Why You Shouldn’t Judge by PnL Alone

I’ve spent years building and evaluating systematic strategies across highly adversarial markets. When you iterate on a trading system, PnL is the goal but a terrible day-to-day signal. It’s too noisy, too path-dependent, and too easy to cherry-pick. A simple framework—form a hypothesis, measure a test statistic, translate it into a probability under a “no-effect” world (the p-value)—helps you avoid false wins, iterate faster, and ship changes that actually stick. Below I’ll show a concrete example where two strategies look very different in cumulative PnL charts, yet standard tests say there’s no meaningful difference in their average per-trade outcome. I’ll also demystify the t-test in plain language: difference of means, scaled by uncertainty.