The verdict first — 34 blind seasons, graded
Each bar is one full season the model predicted before seeing any of it, for whatever view is selected on the left (a refuge, a region, or all 41 refuges pooled, hunter-weighted). Above zero: the model beat the naive baseline that season. Click a bar to load that season below and see the day-by-day curves behind it.
Walk-forward: predicted vs actual
The skill grid — where the model wins and loses
every refuge × season with ≥200 hunter-days · click a cell to load it aboveThe effect dials — build a day, ask the posterior
posterior-mean effects from the fitted Bayesian model · applied to the selected view's own average dayMultipliers combine the fitted posterior means: exp(Σ β·z) against a reference day (average conditions, clear sky, north wind, regular shoot day). The effects themselves come from the one model fit on everything — what follows your selection is the starting rate: the readout applies the multiplier to the hunter-weighted average of whatever refuge/region and seasons you've picked above. Amber dials are the water block — a real in-sample association that added no out-of-sample skill; they're shown so you can feel the size of the paradox, not so you can forecast with them. All effects associational, per the paper.