Gradient-boosted trees for market data, explained simply

The idea\nGradient boosting builds many small decision trees. Each new tree corrects the errors of the trees before it.\n\n## Features we use\n- Recent returns over 1, 5 and 20 sessions\n- Distance from 20 and 50-session averages\n- Recent volatility\n\n## Guarding against overfitting\nTrees are shallow, learning is slow (shrinkage), and every forecast origin in the backtest is trained only on data available at that time. If the model does not beat simpler baselines, its ensemble weight shrinks.