Why your backtest looks better than your live trading

A backtest is a measurement, not a forecast

The confusion starts with the word itself. A backtest measures how a fixed set of rules behaved against a fixed set of historical bars. That is all it does. It cannot tell you that the rules will keep working, because the market that produced those bars is not obliged to produce similar ones.

This distinction is not philosophical. It changes what you should do with the result. If a backtest is a forecast, you trade it. If it is a measurement, you use it to reject bad ideas quickly and to estimate the size of the risk on the ones that survive.

Cause one: the costs you did not include

This is the most common and the most embarrassing, because it is entirely self-inflicted. A backtest that fills you at the exact close of the signal bar, with zero spread and zero commission, is not a conservative estimate — it is a different strategy from the one you will trade.

The arithmetic is brutal at short horizons. If your average winning trade is 0.20% of price and your round-trip cost is 0.05%, you have handed back a quarter of your edge before anything else goes wrong. Tighten the timeframe and the ratio gets worse, because the edge shrinks faster than the cost does.

• Add the real spread of your broker at the hour you actually trade, not a default pip value.
• Add commission per lot, both sides.
• Add a slippage estimate: at minimum one tick on entries, more around news.
• Re-run and compare. If the strategy only lives at zero cost, it does not live.

Cause two: fills that assume a price you could not get

A backtest decides on the close of a bar and, very often, buys at that same close. In live trading the bar has already closed before you know the signal, so the earliest realistic fill is the next bar's open — which may be nowhere near the price your equity curve used.

The gentle version of this mistake costs a fraction of a bar of movement. The severe version happens on breakouts and around scheduled data, where the next open can be several times the normal bar range away. Strategies that specifically trade breakouts are the ones most damaged by honest fill assumptions.

Cause three: measuring the wrong population

If you chose this strategy by looking at a hundred candidates and keeping the best, your backtest result describes the best of a hundred attempts, not the expected behaviour of the idea. The same is true if the strategy came from a source that only publishes winners.

This is the mechanism behind multiple-testing corrections such as the Deflated Sharpe Ratio. The correction sounds exotic, but the intuition is plain: pick the luckiest coin out of a thousand and it will look like a coin with a bias. The larger the search, the less the winner's result means.

Cause four: position sizing that changed the shape of the curve

A backtest run at a constant one lot and a live account run at 'whatever the risk calculator says today' are two different equity curves. Sizing that grows with the account turns a modest drawdown into a severe one, because the losses happen at a larger size than the gains that preceded them.

The asymmetry is worth stating plainly, because it is the reason deep drawdowns are so much worse than they feel on paper.

Cause five: the market changed, and the rules did not

Some of the gap is not a mistake at all. Volatility regimes shift, correlations change, and a rule tuned to one environment behaves differently in another. You cannot eliminate this, but you can avoid being surprised by it: test the same rules on a second market and on a different era, and see whether anything faint survives.

If the strategy only works on one instrument over one period, you have not found an edge. You have found a description.

How big is the gap, typically

You do not have to guess. Run the same rules three times: once with zero costs and perfect fills, once with your broker's real costs, and once with costs plus a one-bar execution delay. The three equity curves tell you how much of your result was the strategy and how much was the assumptions.

For context on what normal risk looks like, the strategies in our own public library currently sit at a median maximum drawdown of 6.7%, with the 90th percentile at 19.1%. Those are drawn from the same records the strategy pages publish. If your own candidate shows a maximum drawdown far below that while also showing an unusually smooth curve, the most likely explanation is an assumption, not a discovery.

What to do with this

Treat a backtest as a filter rather than a promise. Use it to throw away the obviously broken and to size the survivors. Then run the survivor in simulation on live prices for a while, where the fills are real and the costs are real, before it sees any money.

None of this makes the outcome certain. It makes the uncertainty visible, which is the only part you actually control.