Model Fingerprints for Interpreting Financial Machine Learning Predictions
Summary
The article presents Model Fingerprints as a way to describe how machine learning features affect predictions. It estimates partial dependence by varying one feature while averaging predictions over other observations, then separates that dependence into linear and nonlinear effects. Pairwise interactions are estimated by comparing joint partial predictions with the effects of each feature individually. The method is presented alongside mean decrease impurity, mean decrease accuracy, and single feature importance, which have differing limitations, including sensitivity to correlated features.
A trend-following example combines a moving-average crossover signal with triple-barrier labels and a bagged tree classifier. The reported feature importance and fingerprint analyses identify volatility features as influential, including through pairwise effects, and suggest follow-up feature research. This is an illustrative workflow rather than broad evidence of predictive or trading performance. Partial dependence and the example’s model and data choices shape the interpretation, and the article notes that other interpretation methods are available.
Key ideas
- Model Fingerprints divide feature effects on predictions into linear, nonlinear, and pairwise components.
- Partial dependence summarizes prediction changes as a feature varies while other observations are averaged.
- MDI, MDA, and SFI estimate feature importance differently and can respond differently to correlated inputs.
- In the example, volatility features rank strongly across importance and fingerprint analyses.
- The example guides feature research but does not establish that the model will perform profitably out of sample.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.