Backtesting 101: Why Most Retail Backtests Lie

A beautiful equity curve proves nothing

If a strategy looks too good, the bias is in the test, not the alpha. Honest backtesting is mostly about hunting for the ways you fooled yourself.

Lookahead bias

Using information that was not available at decision time. Computing a signal on today's close and acting at today's open is the classic offender. Every input must be known strictly before the trade.

Survivorship bias

Testing only on stocks that exist today silently deletes the ones that were delisted or went to zero. The universe must be the universe as it was on each historical date.

Overfitting

Tune enough parameters and any noise becomes a "system". A rule with ten conditions that perfectly fits the past usually fails out-of-sample. Fewer knobs, tested on data the model never saw, is the cure.

Ignoring costs

Brokerage, STT, slippage and impact turn many paper-profitable systems into losers. Always backtest net of costs, not gross.

The honest workflow

Define the rule, test on in-sample data, validate on untouched out-of-sample data, and only then size small and trade live. If the edge survives all four checks, it might be real.