Six defects none of the summary numbers showed

A paper on one intraday US equity system makes a narrow claim and supports it with specifics. Backtests of intraday strategies can drift away from what the live system actually does, and in this case the drift never showed in the summary statistics. The authors report six defects found by running the system live and comparing each live day against its backtest replay.

Three of the six came from data. A default feed carried only a small fraction of real opening volume. Historical prices were adjusted for dividends, which live orders never see. And the live system downloaded data the backtest never did.

Two came from code and one from configuration. A sizing step exceeded a limit that the live account enforces. UTC timestamps were read as New York time. And settings that were switched on never reached the code that places the orders.

Each of the six is the kind of defect that leaves the aggregate numbers looking reasonable. A feed that under-reports opening volume changes what the strategy appears able to trade. Prices adjusted for dividends move levels a live order would never have seen. Timestamps read in the wrong zone shift the whole session. The authors' point is that such differences are visible in a day-by-day comparison against the replay and not in a summary.

What makes the paper useful rather than anecdotal is its structure. For each defect the authors give the symptom, the size of the error, the cause, and the test or control that now guards against it, and they close with a ten-point checklist. They also name the shape all six share: each sat where two paths that should agree had drifted apart. They state explicitly that they make no claim about whether the system is profitable, which is the right place to leave it.