Causal Inference with Propensity Scores for More Robust Financial Models

Summary

This research summary explains why predictive machine-learning models can mistake unstable associations for causal relationships, especially when selection bias or confounding drives observed patterns. It presents propensity-score analysis as a way to make treated and comparison groups more comparable using observed confounders. The described process estimates propensity scores and effects, checks covariate balance, and uses refutation tests to probe whether estimates are reliable.

The application examines four stock concepts and one-month-ahead returns among CSI 800 constituents from 2016 through March 2020, controlling for fundamental and price-volume factor exposures. It reports a positive relationship for seasonally adjusted fund-heavy concept membership and a negative one for stock-pledge membership; the evidence for earnings pre-announcement and moat concepts is inconclusive. Propensity-score weighting performed best on the paper’s balance and refutation checks. These are observational findings over a defined universe and period: adjustment for measured characteristics does not eliminate possible unmeasured confounding, and the summary does not show that the effects persist outside the sample.

Key ideas

• Predictive models can learn unstable associations caused by selection bias or confounding.
• Propensity scores compress observed confounders to improve comparability between groups.
• A useful analysis checks balance and tests the robustness of estimated effects.
• The reported study found different return effects across stock concepts, with some results inconclusive.
• Its findings are limited to the sampled stocks, concepts, period, and measured confounders.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.