Probability of loss is the chance that a campaign’s profit over a given period comes out negative — the share of plausible outcomes that fall below zero. It answers how often a day loses money, and nothing else. It is a frequency, so it is completely blind to amounts: a campaign that loses fifty baht on four days out of five and makes two thousand on the fifth has a probability of loss of 0.8 and is comfortably profitable. Used as a stop-loss trigger without that caveat, it fires on exactly the campaigns whose demand is sparse rather than on the campaigns that are losing money.
Formula
prob_loss = P( profit < 0 )
For a campaign whose orders arrive sparsely,
prob_loss ≈ the probability of a zero-order day
And note: 1 − prob_loss ≠ recommendation confidenceExample
A high-value Shopee campaign: most days no order, occasionally one
large order.
probability of loss 0.65
average daily profit +฿417
ROAS 9.5
The campaign loses money on nearly two days in three and is one of the
best in the account. A rule that pauses above 0.5 would kill it.Why do sparse campaigns always score high?
Because on a day with no order, profit is exactly the negative of that day’s spend — a loss, by definition. For a campaign that converts infrequently, the probability of loss therefore converges on the probability of a zero-order day, which is a property of how lumpy demand is rather than of whether the campaign pays. High-value products, niche categories, and anything with a long consideration cycle all have sparse conversion and will all show a high probability of loss no matter how profitable they are. Reading the metric as evidence of unprofitability inverts the truth on precisely the campaigns where the stakes are highest.
Is probability of loss the opposite of confidence?
No, and this is the most common misreading. One minus the probability of loss is the chance that profit at the current spend is at least zero. Recommendation confidence is the chance that a proposed change beats leaving things alone. They are anchored to different reference points — zero in one case, the status quo in the other — so they measure different things and do not sum to one. A campaign can have a 65% probability of loss and a 1.7% confidence that cutting its spend would help, and both figures can be correct simultaneously: it loses on most days and is still much better funded than unfunded.
When is it the right metric?
When the frequency of a loss genuinely matters more than its size — a cash-flow constraint where a run of losing days cannot be absorbed, or a compliance limit expressed in occurrences. For deciding whether a campaign should exist, use a magnitude-aware measure instead: expected profit if the distribution is well behaved, or the average of the worst outcomes if it is skewed and the downside is what you are protecting against. Probability of loss is a useful thing to display next to those, because it describes the texture of the campaign, but it is a poor thing to act on alone.
Why it matters
Probability of loss is intuitive, easy to compute, and therefore extremely common as an automated stop rule — and it is the wrong rule for marketplace advertising, where demand is lumpy by nature. A threshold like “pause above 50%” will fire on high-value, low-frequency, high-margin campaigns and leave alone the steady, thin-margin campaigns that lose a little every single day, which is the exact opposite of the intended behaviour.
In DataGlass
DataGlass reports probability of loss as descriptive context on a campaign, alongside the shape of the outcome distribution it comes from, but does not use it as a pause gate. Pause and budget recommendations are decided on a paired comparison against the current setting rather than on the frequency of losing days.