Why ridge is the benchmark
Ridge keeps the model simple while still letting the features carry real weight. That makes it a strong alpha benchmark: if the tree cannot beat ridge on the same inputs, the extra flexibility is probably not buying much.
Where boosted trees help
Boosting helps when the relationship is nonlinear, thresholded, or tied to regime interactions. It can capture pieces of the signal that ridge will flatten away. The catch is obvious: the same flexibility can also memorize noise.
How to compare them
Keep the features, split, costs, and trade rule identical. Then compare the out-of-sample result, not just the training score. If the tree only wins on one lucky window, the comparison is done.
- Same inputs. No hidden advantage for either model.
- Same chronology. Walk-forward or chronological split only.
- Same friction. Costs and slippage have to match.
The blunt rule
Start with ridge. Move to boosted trees only if the simpler model leaves real signal on the table.
Common mistakes
- Giving the tree more help. Different feature sets make the comparison meaningless.
- Ignoring stability. A higher score on one fold is not enough.
- Calling complexity alpha. More knobs is not the same as more edge.