Why Backtest Results Don't Match Live Trading (And What Actually Causes the Gap)

In a backtest, every order is filled at the exact price the candle closed at, or the exact tick your stop hit. In live trading, that almost never happens.

Slippage is the difference between the price you expected and the price you got. In quiet markets, it might be a fraction of a pip. During high-impact news, low liquidity sessions, or fast-moving momentum, it can easily be 5, 10, or 20 pips on forex, or several dollars on an equity or crypto contract. If your backtest didn't account for realistic slippage, your edge was smaller than you thought before you even took the first live trade.

Spread widening works the same way. Most backtesting platforms apply a fixed spread — say, 0.5 pips on EUR/USD. But during the London-New York overlap or rollover, spreads can double or triple. If your strategy relies on tight profit targets (like a 5-pip scalp), a 1.5-pip spread eats 30% of your gross win. In the backtest, it looked like free money. In live execution, it's a slow bleed.

Overfitting — also called curve-fitting — happens when you keep tweaking indicators, thresholds, and filters until the backtest produces the highest possible return on that specific slice of history. The problem: you haven't built a strategy that predicts the future; you've built a description of the past.

A strategy with an EMA(20) cross is simple. A strategy with an EMA(103) that only enters on Tuesdays when the RSI is between 42 and 47 and volume is above the 14-day median is almost certainly curve-fitted. It will look incredible on the data you trained it on, and it will fail almost immediately in live conditions because the market will never reproduce that exact combination of noise again.

Look-ahead bias occurs when your backtesting logic accidentally uses information that wasn't available at the time the trade was supposed to trigger. This can happen in subtle ways: using indicator calculations that repaint after the bar closes, referencing the session's high or low before the session ends, or using split-adjusted data that wasn't known in real time.

Data snooping is related: if you test 50 variations of a strategy and pick the best one, standard statistics say at least a few will look profitable purely by chance. If you then trade that 'best' version live, you're trading a statistical artifact, not an edge.

Commissions, swap fees, exchange financing rates, and platform latency are rarely modeled accurately in retail backtesting software. A strategy taking 40 trades a month with a $0.50 difference per side in commission costs $480 more per year per contract than the backtest assumed. Over time, friction compounds just like profits do.

Latency matters too. If your entry signal fires and it takes 200 milliseconds for your order to reach the broker's server, other market participants with faster connections or co-located servers will take the liquidity ahead of you — especially on breakout or momentum strategies. You get filled worse, later, or not at all.

Markets aren't static. A strategy that crushes a trending market between 2020 and 2021 will likely bleed capital during a choppy, range-bound 2023. If your backtest period was dominated by one type of environment (like a multi-year bull run or zero-interest-rate regime), your live results will suffer the moment the regime shifts.

This isn't an execution failure — it's an adaptation failure. The strategy wasn't broken; it was simply tested under conditions that no longer exist.