Why Most Backtested Strategies Fail in Live Trading
The five gaps between a good backtest and a good strategy — and which gate catches each one.
Look-ahead bias
The most common and least visible failure: a signal that uses information not actually available at the moment of the trade — an indicator computed on a bar that hasn't closed yet, or a fill priced at a bar's own open when the decision was made mid-bar. The result back-tests beautifully because it is quietly cheating.
The engine enforces signal-on-close, fill-on-next-open at the execution layer, not as an optional setting — there is no configuration that reintroduces look-ahead bias by accident.
Curve-fitting to one dataset
A strategy tuned until it produces a smooth equity curve on one specific date range has been fitted to that range's noise, not to a real market inefficiency. It fails live because live markets don't replay the same noise.
Walk-forward analysis and the PBO gate exist specifically to catch this — see the walk-forward and PBO pages for how each one measures it.
Ignoring execution cost
A strategy that trades frequently can look profitable with zero fees and instant fills, then go negative the moment slippage and taker fees are applied. This is especially common with mean-reversion strategies on tight timeframes, where the edge per trade is small enough that costs eat all of it.
Realistic fills (0.05% entry / 0.10% exit slippage, 0.075% taker / 0.020% maker fees by default) are on for every backtest here, not an opt-in toggle.
If a strategy only looks profitable with fees and slippage set to zero, that is not an edge — it is a fee-sensitivity problem wearing an edge costume.
Regime change
A strategy validated entirely on a trending market (or entirely on a ranging one) has been validated against one regime, not against "the market." When conditions flip, the strategy's whole premise can stop applying — this isn't overfitting in the statistical sense, it's a scope problem: the backtest window simply never contained the regime that later broke the strategy.
There is no gate that fully solves this — no amount of historical testing guarantees the future contains the same regimes as the past. The honest mitigation is a long enough backtest window to span multiple regimes, and treating a strategy's validated period as a stated assumption, not a guarantee.
Small-sample luck
A strategy with eight winning trades can look extraordinary and mean almost nothing — eight trades is not enough data for walk-forward, Monte Carlo, or PBO to say anything statistically meaningful, even though the platform will still show you the real numbers rather than blocking them outright. Treat a thin sample as an early read, not a verdict, until the trade count grows.