Avito Ran an Uplift Model Tournament: X-learner Won, Not Causal Forest

Grisha Kryukov, an analyst on Avito's trust and safety team, staged a "championship" in July 2026 among four popular uplift models — S-learner, T-learner, X-learner and Causal Forest — with a group stage and a playoff. X-learner took gold unexpectedly: it left the group stage only third, but won both semifinals and the final.

Why uplift models matter

Uplift modeling estimates the individual effect of an intervention — how much a specific person's behavior changes because of a campaign, rather than the average response of a group. The main difficulty is that you cannot observe both outcomes for a single subject at the same time: with the intervention and without it.

"The main problem: for a single player, we never know both situations at the same time," —

Grisha Kryukov, analyst on Avito's trust and safety team.

The four approaches tackle this problem differently. S-learner feeds the treatment flag into the model as an ordinary feature. T-learner trains two separate models. X-learner adds training on pseudo-effects and propensity weighting. Causal Forest builds trees that maximize the difference in effect.

How the tournament worked

The tournament consisted of a group stage on synthetic data and a playoff on real data. The author measured model quality with PEHE (Precision in Estimation of Heterogeneous Effect) — the lower the value, the more accurate the effect estimate.

• Models: S-learner, T-learner, X-learner, Causal Forest
• Metric — PEHE; group points awarded by place: 3-2-1-0
• Group stage — synthetic data, 5000 observations, 10 features, 70/30 split
• Semifinals — the Hillstrom MineThatData dataset, 64 thousand e-commerce customers
• Final — the FIFA 21 dataset, about 10 thousand players and 26 attributes

After three group rounds, Causal Forest led with 6 points, S-learner and X-learner each had 5, and T-learner closed out the table with 2 points.

Why X-learner won

X-learner won because its architecture learns specifically from the difference between treated and untreated scenarios, not from the overall response level. In the semifinal it crushed S-learner (PEHE 0.215 versus 0.747), and in the final it beat T-learner with a PEHE of 0.579 versus 1.135.

"The architecture works as intended: instead of the overall level, the model learns specifically from the difference between the worlds with and without the intervention," —

Grisha Kryukov.

The author explains the favorites' defeat this way: on synthetic data, where the distribution is ideal, the robust Causal Forest had the edge, but on "dirty" real-world data the more flexible architecture won. T-learner took particular criticism: its two independent models do not share information, and that proved fatal in the final.

What this means

There is no universal winner in uplift modeling: Causal Forest wins on clean synthetic data and small samples, while X-learner wins on noisy real-world data. Choose a model for your specific dataset, not based on a method's reputation.

Frequently asked questions

What is uplift modeling?

It is a method for estimating the individual effect of an intervention — how much a specific user's behavior changes because of a campaign or an email, rather than the average effect across the whole audience. It is used in marketing, retention and anti-fraud.

Which model won the Avito tournament?

X-learner was the champion: it left the group stage only third, but won both semifinals and the final with a PEHE of 0.579 versus 1.135 for T-learner. The reason is its architecture, which learns from the difference in effects.

What does the PEHE metric mean?

PEHE (Precision in Estimation of Heterogeneous Effect) measures how accurately a model estimates the individual effect: the lower the value, the closer the prediction is to the true one. In the tournament, it determined the rankings.