Sundial vs Laplace

Tsinghua (THUML) · Sundial see model card zero-shot one-step change series

Sundial is a generative time-series model; we draw samples per step and score the empirical predictive. Run raw only in this study; the collaborative arms were built for the three strongest models first.

Resources: GitHub · Model card · Paper (arXiv 2502.00816)

Live snapshot  Derived from the week-long round-robin study; counts grow as coverage deepens. Everything scores against Laplace on the identical series and windows.

Standalone

The model's own predictive, run zero-shot, scored per series against Laplace by a one-step-ahead (k=1) Diebold–Mariano test on the log-score differential.

stratumnwin / draw / loss vs Laplace med ΔLLCRPS ratio cov₀₀
economic change-series, business-daily1328-2.771.0780.64
economic change-series, weekly2720-2.561.0220.68
economic change-series, monthly (annual cycle)2720-3.021.0140.56
M4 hourly, strongly seasonal414-3.411.0880.52
asset prices and returns, daily1392-2.351.0950.59
beats Laplace draw loses to Laplace

Median per-series Δ log-likelihood in nats (negative is worse than Laplace); CRPS ratio to Laplace (above 1 is worse); raw central-90% coverage (0.90 target).

Star plot

Sundial standalone against Laplace, on the same six regime axes as the site's standalone radar. Each radius is the log-likelihood ratio, (wins + ½·ties) / n scaled so an even split with Laplace sits on the dashed 1.0 ring; outward beats Laplace more often, inward less. The M4-hourly set splits into soft and hard waveforms by corpus order, matching that radar.

economic (daily)weekly cyclesyearly cyclessoft waveformshard waveformsprice / returnsLaplace = 1.0economic (daily): log-likelihood ratio 0.03 vs Laplace (n=1339)weekly cycles: log-likelihood ratio 0.13 vs Laplace (n=2760)yearly cycles: log-likelihood ratio 0.02 vs Laplace (n=2760)soft waveforms: log-likelihood ratio 0.00 vs Laplace (n=180)hard waveforms: log-likelihood ratio 0.00 vs Laplace (n=234)price / returns: log-likelihood ratio 0.00 vs Laplace (n=1421)

Collaborative use

The recalibration (@lap) and portfolio (&lap) arms were built for the three strongest models first, so Sundial is scored standalone here. The sidecar wraps any per-step predictive, so these arms can be added without retraining. See the sidecar pattern.

Protocol

Fixed 128-length context, rolling one-step test window, no fitting, each model in its own environment. Strata split the cached FRED universe and the M4-hourly set by frequency and regime. Full method on the sidecar pattern page and in the methodology.

Architecture and methodology

Sundial is a decoder-only transformer trained without discrete tokenisation. It embeds continuous-valued patches and learns a generative TimeFlow head by flow-matching, so instead of a parametric density it produces samples of plausible futures. This study draws those samples (the base-128m model) and scores their empirical predictive; it is pretrained on the trillion-point TimeBench corpus.