Files
Airplay-Predictor/docs/LONG.md
T

7.9 KiB
Raw Blame History

The long any-DR model (2019–)

Total DR plays per artist and week (all channels, P4 regions included). An add is τ = 20 or more plays in a week of the next 2 week(s) after 8 weeks under τ, and τ again in at least 2 of the 3 weeks after it (it stays on). Archive era 2019-04-08 to 2024-09-09 (15538 artists observed, 979721 qualifying rows, 1432 adds); current panel 2025-11-17 to 2026-11-02 (293252 rows, 293 adds).

Models: 1-year trains on the 52 weeks of origins before the test period (what the site can do from the current panel alone); all years on every earlier origin; + older history adds four features from the artist's weeks before the last 52 (for today's predictions: the 2019–24 archive); all years + era adds an archive-era dummy, switched off when predicting (the archive over-represents popular artists, so its base rate is higher). Training rows' labels (window plus the weeks checked after it) end before the first test origin, so no label is seen early. Δ < 0 means better than 1-year.

Sanity check (2026-11-17)

The first version of this report counted any week at τ as an add. Checking the current panel showed most of those were blips:

  • The typical "add" at τ = 20 was 2 plays on national P4, repeated on each regional station.
  • 58% fell back under τ in each of the next 4 weeks; at τ = 50 it was 56%.

The target now requires the artist to stay on: τ again in at least 2 of the 3 weeks after. Training labels end before the first test origin, including those 3 weeks.

Other checks:

  • Archive pages that list only weeks with plays (13,277 captures) look genuine. They're new or rare artists, e.g. a debut going 1 → 17 → 45 plays. So unlisted weeks are counted as 0 plays.
  • Scale between the periods: the archive's weekly totals match the current panel with P4 regions summed (99th percentile 136 vs 213, against 45 without the regions). The archive runs somewhat lower, which the era dummy absorbs.
  • Leaks: features only use weeks up to t, except the release and New Music Friday flags for the window. Those are scenario inputs, as on the site. Chart positions use the chart published a week before t.
  • Known biases:
    • Discographies (Deezer) and NMF Denmark lookups exist only for today's pool, so in the archive the "discography known" and release flags favour artists who were still played in 2025–26. That flatters the archive years; it doesn't change the comparison within a year.
    • The archive also over-represents artists whose pages got crawled.
    • Chart matching here uses primary names only (the site also uses aliases).
  • Power: the current-panel test has 132 sustained adds. The log-loss intervals rule out gains larger than about 3–5% of log loss. Differences in average precision (0.035–0.046) are within noise.
  • Skill: adds are rare (0.09–0.25% of rows). An average precision of 0.04–0.19 ranks real adds 40–75 times better than chance, though Brier skill stays near zero.

Findings

  • With sustained adds, all years beats 1-year on the current panel by 0.0001 in log loss (90% CI −0.0003 to −0.0001, about 2%). But it's worse calibrated (mean p 0.0011 vs a rate of 0.0009) and has lower Brier skill.
  • In the archive years it's a wash: −0.0002 in 2023, +0.0002 in 2024, 0 elsewhere.
  • Older-history features: within ±0.0001 for today.
  • τ = 50 shows the same: no reliable gain.
  • More years and older history help little or not at all. The one-year limit is not the bottleneck; better release and promo timing is the more promising lever.

Out of sample

test period rows adds model log loss Brier skill avg precision mean p / rate Δ log loss vs 1-year (90% CI)
2021 314613 392 1-year 0.0072 +0.012 0.054 0.0018 / 0.0012 —
2021 314613 392 all years 0.0072 +0.008 0.053 0.0018 / 0.0012 +0.0000 (+0.0000 to +0.0000)
2021 314613 392 all years + era 0.0072 +0.008 0.053 0.0018 / 0.0012 +0.0000 (+0.0000 to +0.0000)
2021 314613 392 1-year + older history 0.0071 +0.024 0.054 0.0014 / 0.0012 -0.0001 (-0.0002 to -0.0000)
2021 314613 392 all years + era + older history 0.0071 +0.021 0.053 0.0014 / 0.0012 -0.0001 (-0.0002 to -0.0000)
2022 135465 263 1-year 0.0108 +0.004 0.050 0.0022 / 0.0019 —
2022 135465 263 all years 0.0108 +0.015 0.051 0.0024 / 0.0019 +0.0000 (-0.0002 to +0.0001)
2022 135465 263 all years + era 0.0108 +0.015 0.051 0.0024 / 0.0019 +0.0000 (-0.0002 to +0.0001)
2022 135465 263 1-year + older history 0.0108 -0.015 0.048 0.0022 / 0.0019 -0.0000 (-0.0002 to +0.0001)
2022 135465 263 all years + era + older history 0.0107 +0.013 0.049 0.0021 / 0.0019 -0.0001 (-0.0003 to +0.0000)
2023 126845 204 1-year 0.0079 +0.056 0.106 0.0015 / 0.0016 —
2023 126845 204 all years 0.0077 +0.069 0.124 0.0015 / 0.0016 -0.0002 (-0.0003 to -0.0000)
2023 126845 204 all years + era 0.0077 +0.069 0.124 0.0015 / 0.0016 -0.0002 (-0.0003 to -0.0000)
2023 126845 204 1-year + older history 0.0078 +0.059 0.108 0.0017 / 0.0016 -0.0001 (-0.0002 to +0.0000)
2023 126845 204 all years + era + older history 0.0077 +0.067 0.118 0.0016 / 0.0016 -0.0002 (-0.0003 to +0.0000)
2024 16092 41 1-year 0.0116 +0.090 0.185 0.0026 / 0.0025 —
2024 16092 41 all years 0.0118 +0.057 0.148 0.0026 / 0.0025 +0.0002 (+0.0001 to +0.0011)
2024 16092 41 all years + era 0.0118 +0.057 0.148 0.0026 / 0.0025 +0.0002 (+0.0001 to +0.0011)
2024 16092 41 1-year + older history 0.0114 +0.082 0.169 0.0028 / 0.0025 -0.0003 (-0.0006 to +0.0000)
2024 16092 41 all years + era + older history 0.0118 +0.055 0.135 0.0028 / 0.0025 +0.0002 (+0.0002 to +0.0010)
current panel, origins from 2026-05-18 150981 132 1-year 0.0053 +0.012 0.038 0.0007 / 0.0009 —
current panel, origins from 2026-05-18 150981 132 all years 0.0052 +0.003 0.046 0.0011 / 0.0009 -0.0001 (-0.0003 to -0.0001)
current panel, origins from 2026-05-18 150981 132 all years + era 0.0053 -0.005 0.046 0.0013 / 0.0009 -0.0000 (-0.0002 to -0.0001)
current panel, origins from 2026-05-18 150981 132 1-year + older history 0.0053 +0.007 0.035 0.0007 / 0.0009 -0.0000 (-0.0000 to +0.0000)
current panel, origins from 2026-05-18 150981 132 all years + era + older history 0.0054 -0.027 0.046 0.0016 / 0.0009 +0.0001 (-0.0002 to +0.0000)

Coefficients (all years + era + older history, every row)

feature β
intercept -9.607
log plays week t +1.493
log mean plays last 4w -0.240
log mean plays last 12w +0.653
log max weekly plays last 52w +0.333
on DR 9-25 weeks ago +0.323
on DR 26-52 weeks ago +0.159
weeks observed of last 52 (share) +1.014
weeks with plays of last 12 (share) -1.882
trend: last 4w vs the 8 before +0.506
discography known +0.577
release in last 4w +0.692
release 5-12w ago +0.294
release in the window +0.638
log releases last 52w -0.169
on Track Top-40 now +0.783
on Airplay Top-20 now +0.887
log Top-40 hits so far -0.059
log Airplay hits so far +0.246
new chart entry in last 4w +0.500
NMF data covers t -0.420
NMF Denmark in last 4w +1.052
NMF Denmark in the window +1.587
log NMF placements last 52w -0.150
Danish +0.451
group +0.207
female +0.602
season index of the window +1.000
archive era -0.032
older history: observed before the last 52w -0.666
older history: log max weekly plays +0.040
older history: log adds +0.192
older history: share of weeks on DR -0.265