Backtesting that survives live markets

Almost every strategy looks profitable in a backtest. That is the problem. The gap between a beautiful equity curve in simulation and a flat or negative one in production is where most systematic trading efforts quietly fail. A backtest that survives contact with live markets is a different, far more demanding artefact than one that simply looks good.

§01Lookahead is the silent killer

The most common flaw is lookahead bias: the backtest uses information that would not have been available at decision time. It can hide in something as innocent as using a bar’s close to make a decision inside that bar, or aligning a signal to data that only arrived later. A single microsecond of accidental foresight can turn a losing strategy into an apparent winner. Event-driven replay, where the simulator only ever sees what the live system would have seen at that instant, is the defence.

§02Model the queue, not just the price

A naive backtest assumes you get filled whenever the price touches your level. Reality is a queue: at a given price you sit behind existing size, and you only fill when everything ahead of you does. Ignoring queue position massively overstates passive fill rates and flatters any market-making or limit-order strategy. Realistic simulation replays full order-book depth and tracks where your order would actually stand.

§03Cost honesty

Fees, rebates, borrow costs, and — above all — market impact must be modelled honestly. A strategy that trades often is exquisitely sensitive to per-trade cost; shaving a basis point off assumed slippage can flip the whole result. If anything, a good backtest is pessimistic about costs, because live markets rarely give you the best case.

§04Out-of-sample discipline

Finally, a backtest is only as trustworthy as the discipline around it. Tuning parameters until the curve looks perfect is curve-fitting to noise. Reserving genuinely out-of-sample data, limiting the number of trials, and being suspicious of results that are too clean are what separate research from wishful thinking. Simulated performance has inherent limitations and is not indicative of future results; this article is educational only.