Challengers

Everyone who has stepped up, and how it went.

skaters cooperates as well as it competes: sandwiched in laplace's coordinates, every challenger below improves dramatically (the sandwich), so a bare horserace can mislead. These pages exist as counterpoint to overly aggressive marketing in autonomous univariate distributional prediction.

Featured bouts

TabFM — David v Godzilla

A 12 GB tabular foundation model, pre-registered, eleven arms, every one behind laplace on likelihood, at one ten-thousandth the footprint. The bout →

Prophet

Three scales, 30 to 921 series, pre-registered at the largest. Raw: 4 wins in 921, and −11 nats median on repeat-heavy series. The tape →

The record

Numbers are per-series win-rates for laplace, log-likelihood first, CRPS second. Everyone is converted to the same Dist and scored by the same code on the same held-out points. Every row links its tape. The one belt we do not hold is at the bottom, stated rather than averaged away.

challengercornerLLCRPStape
AutoARIMA (statsforecast)classical82%53%study
auto.arima (real R)classical79%51%bout
AutoETSclassical97%83%study
ETS / SARIMAX (statsmodels)classical96% / 87%81% / 53%study
Theta (the M3 winner, R)classical87%63%bout
ADAM (smooth, R)state space96%82%bout
BSTS (full Bayesian posterior, R)state space97%80%bout
nnetar (neural AR, R)neural96%76%bout
NF-StudentT (NeuralForecast)neural100%78%study
CSP (incl. adaptive)conformal99%100%bout
AutoARIMA + conformal / ACIconformal86% / 88%31% / 29%study
TimesFM 2.5, 200M (Google)foundation100% cont.72% cont.David v Goliath
Chronos-Bolt (Amazon)foundation100% cont.88% cont.study
Moirai (Salesforce)foundation97% cont.93% cont.study
Lag-Llamafoundation99% cont.94% cont.study
TabFM 1.0 (Google, tabular)foundation71–88% (11 arms)38–60%David v Godzilla
Prophet (Meta)calendar GAM99.6%study
GARCH-t, non-price seriesheavy-tail SOTA68–82%53–54%bout
GARCH-t, price/return seriesheavy-tail SOTAthe belt is theirs; we recommend it therestudy

Featured bouts

David v Goliath →

Laplace vs TimesFM, Google’s 200M-parameter foundation model: 69 of 69 continuous series on likelihood, a 2.26-nat median gap, the sorted-gap chart, the footprint table, and where Goliath actually wins (predicting zeros).

One bout page so far. Challengers that earn a full treatment get one; the next candidates are the native-density foundation models, Moirai and Lag-Llama, which came closest. Rematch conditions for TimesFM are tracked in skaters#97.

Send a challenger

The protocol is packaged and copyable: the benchmark-against-laplace skill for any distributional forecaster, and the timesfm-study skill for foundation models specifically. Same Dist, same code, same held-out points, both metrics, splits disclosed. Run it, and if your method wins, open an issue with the tape; the table takes new rows in either direction.

New bouts are pre-registered: the protocol, parameters and analysis plan are committed to benchmarks/preregistrations before the results are read, with everything already observed at filing disclosed. The first filed statement is the TabFM bout.