Forecast model v2.2
-
The weekly forecast no longer flatlines after an idle stretch. When the
recent snapshot window was a plateau (common for the slow-moving weekly
gauge), the Monte Carlo's rate draw could degenerate and the forecast
modal showed a single flat line with a zero-width interval ("80% CI
49%-49%"). The per-path rate is now drawn from a rectified Gaussian
instead of a moment-matched Gamma: "no further growth" remains a possible
outcome with honest probability, while the fan stays open according to the
historical rate variance. Path increments are unchanged (still the v2.0
Gamma process), so trajectories remain monotone. -
Better-calibrated intervals across the board. On the real-data
benchmark (LOSO replay), 80% CI coverage moves from 55% to 89% on the
session gauge and from 73% to 74% on the weekly gauge, with CRPS slightly
improved on both - the wider coverage is earned, not bought with a blanket
widening. -
Spec updated. MODEL.tex and MODEL.pdf describe the new rate law, the
v2.1 spec is archived underinternal/forecast/archive/, and the forecast
CHANGELOG has the full rationale. New regression tests pin the plateau
scenario at both the sampler and the store-to-payload integration level.