Gradient Boosting
Chains decision trees together, where each new one is trained specifically to correct the errors the previous ones made.
Why it matters
Unlike Random Forest, the trees here don't vote independently: they're built one after another, and each one focuses on the days the previous ensemble predicted worst. That usually beats averaging independent trees, at the cost of being easier to overfit if too many correction rounds are allowed to grow.
Build it yourself
Start from a fixed 50% base and add two correction rounds — each real tree corrects a little more, hundreds of times over.
Prediction % = 50 (base) + Correction 1 + Correction 2
Combined prediction
Example with only two correction rounds — the real model chains hundreds, each one smaller than the last.
Illustrative example numbers for practice — not real data.
How it works, step by step
- The first tree makes a simple starting prediction — for example, the historical average probability of a rise.
- That prediction's error is computed for every historical day: how far off it was, and in which direction.
- A second tree is trained, but not to predict the original outcome — instead, to predict THAT error, i.e. to learn how to correct it.
- The combined prediction (the first tree plus the correction) beats the first tree alone. The process repeats hundreds of times, each correction a little smaller than the last.
Honest limitation. Train too many correction rounds and the model can end up memorizing noise specific to the training data instead of a real pattern — guarded against with the same "walk-forward" validation care explained on the main models page. And a note on scope: it's the only model in this comparison whose predictions the project does not store day by day — the version it does store is LightGBM, which uses this same mechanism — so this page doesn't end with a chart.
Predictions
This particular model's daily predictions aren't stored, so this page doesn't end with a chart — that would mean faking a number that doesn't exist.