Numbered working papers on how a backtest lies and how to stop it. No signals, no market commentary, no “X % per month”. Every paper ends in a setting you can change yourself.
What the agent runs, why its own trials count against your multiple-testing bar, and four documented weaknesses: it inherits the sample, it can be right about the statistics and wrong about the strategy, its autonomous checking spends your budget, and it refuses things you may actually want.
Keywords validation · agents · multiple testing · deflated Sharpe
What the agent runs, why its own trials count against your multiple-testing bar, and four documented weaknesses — sample dependence, unmeasurable artefacts, the budget its autonomous checking spends, and the things it refuses by policy.
Repainting indicators, bar-magnifier gaps, zero slippage and a 20k-bar ceiling: four reasons the Strategy Tester flatters your idea — and the exact settings that stop it.
One in-sample fit is a guess. Six rolling out-of-sample folds are evidence. How anchored and rolling walk-forward work, what the efficiency ratio tells you, and where the cut-off should be.
Why 20 bps is a sane default, when it is not enough, and how portfolio-level slippage changes the answer for a cross-sectional momentum book.
Why the number of trials matters more than the best result, and how Backcast Labs counts them for you.
Regimes cluster. Resampling that ignores it understates drawdowns by a third in our tests.
What is in the export file, what a runner should check before it accepts one, and how drift monitoring closes the loop.
Backcast Labs is in early access and not on sale yet. Working papers are published here; there is no mailing list to subscribe to.
Send us a short e-mail and we will let you know when a place opens.