The goal
Most trading backtests look great because they quietly overfit to the past. The point of this project is to learn strategies and then test them in a way that resists that trap.
How it works
Market data→
Q-Learning→
Ensemble→
Walk-forward test
Learn, combine, then validate out of sampleresults you can trust
- Q-Learning treats trading as a sequence of decisions and learns which action to take in each market state from the rewards it earns.
- Ensemble methods combine several models so no single one's quirks drive every trade.
- Walk-forward optimization trains on one window of history, tests on the next unseen window, then rolls forward, which mirrors how a strategy would actually be used.
The lesson
Evaluation design matters more than model choice. A simple strategy tested honestly is worth more than a clever one tested on data it has already seen.