Interpreting Gradient-Boosted Trees for Financial Price Forecasts
Summary
This article explores how to inspect a CatBoost gradient-boosted tree model trained to forecast a future price from market data and technical indicators. It motivates feature interpretation as a way to debug a model, evaluate engineered inputs, guide future data collection, and better understand predictions. The article notes that tree models divide inputs into groups and use fixed group averages, which can produce flat forecasts and limit extrapolation beyond the training range.
The demonstration uses MetaTrader data for a volatility index, indicator features, and a target price shifted into the future. The author examines feature contributions and interactions, then describes combining the tree forecast with a linear model in a trading example. The text provides implementation details but limited evidence for trading effectiveness: it does not report a systematic out-of-sample performance evaluation, transaction costs, or risk-adjusted results. Its observations about model limitations and interpretation should therefore be treated as a demonstration, not proof of a profitable forecasting method.
Key ideas
- Gradient-boosted trees can produce repeated or flat predictions because their learned group averages are fixed.
- The article uses a future price target and technical indicators as inputs to a CatBoost regression model.
- Feature importance and feature impact analysis can help identify influential inputs and possible sources of noise.
- Interpretation can inform debugging, feature engineering, and decisions about collecting additional data.
- A trading example is shown, but the document does not establish its out-of-sample profitability or account for trading costs.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.