A backtest can be mathematically correct and still overstate what a live trading bot can achieve. The difference usually comes from information leakage, unrealistic execution or a strategy tuned too closely to one historical period.
Timestamp alignment, revised data and incomplete delisting history can quietly change a result. The test must reconstruct what the bot could actually have known.
Order size also matters. Historical midpoint prices do not guarantee that the desired quantity could have traded there.
Live deployment should begin with restricted API keys, small exposure, monitoring and an automatic shutdown condition. A backtest is evidence, not a performance guarantee.
Data can leak the future
Look-ahead bias occurs when a strategy uses information that was not available at the simulated decision time. Survivorship bias appears when the dataset contains only markets or assets that remained available.Timestamp alignment, revised data and incomplete delisting history can quietly change a result. The test must reconstruct what the bot could actually have known.
Execution has a cost
Fees, spread, slippage, funding, latency and partial fills should be modeled. A strategy that trades frequently may lose its entire theoretical advantage after realistic execution assumptions.Order size also matters. Historical midpoint prices do not guarantee that the desired quantity could have traded there.
Separate discovery from validation
Use different periods for strategy development and evaluation. Walk-forward tests and paper trading can expose sensitivity to changing regimes. Report drawdown and losing periods, not only total return.Live deployment should begin with restricted API keys, small exposure, monitoring and an automatic shutdown condition. A backtest is evidence, not a performance guarantee.