Off-Policy Evaluation
Definition
Off-policy evaluation estimates a new strategy based on data from another strategy.
The estimation of the expected outcome of a target policy based on logged data collected under a different behavior policy.
Simply explained
One tries to estimate from past decisions and outcomes how a different approach would have performed.
Why it matters in practice
Helps realistically assess model quality, limits, and technical decisions.
Example from business
A new recommendation strategy is evaluated with logged actions, outcomes, and the probabilities required for the chosen estimator before going online.
Limits and pitfalls
- Ignoring missing overlap between target and behavior policy, using unreliable propensity scores in weighting estimators, or concealing model errors, variance, and uncertainty.
Origin & history
OPE originated in reinforcement learning, causal inference, and contextual bandits.
Use in mid-sized companies
- Possible application: A new recommendation strategy is evaluated using logged actions, outcomes, and the probabilities required for the chosen estimator before going live.
Frequently asked questions
What does Off-Policy Evaluation mean?
The estimation of the expected outcome of a target policy based on logged data collected under a different behavior policy.
Which mistakes should be avoided?
Ignoring missing overlap between target and behavior policy, using unreliable propensity scores in weighting estimators, or concealing model errors, variance, and uncertainty.
Sources
Source checked on 10.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: