How to backtest a trading strategy without fooling yourself

Entry condition, exit condition, holding period, and what counts as a win. Fixing this in advance is what stops a test from being quietly adjusted until it looks good. Our rules: enter at the closing price on a signal day, exit at the close a fixed number of trading days later, win means the exit close moved in the signalled direction.

On each historical bar the engine must see only prices up to that bar. This sounds obvious and is violated constantly — by indicators computed over the whole series at once, by survivorship-filtered ticker lists, and by split or dividend adjustments applied with hindsight. Use split-adjusted session opens, highs, lows and closes, and recompute every indicator inside the loop.

This is the error that produces the absurd numbers. If a strategy signals almost daily with a 20-day hold, twenty positions overlap; compounding them as though each were fully funded is arithmetically impossible. Before we fixed it, one ticker reported +19,812%. With one fully-allocated position at a time, and a count of how many signals were skipped because capital was already committed, the results became something a real account could have produced.

Split your universe. We used 15 tickers to choose filters and held 15 back entirely, measuring those once, afterwards. The held-back numbers were lower — 57.3% hit rate and 0.38 Sharpe versus 58.8% and 0.59 — and they are the ones worth believing. Any change that helps in sample and hurts out of sample gets discarded, which is exactly what happened to a strategy re-weighting scheme that looked excellent in testing.

• Hit rate — share of positions that moved the signalled way.
• Average return per position — the honest per-trade figure.
• Maximum drawdown — the worst peak-to-trough fall you would have sat through.
• Sharpe ratio — return relative to volatility; 0.5 is modest, above 1.0 is strong.
• Buy and hold — the benchmark that most strategies quietly lose to.

Commissions, bid-ask spread, slippage on thin names, taxes and missed dividends all sit outside a backtest. A 0.2% average gain per position rarely survives them. Treat a marginal result as a negative one.

The Backtesting page runs this exact procedure on any US-listed ticker over one, two or five years with holding periods of 5 to 30 trading days, reporting hit rate, average and compounded return, drawdown, Sharpe, per-strategy accuracy and the buy-and-hold comparison, plus how many signals were skipped and why. Our own results across 30 tickers are published on the measured performance page, losing names included.