Backtest vs live: reading results without fooling yourself

Every EA for sale comes with a backtest that rises from bottom-left to top-right. Producing one of those is easy. Producing one that means something is not. Here is what we check before a strategy is allowed near live money.

Modelling quality is a floor, not a verdict

A 99% modelling quality report only says the tick data was dense. It says nothing about whether the spread used matched reality, whether swap was applied, or whether the strategy would have been filled at those prices. Set the spread in the tester to your broker's average, not its best, and switch commission on.

Test through the periods that hurt

A backtest starting in 2021 tells you how a strategy handled a trending market. Include 2020 and 2022. Include the weeks around major central bank surprises. If a system has never been tested through a regime it will eventually meet, you do not know how it behaves there — you only know it has not met it yet.

A curve with no ugly stretch in it usually means the test period was chosen after the fact.

Four numbers past the equity curve

• Maximum drawdown, in percent and in duration. Twelve percent is manageable. Twelve percent that lasted five months is a different experience entirely.
• Number of trades. Under a few hundred and the result is anecdote, not evidence.
• Profit factor between 1.2 and 2.0. Below that is fragile; far above it usually means overfitting.
• Largest single win as a share of total profit. If one trade made a third of the money, the strategy is not what the curve suggests.

Change one input and see what happens

Take a working setting and move it ten percent in each direction. A robust strategy gets slightly worse. A curve-fitted one falls apart. The same applies to the start date — shifting the test window by a month should not change the conclusion.

Expect live results to be worse

They always are. Slippage, requotes, the occasional missed entry and the weeks you interfere by hand all take their share. If a strategy is only viable when the backtest's numbers hold exactly, it is not viable. Give yourself a margin: assume live performance around 70% of the tested figure and check that you would still be satisfied at that level.

The question worth asking

Before the numbers, ask what the strategy actually exploits. A trend follower on gold has a reason to work — gold trends. A system with no explanation beyond "the backtest is good" has found a pattern in noise, and noise does not repeat on schedule.