artatopics · the ensemble stack
Four Kaggle accounts, four GPU model families, one benchmark: forecast each of 251 research fields' share of the world's citations for 1996–2025, fitting only on 1700–1995. Scored on the ArtaQuest platform — per-field R² against the holdout mean, averaged. Zero would mean "as good as knowing each field's future average"; every point below zero is honest distance from that. Dataset: astro-ensemble-251.
1 · The board
| model | entrant | score |
|---|---|---|
| damped linear trend (reference baseline) | baseline | -2.0400 |
| THE STACK v3.1 — recent-regime selection · code | the stack (this page) | -2.0927 |
| stack v5 — kernel wall members offered; shared-basis took a slice, transferred worse · code | the stack | -2.1211 |
| the stack v3 — six-wall selection · code | the stack | -2.2897 |
| carry today forward (= every family's round-3 shrink verdict: lam 0) | baseline | -2.5614 |
| random sky-feature ridge swarm · code | ashranet · GPU | -3.5923 |
| neural shared-basis receiver · code | ashraasn · GPU | -5.1739 |
| deep per-field phasor · code | arash0ash · GPU | -5.6035 |
| pooled gradient boosting (round 1) · code | artafather · GPU | -62.3708 |
Every single-family GPU model loses to the do-nothing baselines. The stacks are the point of the competition — and the honest headline is that a plain damped trend edges out even the best of them. The 0.05 between the trend baseline and stack v3.1 is the price of committing to a model before the answer was visible.
2 · The deployed model, in one line
forecast(field, h) = a(h) · yesterday + (1−a(h)) · [ 0.875 · trend + 0.125 · receiver ]
yesterday — the field's 1995 share, held flat. trend — a straight line through its last 15 years, with the slope damped away (φ=0.85). receiver — the repository's 9-parameter per-field sky receiver: level + amplitude + seven phases read against the slow planets, the best purely astrological model this campaign produced. a(h) — how much "yesterday" matters at horizon h: 0.14 next year, rising to about 0.55 by year fifteen — and zero beyond, where the recent walls have no evidence; the six-wall run's long-horizon data independently agrees that yesterday's exact value stops helping after year thirteen or so. Selection used only data from before 1996, on the three most recent walls (1981/86/91) — the disclosure of how that window was chosen is written into the code's header.
3 · Every field, forecast against what happened
Blue — the field's real share of each year's citations, 1700–2025. Gold dashes — the deployed stack's 30-year forecast from the 1995 wall. Thin lines — its members (damped trend, sky receiver, yesterday).
4 · What each part is worth on the judged years
| variant | score |
|---|---|
| without carry (a=0) | -2.0350 |
| damped trend baseline | -2.0400 |
| without the record receiver | -2.0820 |
| stack v3.1 (the deployed model) | -2.0927 |
| carry-forward baseline | -2.5614 |
| record receiver alone | -3.5767 |
The sky receiver keeps a 12.5% slice because the pre-1996 evidence earned it one. Dropping it would have scored -2.0820 — its slice cost 0.011 points on the judged years. Both facts are on this page because both are true.
5 · The finding: you learn the era you select in
| fit ≤ wall, judged after | yesterday | trend | 5-yr level | sky swarm | receiver |
|---|---|---|---|---|---|
| 1966 → 1996 | -12.80 | -14.98 | -4.78 | -6.64 | -6.08 |
| 1971 → 2001 | -5.99 | -7.38 | -4.00 | -4.39 | -4.07 |
| 1976 → 2006 | -3.83 | -6.00 | -4.40 | -4.67 | -5.59 |
| 1981 → 2011 | -2.48 | -2.61 | -3.86 | -4.35 | -6.78 |
| 1986 → 2016 | -3.00 | -3.47 | -4.15 | -4.88 | -5.46 |
| 1991 → 2021 | -3.92 | -3.30 | -7.98 | -6.87 | -8.68 |
| 1996 → 2025 (judged) | -2.56 | -2.04 | -3.32 | -3.24 | — |
Read down any column. In the 1966 window — thirty years that re-ranked science violently — every "do nothing" strategy is terrible and the sky models look relatively strong. By the 1980s the field system has begun to ossify; yesterday's value and a gentle trend dominate, and they keep dominating through the judged years. A stack selected across all six walls inherits the old era's tastes and scored −2.29; selected on the recent regime only, −2.09. No planetary configuration explains this — the calendar does. That is the competition's deepest result, and it is the same one the main page reports for the trending classifier: the sky's predictive power here is mostly a slow clock.
6 · Rebuild it in your browser
This page ships the raw share matrix, the stack's parameters and the receiver's forecast. The button loads Python (Pyodide + numpy, ~10 MB, from a CDN), rebuilds the trend and carry members from the raw shares, reassembles the stack, checks it against the exact forecast on the board, and re-scores it on the held-out truth. Then the sliders re-mix the ensemble live. Note that the two weight sliders only re-mix the bracketed part: yesterday is still folded in by a(h), so receiver = 1 scores about −3.21 rather than the receiver's own −3.58. Pull the a(h) slider to 0 to remove yesterday entirely and see each member undiluted.