CauKer: Classification Time Series Foundation Models Can Be Pretrained On Synthetic Data Only

Source

Core Claim

CauKer argues that classification time-series foundation models can be pretrained sample-efficiently on synthetic data generated from Gaussian-process kernel composition and structural causal models.

Key Contributions

  • Generates diverse, causally coherent synthetic time series with trend, seasonality, and nonlinear interaction structure.
  • Targets pretraining for classification TSFMs rather than forecasting alone.
  • Reports scaling laws over synthetic dataset size and model capacity.

Method Notes

CauKer is a bridge between Synthetic Data For Time Series, Causal Time Series, and Time-Series Foundation Models.

Evidence And Results

The abstract reports synthetic dataset scaling from 10K to 10M samples and model scaling from 1M to 783M parameters, with clearer scaling behavior than real-world datasets.

Limitations

The source focuses on classification TSFMs. It does not settle whether synthetic causal generation transfers to reasoning, forecasting, or high-fidelity generation tasks.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Data diversity, curriculum, and long tailpartially closesGenerates arbitrarily large classification pretraining data with GP kernels plus SCM structure and reports data/model scaling laws.Synthetic factors are hand-designed and classification-only; rare real regimes and interventions are not validated.
Augmentation-free or dataset-aware self-supervisionpartially closesPretrains Mantis and MOMENT-style encoders on synthetic causal-kernel series rather than relying on a real classification corpus.Does not cover forecasting, generation, editing, or action-conditioned objectives.
Causal and counterfactual structureadjacentUses DAG propagation and nonlinear edge functions to create causally coherent synthetic series.No logged actions, interventions, or counterfactual rollout benchmark.

Open Questions

  • Which synthetic causal assumptions produce robust out-of-distribution transfer?
  • Can CauKer-style data help reasoning models such as TimeOmni-1?