ML Ad Forecaster: Estimating Results Before Launch

Vladimir Davidko23 May 2026

Problem

An advertiser at a marketplace sets a budget, targeting and a bid, and wants to know before launch what they will get. Avito's first forecaster ran ClickHouse aggregates at request time. It took seconds, or minutes in complex scenarios, had no answer for new audience segments, and needed a growing zoo of query templates. Advertisers complained about forecast quality, and large advertisers will not allocate budgets on a forecast they do not trust.

Example from the article: men with premium income, interested in business and real estate, in three regions, weekdays from 10 to 19, for 7 days. The cabinet shows 1240 clicks, 5 days to complete delivery and a recommended bid of 23 ₽.

Solution in brief

Train CatBoost on an hourly dataset to predict the 10th, 50th and 90th quantiles of target actions. Sum the hourly forecasts into a campaign trajectory capped by the remaining budget. Campaigns with enough history are extrapolated from their own past instead. Verify with a shadow test, since an A/B test was not feasible.

Questions this case answers
  • How to forecast the delivery and outcome of a campaign with a prediction interval before it starts, when the customer can change inputs at any moment
  • How to replace a slow query-time forecaster with a model trained on hourly aggregates and validate it with a shadow test instead of an A/B test
  • How to forecast clicks or impressions of an ad campaign, with an interval, before it launches
  • How to build a prediction interval with quantile loss and measure it with MAE, coverage and Winkler Score

Tools and methods

  • CatBoost with quantile loss — forecast q10, q50 and q90 of clicks or impressions for cold campaigns
  • History extrapolation with seasonality — for campaigns above a history threshold
  • MAE, coverage, Winkler Score — offline quality of the point estimate and the interval
  • Shadow test — ML forecast logged next to the live analytical one, compared with actuals
Continue reading — it's free

Google or a one-time email code. No password.