Leakage is why most backtests look good

A pipeline that can see the future will look excellent on historical data and lose money in production. This is the most reliable failure in quantitative work, and it is almost always invisible from the results.

Crossview is held to a set of rules that make leakage hard to introduce by accident.

Every feature uses a trailing window

Nothing sees the bar it is scoring. A 30 day volatility is computed from the 30 days before the decision point. A trailing window is not a stylistic choice, it is the difference between a measurement and a memory.

The universe is what was observable

Names qualify from what was listed and liquid before the window, not from what survived to today. Screening on present-day liquidity quietly selects the winners and flatters everything downstream.

Published artefacts are immutable

A dated build is never rewritten in place. Successive builds accumulate as separate dated folders. If yesterday's numbers change, that is a bug, not an update.

Models are frozen between refits

The weekly model is trained on data older than the replay window and then applied unchanged to every snapshot in that window. Refitting per snapshot would let the model absorb the period it is being scored on.

What the discipline costs

The headline numbers get worse. A leaking pipeline will show a smoother equity curve than an honest one, every time, and the honest pipeline looks like the weaker product on a slide.

How you check it without trusting us

Model manifests carry their holdout metrics, including the walk-forward information coefficient. The standing IC report runs as a model health check and is visible in the terminal. When a model's edge is weak, that is recorded rather than smoothed over. One recent replay measured a rank correlation between forecast and realised 24 hour move of about minus 0.065, which is not an edge. Anyone quoting a better number from the same replay should be asked how it was computed.