Active Learning — bpleone / trade

Active learning = picking which samples to label first to make the model learn fastest. The model is most uncertain about findings where P(win) is near 50% — those are the ones where labeling gives the biggest accuracy boost per row.

Below: 30 most-uncertain unlabeled findings. Manually click HIT or MISS — model retrains immediately on each label. This is how you teach the model from your own knowledge faster than waiting for automatic outcome rating.

Total Uncertain

|prob - 0.5| < 0.15

Labeled This Session

Model Acc Before

Model Acc Now

📋 Most Uncertain Findings (sorted by |prob - 50%|)

SYM

P(WIN)

UNCERTAINTY

SETUP

LABEL

🧠 Why this works

Suppose the model is 90% confident on one finding and 50% confident on another. The 90% one already gives the model little to learn — outcome is mostly expected either way. The 50% one is genuinely undecided — labeling it teaches the model the most. In practice, labeling uncertain samples first can yield 2-5× more accuracy improvement per label than random sampling.

How to use: Trust your trading intuition. If you'd bet on the setup, click HIT. If you'd skip / fade, click MISS. If you can't tell, skip it. Don't agonize — the model just needs a directional signal.