AutoTune Lab
Four ways to search a parameter space, and the machinery to tell you when the winner is just an artefact of the sample.
Search modes
- Grid Search — sweep the space exhaustively
- AI Bayesian — spend the trial budget where the surface looks promising
- Walk-Forward — rolling in-sample fit, out-of-sample verdict
- Walk-Forward 2 — same windows, each fold carrying its own portfolio picks forward
Ranking
- Nine ranking metrics — you choose which one the leaderboard obeys
- Composite score blends Sharpe, Sortino, profit factor, recovery and expectancy
- Rank by drawdown or Ulcer when survival matters more than total return
- Optimal portfolio picks the top five, and the top five that are not correlated
Overfit defence
- In-sample and out-of-sample windows kept strictly apart
- Walk-forward efficiency — out-of-sample return as a share of in-sample
- Overfit risk flagged on the report, not buried in a log
- Correlation analysis across the leaderboard, so five clones do not read as five ideas
The point of walk-forward here is to be allowed to fail. A parameter set that only worked on the window it was fitted to is reported as such.