The studies I'd show a risk desk: failures included.

The question: A multi-timeframe trend system had validated beautifully: a 10-instrument portfolio showing +0.185R per trade over ~3,000 trades, a 57.6% win rate, eight instruments cleared for deployment. Before capital went anywhere near it, one question had to survive an audit: does the backtest know anything at decision time that the live system couldn't have known?

The method: Line-by-line reconstruction of what information each decision could legally see. The bug, once found, was quiet: the higher-timeframe trend windows were sliced by bar open time, which admitted the still-forming daily bar (carrying its final close, up to ~23 hours of future information) into the trend filter on every single evaluation. Because a daily bar's eventual close correlates with the outcome of a breakout taken during that day, the filter was subtly grading entries with the answer key.

The result: The corrected backtest collapsed the portfolio edge from +0.185R to +0.053R; roughly 70% of the claimed edge was the bias. Win rate fell from 57.6% to 51.9%. Of ten instruments, exactly one survived the acceptance bar.

The step most backtests never get: a control run. The old, biased slicing was re-emulated bit-for-bit on identical fresh data and reproduced the original results to within a few hundredths of an R, proving the collapse was caused by the bias fix, not by a shifted data window. That's the difference between "the numbers changed" and "the cause is isolated."

What it became: The invalidated benchmark was retired in writing so it could never be used as a target again. The surviving instrument was rebuilt around what a follow-up exit study proved actually carried the edge, revalidated on the corrected engine, and is the system that trades live today. The bias class itself became a permanent item on my audit checklist, including checking that the live scanner's bar handling matches the corrected backtest's semantics exactly, because a backtest and a live feed can disagree about time in both directions.

The question: The EURCHF mean-reversion system had already passed a full validation: positive every year in the test window, confirmed on a held-out year, statistically significant. A later strategy review asked a question the original 288-configuration grid search structurally couldn't ask: the design fades dislocations in both directions by construction, but is the edge actually symmetric?

The method: No new backtest, no new parameters: just re-cutting the existing validated trade log by direction. That matters: when a split is decided after seeing the data, the honest protocol is to treat the result as a hypothesis, demand year-by-year sign consistency, and plan from the conservative end of a bootstrap interval rather than the point estimate. All of that is in the write-up.

The result: The split was not subtle. Buying the downside dislocations: +0.392R per trade across 192 trades, positive in all six years, including the low-volatility 2026 patch that drags the combined system toward zero. Fading the upside: −0.062R across 126 trades, negative in four of six years. The system's entire edge lives on one side of the book; the other side has been quietly taxing it the whole time. The economic story is coherent: the two directions face different structural flows in that pair, which is what separates a finding from a fluke.

What it became: Not an instant "upgrade." One-siding the book concentrates exposure to the pair's known catastrophic scenario (a central-bank regime break of the kind that pair has produced before), so the finding is coupled, non-negotiably, to event blackouts and reduced sizing before it ships. A bigger lesson got written into the process: grids explore the axes a template exposes and silently ignore the ones it doesn't. Direction, regime admission, stop logic: enumerate the frozen axes before refining the searched ones. The money is usually in an axis nobody varied.