Backtesting won't save you, but it will humble you

Backtesting has a seductive quality. You take a strategy, run it over historical data, and out comes a curve that climbs up and to the right. There is something deeply reassuring about that curve. It feels like proof. It feels like you have discovered something true about the market. And then you trade it live and it falls apart, and you are left wondering what you did wrong. Usually the answer is that you believed the curve.

This is not an argument against backtesting. It is a genuinely valuable tool. It is an argument against the specific, very common mistake of treating a backtest as a prediction rather than as what it actually is — a description of how a set of rules would have behaved over one particular slice of the past, which has already happened and will never happen again in the same way.

The past is not a sample of the future

The core problem is subtle. When you optimise a strategy against historical data, you are tuning it to fit that specific history, including all the noise and accidents that history contained. The more you tweak, the better it fits — and the more of what it is fitting is coincidence rather than signal. This is overfitting, and it is the quiet killer of backtested strategies.

Given enough parameters to fiddle with, you can make almost any strategy look brilliant in hindsight. That is not a sign of a good strategy. It is a sign that you have memorised the answer sheet.

A strategy that has been heavily optimised over past data has effectively memorised that data. It knows exactly which trades worked because it was built after those trades were already known. The future, inconveniently, has not read the answer sheet. It will present conditions the backtest never saw, and the overfitted strategy, having learned the specifics rather than a general edge, has nothing useful to say about them.

What a backtest can honestly tell you

Used carefully, a backtest answers real questions. It tells you whether a strategy is even coherent — whether the logic does what you think it does. It shows you how the strategy behaves in different regimes: what happens in a trend, in a chop, in a crash. Most valuably, it shows you the shape of the pain. How deep do the drawdowns get? How long are the losing streaks? Can you actually stomach the equity curve, or does it have stretches that would make you quit?

That last point is underrated. The number that matters is rarely the total return. It is the worst stretch along the way. A strategy that doubles your money but has a period where it is down 60% for four months is a strategy almost nobody can actually trade, because they will abandon it at the bottom. The backtest is most useful not for promising returns but for warning you about the suffering required to earn them.

Testing it with a little honesty

A few habits separate a backtest that informs from one that flatters:

• Hold out data. Build on one period, then test on a period you never looked at while building. If it falls apart out of sample, it was overfit.
• Keep it simple. Fewer parameters mean fewer ways to accidentally fit noise. A rule with two knobs is more trustworthy than one with twelve.
• Include the frictions. Fees, spread, and slippage are not rounding errors. A strategy that is profitable before costs and underwater after them is just a generator of broker revenue.
• Be suspicious of perfection. If the equity curve is a smooth line with tiny drawdowns, you have almost certainly fooled yourself. Real edges are lumpy.

The right relationship with backtesting is the same as the right relationship with a signal. It is an input, not an oracle. It can tell you that an idea is worth trying, show you how it tends to behave, and prepare you for the drawdowns ahead. It cannot tell you that tomorrow will cooperate. Hold it loosely, assume live results will be worse than the test, size accordingly, and you will get the benefit of the tool without falling for its most flattering lie.