Every number here is scored on a held-out window whose horizon has already fully elapsed. The forecast on the chart starts at the newest bar, so nothing on it has happened yet and it cannot be scored — this is a separate run with an origin far enough back that the whole horizon is known. The metric is mean 9-level pinball loss divided by a naive baseline (flat last close, widened by the empirical distribution of step changes). Below 1.000 beats naive.
Per-subnet ratios vary far more than the aggregate. A subnet can beat naive on one timeframe and lose on another; both are shown rather than averaged away. Click a heading to sort.