Running 9,000 Backtests Through Claude — Only Mean Reversion Survived in Brendan's Method

AI Pathways' Brendan publishes the strategy-validation system he built on Claude. Trend following, mean reversion, momentum, breakouts — every major family — tested across 30 assets and 15 years of data, with 9,000 backtests in total. The frame is to show quantitatively what survives.

His background: math and econ at UCLA, three years at Raymond James in investment banking, and two years building trading systems for clients. The video opens with a clear disclaimer that nothing here is financial advice, and is then disciplined in sticking to data and methodology. The test surface is daily bars on liquid assets — major ETFs like S&P 500 and NASDAQ, the sector ETFs, gold, oil, bonds, Bitcoin and Ethereum. Six filters whittle the candidate set down in stages.

The result lands cleanly. Among the major categories, mean reversion was the only family that came out positive, and shuffling the trade order didn't break it — a sign the result wasn't a luck artifact. 524 strategies cleared all six filters; restrict that to assets with over 10 years of history and it drops to 478. Of those 478, 64% are mean reversion strategies.

Why mean reversion dominates is consistent with the academic literature. After prices overshoot the mean, both human psychology (the unwinding of overreaction) and market structure pull them back. Decades of research underwrite the mechanism. By contrast, a moving-average crossover — one line crossing another — has no theoretical reason to predict anything, and the 9,000 backtests didn't support it either.

Brendan publishes this in a form anyone can replicate. Screenshot the prompts and hand them to an AI model — you can build it in the claude.ai web app or in Claude Code. Claude Code has the advantage of letting you manage and update the codebase in one place.

The construction is layered. The first layer is the data and strategy library — copy it into Claude Code and you're up. The second is the backtest funnel — walk-forward / out-of-sample validation plus the six filters. The third layer is robustness checks — parameter sensitivity and bootstrap stress testing. The final layer is the cross-sectional check, specific to momentum-based strategies. Together the layers give you a tool to validate your own ideas quantitatively, rather than trusting a backtest you happened to see on TradingView.