trading-box — AI-Augmented Multi-Strategy Trading

Why I built it

I wanted a concrete answer to the question "does any of this AI/ML actually help when there are real consequences for being wrong." Markets are a fair test: you can't cherry-pick examples after the fact, there's a clean backtesting protocol if you respect it, and the metrics are unambiguous.

How it works

A research-first architecture. Every strategy lives behind a Protocol: inputs are bars + features, outputs are signed position sizes. That makes it trivial to add a new strategy, run it against the full pair × timeframe matrix, and compare on the same scoring function.

Walk-forward adaptive whitelisting. Instead of picking pairs once and hoping, the system periodically re-selects the active pair whitelist from a trailing window — win rate above a threshold, sample size above another. Pairs that stop working drop out; pairs that start working get included.

Per-asset trend filtering. Different assets respond to different regime filters (ADX vs SMA-slope vs Donchian-channel). The system runs a small filter selection step per asset; in backtest the winner varied by asset, and forcing a single global filter would have left money on the table.

Lookahead-bias-clean backtesting. This is the part most homemade backtesters get wrong. Features are computed strictly from past bars; the walk-forward boundary is enforced in code, not by convention. Every metric in the report has a corresponding test that fails if a future bar leaks in.

Results

Over a three-year backtest on 53 crypto pairs at 1h timeframe, the hybrid strategy (per-asset baseline if its in-window profit factor is healthy, otherwise the dynamic whitelist) returned roughly +$103k against a baseline single-strategy approach that lost roughly $150k. The 4h timeframe was cleaner still — about 60% better than 1h, with traditional assets (forex, gold) only viable on 4h.

These are research-grade results from a personal-scale system. The point is not the dollar number — it's the discipline that produced it: protocols, holdout tests, no leakage, no overfit-to-narrative.