Excess above noise: the signal quality floor

1 · The problem with a bare IC threshold

Information coefficient — the rank correlation between predictions and realised forward returns — is our primary screening statistic. The temptation is to gate on a fixed threshold: promote everything above +0.02, say. But an IC is an estimate with a sampling distribution, and its standard error varies enormously across cells: a coin with two years of clean non-overlapping bars pins its IC down tightly; a coin with six months of patchy history does not. A fixed threshold passes lucky short-history cells and fails unlucky long-history ones — exactly backwards.

2 · The definition

For each (model, coin) cell we compute the per-bar IC series on h-stride non-overlapping windows (overlapping windows understate the standard error — see the overlap bug in Paper № 07), estimate SE from that series, and define:

excess_ic = |IC| − SE(IC)

A cell with excess IC ≤ 0 is statistically indistinguishable from noise regardless of its headline IC, and is treated as such: it cannot pass screening, it cannot enter an ensemble, and it does not count as evidence that its architecture “works”.

3 · What the floor did in practice

Applied retroactively to a full screening grid, the excess-IC floor cut roughly half of the cells that a bare positive-IC gate would have passed. The cells it removed were concentrated exactly where theory predicts: short-history coins, high-volatility names with fat-tailed per-bar ICs, and cells whose headline IC was carried by one regime. The cells it kept went on to have a visibly better survival rate at the later, more expensive gates of the promotion funnel (Paper № 11).

The deeper point is about multiple testing. A grid of hundreds of cells will produce dozens of positive ICs by chance; requiring each IC to clear its own standard error is the cheapest possible first defence — it prices every cell's evidence in that cell's own units before any cross-cell selection happens.