The Backtest Trap · Lesson 1 of 5

The Seductive Curve

A backtest is not evidence that a strategy works. It is evidence that a strategy can be made to have worked.

6 min read · free forever · no sign-up


Load ten years of price data, test an idea, adjust it, test again. Two evenings later a beautiful equity curve climbs across your screen — steady, smooth, victorious. Here is the uncomfortable truth about that curve: its beauty is evidence of your editing, not the market's generosity.

The mechanism is called overfitting, and it needs no dishonesty — only iteration. Every time you tweak a parameter and re-run, you are not testing a strategy against the past; you are selecting for whatever combination best matches the past's accidents. Historical data contains real, repeatable structure — and mountains of one-off noise. A strategy fitted hard enough absorbs the noise, and noise, by definition, does not repeat. The curve is a portrait of a decade that will never happen again.

The tell is degrees of freedom. Count your adjustable choices: indicator settings, thresholds, session filters, stop and target sizes, "and skip December". Each is a dial, and with enough dials any dataset can be fitted perfectly — statisticians joke that with four parameters they can fit an elephant. A strategy with nine tuned conditions did not discover nine truths about markets; it made nine concessions to one particular history.

Two questions expose an overfit curve in seconds:

"Why does this rule exist?" A robust rule has a market reason (news moves prices; liquidity dries at rollover). An overfit rule has a data reason ("removing Tuesdays improved the Sharpe"). If the only justification is that it helped the backtest, it IS the backtest.

"What happens if I nudge it?" Change the 14 to a 15, the 40 pips to 38. A real edge degrades gently; an overfit one falls off a cliff, because the exact values were the fit. (Lesson four turns this into a formal tool.)

From our own ledger

Our research phase produced exactly one seductive discovery: an exit combination showing +151 pips — the best cell in the whole grid — which happened to involve trading with no stop-loss at all, on a sample that contained no catastrophe. Perfectly fitted to a past with no disasters in it. We refused it, in writing, and the reasoning is on our site. That refusal is this whole course in miniature.

The lesson in one line Iteration selects for flattery. Every dial you tune fits the noise a little tighter — and noise does not repeat.
← Course contents 2. The Five Ways a Test Lies →
Written by the team behind a real mechanical trading system with a live, verifiable public record — losses included. See the record · Free tools · All free courses · Join the free list