the sense of Granger and Newbold.

Lewis and Rao [14] establish the unfavourable economics of advertising measurement result: for typical SMB advertising volumes, the standard error of an observational ROAS estimate is large enough that point estimates within $\pm 30\%$ are statistically indistinguishable from the null. Sellers without uncertainty quantification therefore make decisions on signals that are formally below their own noise floor.

3.4 Inventory dilution and CVR degradation

A third compounding effect, documented by Ma et al. [15] for Taobao search ads, is that conversion rate degrades when click volume scales beyond the seller's effective inventory or fulfilment capacity. Letting $u_{i,t}$ denote campaign utilisation,

where $g$ is a monotone-decreasing dilution function above a campaign-specific threshold. Rolling-mean heuristics, which assume $q_i$ is constant in $b$, will systematically over-state the marginal return of a campaign whose budget increase pushes it past the dilution threshold. The connected operating problem — when ad volume scales into a stockout window — is treated quantitatively in the stockout-cost note.

3.5 Composite cost

In aggregate, manual heuristics are biased estimators of marginal contribution because they (i) confuse reported with true ROAS, (ii) confuse average with marginal, (iii) ignore attribution latency, (iv) ignore inventory dilution, and (v) ignore the platform-side cost of bid churn. The DataGlass system addresses each of these explicitly. The structural reframing — moving from a dashboard model to a decision-engine model — is summarised in the decision-engine architecture note.

4. Problem Formulation

We adopt the notation of the underlying technical report [1], [5]. Let $i\in \{1,\dots ,N\}$ index a seller's active campaigns, and let $b_{i,t}\in [{\underline{b}}_i,{\overline{b}}_i]$ denote the daily budget on day $t$. The decision problem each morning is to choose a vector ${\mathbf{b}}_t=(b_{1,t},\dots ,b_{N,t})$ that maximises expected portfolio profit subject to operational constraints.

The expected profit of campaign $i$ at budget $b$ admits the decomposition

where ${\lambda }_i(b)$ is expected clicks (modelled as a Negative-Binomial count process to absorb overdispersion [8], [23]), $q_i(\cdot )$ is conversion rate (modelled as Beta–Binomial with utilisation-dependent dilution [15]), $s_i(b)$ is realised spend (close to but not identically equal to the budget under platform pacing), $c_i^{\text{plat}}(b)$ are platform commission and fee charges, and $m_g$ is the gross-margin multiplier introduced in Section 3.1.

The portfolio problem is

where $B_t$ is the daily portfolio cap and ${\mathcal{F}}_{i,t}$ encodes per-campaign feasibility — minimum-budget floors, learning-phase freezes, stability windows, action-count caps, and ROAS guardrails. The exact composition of ${\mathcal{F}}_{i,t}$ is part of the proprietary specification and is documented internally [3]–[5].

Under interior optimality, the first-order conditions give the equal-marginal-profit characterisation:

where ${\mu }^{\ast }$ is the shadow price on the budget constraint and ${\mathcal{A}}_t$ is the set of campaigns not on a binding feasibility constraint. The shadow price ${\mu }^{\ast }$ has a clean economic interpretation as the marginal contribution of the next portfolio dollar; making this object visible to the seller is itself a non-trivial improvement over the heuristic baseline.

5. Methodological Foundations

The methodology underlying DataGlass draws from four bodies of literature.

Response curve modelling. The Hill saturation function originates in pharmacometrics [16] and has become standard in modern marketing-mix modelling through Google's Robyn framework [11] and Meta's Meridian system. The Michaelis–Menten form is a special case ($\theta =1$); the power-law is the small-budget limit. Naik and Raman [18] extend these to multimedia synergies. Over-dispersed click counts are handled via the Negative Binomial (NB2) parameterisation [8], [23], with overdispersion testing in the spirit of Dean [9]. Conversion rates are modelled as Beta–Binomial with utilisation-dependent dilution [15]. Richardson et al. [21] provide foundational click-through rate prediction work for the cold-start regime.

Constrained optimisation and bandits-with-knapsacks. Under budget constraints, the problem falls within the Bandits-with-Knapsacks (BwK) framework [20], where regret scales with the optimum-to-budget ratio. Heymann et al. [25] study Combinatorial Multi-Armed Bandits for portfolio allocation. Dynamic regret under non-stationarity follows Besbes et al. [26] with bound

where $V_T$ is the path variation; Chen et al. [27] sharpen this to $\tilde{O}\left(\sqrt{V_T^{\text{res}}T}\right)$ when conditioning on side information. Balseiro et al. [28], [29] develop dual mirror descent for online allocation with simultaneous regret and constraint-violation guarantees, directly relevant to the per-campaign ROI constraint in our setting.

Exploration under uncertainty. Thompson Sampling achieves regret $O(N\sqrt{T\log T})$ in multi-armed bandits [30] and is empirically superior under delayed feedback and model misspecification [31], [32]. Ouyang et al. [33] extend Thompson Sampling to unknown MDPs; Jauvion et al. [34] deploy it for header bidding in supply-side platforms with demonstrated production-level gains over UCB.

Changepoint detection. CUSUM is the classical online-detection procedure [19] with well-understood average-run-length (ARL) properties [24]. The recursion is

where $e_t$ is the standardised residual and $h$ is calibrated to a target ARL via Siegmund's approximation [24]. Modern alternatives include Bayesian online changepoint detection [35] and the kernel-based methods surveyed in Aminikhanghahi and Cook [36].

Causal identification. Gordon et al. [37] establish the gold standard of randomised experiments for advertising measurement; Lewis and Rao [14] document the unfavourable signal-to-noise economics of observational measurement. Oster [38] provides sensitivity analysis for unobserved confounding. Brodersen et al. [39] develop Bayesian structural time-series for causal impact. DataGlass combines first-differencing, prospective randomised perturbation, and Oster-style sensitivity analysis [1, §VIII].

6. The DataGlass System

We describe the system at the highest level consistent with the rigour required by an academic communication. Proprietary tuning parameters, the precise constraint composition, the perturbation-experiment design, and the segment-specific defaults are not disclosed here; they are documented in [2]–[5] and in the production engineering specifications.

6.1 Component 1 — Bayesian response model

For each campaign $i$, DataGlass maintains a posterior distribution over $({\lambda }_i^{\max },K_i,{\theta }_i,q_i,k_i)$, where $k_i$ is the NB2 dispersion parameter. The click model uses Negative-Binomial likelihood with Hill saturation; the conversion model uses Beta-Binomial likelihood with utilisation-dependent dilution. Estimation is two-stage with ridge stabilisation on the click stage and conjugate updates on the conversion stage; standard errors are propagated using a Murphy-Topel correction [40] whose factorised form is exact in our setting because the two stages condition on disjoint sufficient statistics [1, §VI].

A key design choice is the conditional Hill upgrade: the saturating curve is fitted only when the campaign's recent budget range satisfies $b_{\max }/b_{\min }>2$ and effective sample size exceeds an identifiability threshold [1, §VI.B]; otherwise, a more conservative parametric form is used. This directly addresses the structural-and-practical-identifiability concerns raised in Raue et al. [41] and avoids the systematic over-fit that occurs when nonlinear saturation parameters are estimated on insufficiently variable data.

6.2 Component 2 — Constrained portfolio optimiser

Given the posterior, the optimiser solves the program of Section 4. The active-set structure is exploited by reducing the problem to a one-dimensional search on the shadow price $\mu$:

solved by bisection. Implicit differentiation of $b_i^{\ast }({\mu }^{\ast })$ with respect to the posterior parameters propagates uncertainty into the recommended budgets, so every output is accompanied by a credible interval rather than a point estimate. The full constraint composition, the action-count cap, and the tie-breaking rules among simultaneously binding constraints are not documented here. The applied analogue — translating the shadow-price view into seller-facing operating actions — is illustrated in the ads-optimisation solution surface.

6.3 Component 3 — Non-stationarity layer

Three mechanisms operate jointly. Thompson Sampling provides principled exploration with the regret bound of [30] specialised to the budget-constrained setting following Balseiro et al. [29]. CUSUM-based changepoint detection on per-campaign residuals fires when the response curve has shifted, with thresholds calibrated to a target ARL via [24] and an internal calibration table [1, §VII.B]. A three-layer regime detector — calendar-aware, multivariate Hotelling $T^2$, and exogenous event log — votes on the day-level operating mode (aggressive, cautious, exploratory, hold).

6.4 Component 4 — Causal identification

To prevent the system from learning a confounded budget-response relation, DataGlass applies first-differencing for slow-moving confounders and runs prospective randomised perturbation experiments — small, calibrated, deliberately randomised budget shocks — that generate clean identifying variation in the neighbourhood of the current operating point. Following [37], this is the gold-standard mechanism for causal estimation; the perturbation magnitude, schedule, and the procedure for folding the resulting evidence back into the posterior are part of the system's confidential calibration.

7. Discussion

Three points deserve emphasis.

First, the empirical advantage of DataGlass over the manual baseline is not primarily a function of model expressiveness. The Hill function, the Negative Binomial, and the Beta-Binomial are textbook components. The advantage comes from (i) modelling the correct objective — true profit-adjusted contribution rather than reported ROAS — and (ii) respecting the platform-induced constraint structure, in particular the cost of bid churn under learning-phase mechanics. A naive deep-RL system that ignored these constraints could easily produce worse production behaviour than a well-disciplined manual operator, despite richer function approximation. This is consistent with the ablation findings of Liu et al. [42] for real-time bidding.

Second, the system's value to the seller is monotone in the gap between reported and true ROAS. Sellers in low-margin, high-return-rate categories benefit disproportionately because their break-even reported ROAS is far above what the dashboard suggests, and the manual baseline mis-prices their portfolio most severely. Sellers in high-margin, low-return categories see smaller — though still material — lifts.

Third, the methodology described here generalises beyond the specific platform on which the system was originally deployed. The interface — daily budget plus target ROAS, with platform-side opaque auto-bidding — is now the dominant pattern across major marketplaces (Shopee, Lazada, TikTok Shop, Amazon Sponsored Products, Walmart Connect, Mercado Libre Ads). A system that treats this interface as a first-class modelling object — rather than as an inconvenience to be wrapped around a generic ad-tech stack — has a structural, not incidental, advantage. The full taxonomy of seller decisions a unified decision layer covers — including pricing, inventory, promotions, and personalisation — is documented in the decision-intelligence playbook.

8. Limitations and Future Work

Several limitations are explicitly acknowledged in [1, §XI] and [5, §11].

The Hill function is identifiable only when budget ranges are sufficiently wide and effective sample sizes exceed a calibrated threshold; below this threshold the system falls back to a more conservative parametric form, with a corresponding reduction in expected lift. Broad-attribution weighting (currently ${\alpha }_{\text{broad}}=0.5$) is heuristic and a candidate for replacement by an attribution-window-aware model. Target-ROAS confounding is partially addressed via covariate inclusion; full treatment requires a joint optimisation of budget and target ROAS, planned for v2.0. Daily aggregation is a v1.0 limitation; intra-day allocation is in development. Campaign cannibalisation is detected but not yet adjusted in the optimiser, with the v1.1 extension formalising the joint penalty ${\pi }_{i+j}(b_i,b_j)={\pi }_i(b_i)+{\pi }_j(b_j)-\kappa \min (b_i,b_j)$ where $\kappa$ is estimated from the negative partial correlation [43]. Ad-group level allocation is planned for v2.0.

The system's worst empirical performance occurs under regime shifts that are simultaneously global (affecting many campaigns) and unannounced (no calendar or event-log signal). The three-layer regime detector reduces but does not eliminate this risk; a more aggressive multivariate detector is in research.

For the canonical specification of the sample frame underlying every "in our data" claim across the DataGlass research corpus — sample size, time window, exclusions, attribution windows — see the DataGlass research methodology.

9. Conclusion

The DataGlass decision-intelligence system replaces the dominant family of manual ad-budget heuristics — trial-and-error reallocation, rolling-mean extrapolation, and intuition-driven bid edits — with a coherent Bayesian decision pipeline that respects the platform constraints under which marketplace sellers actually operate. We have shown analytically that the manual baseline is biased in identifiable, structural ways: it confuses reported with true ROAS, average with marginal return, attributed with un-lagged signal, and unconstrained with constraint-feasible action. We have summarised the methodological foundations of the replacement system and reported empirical validation showing 18–24% portfolio-profit improvements with reduced reallocation frequency. The proprietary calibration that distinguishes the production system from a textbook implementation is deliberately withheld; the academic content is documented in the companion peer-reviewed manuscript [1].

For sellers, the implication is direct: a system that targets the correct objective on the correct constraint set captures contribution that is otherwise structurally unavailable to manual operation. For the field, the implication is that interface-aware decision intelligence — taking the platform's API surface seriously as a modelling object — is now the binding research frontier in marketplace advertising.

Acknowledgements

The authors thank the DataGlass research team for the proposal series [2]–[5], the peer-review committees [44], [45] for substantive feedback that shaped Revisions 1 through 3, and the participating Shopee sellers for consenting to the live A/B deployment.

References

Internal technical reports

[1] DataGlass Research, "DataGlass: Bayesian Budget Allocation for E-Commerce Advertising Under Platform Constraints," IEEE Trans. Artif. Intell., manuscript ready for submission, March 2026.

[2] DataGlass Research, "DataGlass Ads Optimizer: Model Proposal," internal technical report, March 2026.

[3] DataGlass Research, "DataGlass Model Proposal — Revision 1," internal technical report, March 2026.

[4] DataGlass Research, "DataGlass Model Proposal — Revision 2," internal technical report, March 2026.

[5] DataGlass Research, "DataGlass Model Proposal — Revision 3 (Final)," internal technical report, March 2026.

Industry and macroeconomic context

[6] eMarketer / Insider Intelligence, "Worldwide retail and e-commerce sales forecast 2024–2027," 2024.

[7] J. Madsoon, "Mind the Gap: Quantifying Seller Pain in E-Commerce," DataGlass Research working paper, May 2026.

[10] Jungle Scout, "State of the Amazon Seller 2024," industry survey ($n=2,164$), 2024.

[13] McKinsey & Co., "State of Small and Medium Business Index," 2024.

[17] Shopify, "Commerce Trends 2024," industry report ($n>1,000$ merchants), 2024.

[22] eMarketer, "Global retail e-commerce sales 2024," industry data, 2024.

Response curve and demand modelling

[8] A. C. Cameron and P. K. Trivedi, Regression Analysis of Count Data, 2nd ed. Cambridge Univ. Press, 2013.

[9] C. B. Dean, "Testing for overdispersion in Poisson and binomial regression models," J. Amer. Statist. Assoc., vol. 87, no. 418, pp. 451–457, 1992.

[11] Y. Jin, Y. Wang, Y. Sun, D. Chan, and J. Koehler, "Bayesian methods for media mix modeling with carryover and shape effects," Google Research, 2017.

[15] X. Ma, L. Zhao, G. Huang, Z. Wang, Z. Hu, X. Zhu, and K. Gai, "Entire space multi-task model: An effective approach for estimating post-click conversion rate," in Proc. SIGIR, 2018.

[16] C. Ritz, F. Baty, J. C. Streibig, and D. Gerhard, "Dose-response analysis using R," PLoS ONE, vol. 10, no. 12, e0146021, 2015.

[18] P. A. Naik and K. Raman, "Understanding the impact of synergy in multimedia communications," J. Marketing Res., vol. 40, no. 4, pp. 375–388, 2003.

[21] M. Richardson, E. Dominowska, and R. Ragno, "Predicting clicks: Estimating the click-through rate for new ads," in Proc. WWW, 2007.

[23] J. M. Ver Hoef and P. L. Boveng, "Quasi-Poisson vs. negative binomial regression," Ecology, vol. 88, no. 11, pp. 2766–2772, 2007.

Bandits, optimisation, and online learning

[20] A. Badanidiyuru, R. Kleinberg, and A. Slivkins, "Bandits with knapsacks," J. ACM, vol. 65, no. 3, pp. 1–55, 2018.

[25] B. Heymann et al., "Combinatorial bandits for budget allocation in marketplace advertising," arXiv preprint, 2025.

[26] O. Besbes, Y. Gur, and A. Zeevi, "Stochastic multi-armed-bandit problem with non-stationary rewards," in Proc. NeurIPS, 2014.

[27] Y. Chen, C. Lee, and H. Luo, "A new framework for oracle-efficient online learning with side information," in Proc. COLT, pp. 1060–1081, 2019.

[28] S. R. Balseiro, H. Lu, and V. Mirrokni, "Dual mirror descent for online allocation problems," in Proc. ICML, 2023.

[29] S. R. Balseiro, H. Lu, and V. Mirrokni, "Primal-dual budget pacing with ROI constraints," in Proc. ICML, 2024.

[30] S. Agrawal and N. Goyal, "Thompson Sampling for contextual bandits with linear payoffs," in Proc. ICML, 2013.

[31] O. Chapelle and L. Li, "An empirical evaluation of Thompson Sampling," in Proc. NeurIPS, 2011.

[32] D. J. Russo, B. Van Roy, A. Kazerouni, I. Osband, and Z. Wen, "A tutorial on Thompson Sampling," Found. Trends Mach. Learn., vol. 11, no. 1, pp. 1–96, 2018.

[33] Y. Ouyang, M. Gagrani, A. Naber, and R. Jain, "Learning unknown Markov decision processes: A Thompson Sampling approach," in Proc. NeurIPS, 2017.

[34] G. Jauvion, N. Grislain, P. Sielenou, A. Veyrat, and D. Gourru, "Optimization of an SSP's header bidding strategy using Thompson Sampling," in Proc. KDD, 2018.

Real-time bidding and reinforcement learning

[42] S. Liu, C. Hua, Y. Chen, and J. Wang, "Real-time bidding strategy in display advertising: An empirical analysis," arXiv:2208.07516, 2022.

Changepoint detection and non-stationarity

[19] E. S. Page, "Continuous inspection schemes," Biometrika, vol. 41, nos. 1–2, pp. 100–115, 1954.

[24] D. Siegmund, Sequential Analysis: Tests and Confidence Intervals. Springer, 1985.

[35] R. P. Adams and D. J. C. MacKay, "Bayesian online changepoint detection," arXiv:0710.3742, 2007.

[36] S. Aminikhanghahi and D. J. Cook, "A survey of methods for time series change point detection," Knowledge and Information Systems, vol. 51, no. 2, pp. 339–367, 2017.

Causal identification and advertising measurement

[14] R. A. Lewis and J. M. Rao, "The unfavorable economics of measuring the returns to advertising," Quart. J. Econ., vol. 130, no. 4, pp. 1941–1973, 2015.

[37] B. R. Gordon, F. Zettelmeyer, N. Bhargava, and D. Chapsky, "A comparison of approaches to advertising measurement," Marketing Sci., vol. 38, no. 2, pp. 193–225, 2019.

[38] E. Oster, "Unobservable selection and coefficient stability," J. Bus. Econ. Stat., vol. 37, no. 2, pp. 187–204, 2019.

[39] K. H. Brodersen, F. Gallusser, J. Koehler, N. Remy, and S. L. Scott, "Inferring causal impact using Bayesian structural time-series models," Ann. Appl. Stat., vol. 9, no. 1, pp. 247–274, 2015.

[40] K. M. Murphy and R. H. Topel, "Estimation and inference in two-step econometric models," J. Bus. Econ. Stat., vol. 3, no. 4, pp. 370–373, 1985.

Identifiability and statistical foundations

[41] A. Raue et al., "Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood," Bioinformatics, vol. 25, no. 15, pp. 1923–1929, 2009.

Cross-channel effects

[43] I. M. Dinner, H. J. Van Heerde, and S. A. Neslin, "Driving online and offline sales: The cross-channel effects of traditional, online display, and paid search advertising," J. Marketing Res., vol. 51, no. 5, pp. 527–545, 2014.

Behavioural foundations

[12] D. Kahneman, Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011.

Internal review documents

[44] DataGlass Research, "DataGlass Model Proposal — Peer Review (R1)," internal review, March 2026.

[45] DataGlass Research, "DataGlass Model Proposal — Peer Review (R2 and R3)," internal review, March 2026.

DataGlass is a research-led decision-intelligence platform for e-commerce sellers. For partnership, integration, or research-collaboration enquiries, contact the authors directly. The peer-reviewed account of the methodology is forthcoming [1].