IC Dispersion is now defined on the leaderboard, and the data package ships what it needs

The IC Dispersion column on the leaderboard used to be blank for most signals, and where it showed a number, that number did not describe your signal. As of 8 September the column is filled for every signal, and the same statistic can be reproduced locally with the updated data package.

What the metric asks. Is your signal's IC higher when the cross-section is moving? Its σ was always meant to be the cross-sectional dispersion of returns. In the earlier implementation σ was taken from the target after its per-hour standardization, and a standardized cross-section has the same dispersion every hour. A correlation with a constant is undefined, so the column stayed empty; the few values that did appear came from an edge case in the data, not from the signals.

The definition now. σ(i) is the population standard deviation, across the tickers, of the forward return over the target's window starting at hour i, taken before the standardization. It is the standardization's own denominator. IC Dispersion is the correlation over time between your hourly IC and σ, over the hours where both are realized. The reading is unchanged: positive means your IC is higher in dispersing markets, negative the opposite, near zero means uniform performance across regimes. The full definition is on the Metrics page.

What it means for standings. Nothing. IC Dispersion is a diagnostic column, shown and sortable; it does not enter the ranking or the quality criteria. Comparisons made on the old column should be discarded.

What to do. Download the updated data package. It ships σ for every training period as data/train/.small.dispersion.parquet, and the walk-forward runner computes IC Dispersion from it, so your local number uses the board's arithmetic on the board's values over the periods you hold. As with every metric, the board's number also covers history you do not receive, so expect the two to differ rather than match. A runner from the previous package prints NaN for this metric together with a note saying the file is missing: that is the old package, not your signal.