The First Pattern v2.2: A Candidate Law-Like Regularity Underlying Generalist Time Series Forecasting

This record contains a preprint presenting The First Pattern (TFP) v2.2, a theory-driven generalist time series forecaster evaluated across 11 diverse domains using a single frozen configuration.

TFP is tested against widely used statistical baselines (including SimpleTheta and Naive2) across domains spanning epidemiological forecasting, technology adoption, energy load, retail benchmarks, web traffic, and finance. The study reports a consistent advantage versus SimpleTheta in the primary cross-domain summary metric, and documents clear boundary conditions where seasonal and benchmark-tuned baselines remain difficult to beat.

Model and hypothesis framing. TFP’s design is organized around a four-stage Information Emergence Cycle (IEC): Q1 (Potential), Q2 (Selection), Q3 (Transformation), and Q4 (Propagation). The paper frames “law-like regularity” as a falsifiable hypothesis: (1) a single configuration should generalize across diverse domains, (2) performance patterns should be mechanistically interpretable, and (3) specific regimes should defeat it.

Synthetic tests. A controlled synthetic benchmark is included to probe when the observed advantage aligns with S-curve dynamics. Results show a short-horizon edge on logistic growth and logistic saturation, while trend and seasonal families favor Theta-style baselines, and longer horizons reduce or reverse the logistic advantage.

Reproducibility and limitations. Key limitations are stated directly, including that comparisons emphasize classical statistical baselines rather than neural SOTA, and that some domains remain challenging for any generalist method. The manuscript includes rolling-origin protocols, domain-by-domain breakdowns, and reproducibility details. Code for TFP is available from the author on request, and supplementary materials and related resources are linked in the manuscript.