Release-event model: channels capped at the any-channel chance; stratified calibration and competition scaling tested (rejected); slots analysis; docs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-11-20 01:33:26 +00:00
co-authored by Claude Opus 5.5
parent a3b225fa48
commit c61800a51f
2 changed files with 116 additions and 12 deletions
+18
View File
@@ -62,3 +62,21 @@ memory down. Switching it on for the site was reverted after the service hit its
Also found: per-channel probabilities above ~0.2 are overconfident (40-50% predicted, ~8% observed),
mostly P4 regions for artists already in rotation (Mumle: 44% on P4 vs 6% any channel). A floor making
"any" >= the best station group made "any" worse (0.0664 -> 0.0690) until those are recalibrated.
## Calibration, competition and slots (2026-11-20), history on
- **Site**: history on (`AIRPLAY_HISTORY=1` for airplay-web and airplay-collect); fitting peaks at ~2.5 GB,
so both services have `MemoryMax=3584M` (drop-in history.conf).
- **Per-channel calibration by station group x in rotation (isotonic)**: no gain (release events per-channel
log loss 0.0150 vs 0.0148 pooled Platt); the 50-60% bin still came true 0 of 16 times. P4's regions repeat
the same few events, so the tail cannot be learned.
- **Cap each channel at the any-channel chance**: release events per-channel PR-AUC 0.105 -> 0.165, log loss
0.0148 -> 0.0143, "any" unchanged; general model neutral to slightly worse (PR-AUC 0.285 -> 0.267).
Applied to the release-event model only (the site's "if they release" numbers: Mumle 44% on P4 vs 6% any).
- **Competition scaling** (each week, scale a channel's chances to its usual number of adds): worse everywhere
(any 0.0736 -> 0.0760). Over 8-week windows actual adds rise with predicted supply (1.4-1.7 per predicted
add, correlation 0.72-0.82 at h=4): no fixed capacity to share. Rejected.
- **Slots per channel** (Poisson deviance on 26 weeks May-Nov 2026, five main channels): flat recent average
232, week-of-year average 833, recent level x shrunk calendar blocks 217. Blocks: summer holiday (weeks 28-31:
P3 0.66, P4 0.77, P5 0.83), New Year (weeks 1-2), December (not trustworthy until song-level data separates
Christmas music). Not yet in the model.