Time Series Prior for Tabular data FM

[thomaspdl]

No description provided.

[thomaspdl]

Description

This PR introduces time series priors to TFM-Playground, enabling pretraining on synthetic data with temporal patterns to improve model performance on forecasting downstream tasks.
A more detailed report is here: https://docs.google.com/document/d/1MCXC9M6t6hPChPQY9ub_CwukymEby-zIxagpfqTMqZc/edit?tab=t.0#heading=h.ua81y23aj4x5

Motivation

Existing priors (TabICL, TICL) generate i.i.d. tabular data with no temporal structure. This is suboptimal for forecasting tasks where data has trends, seasonality, and autocorrelation. Our time series priors expose the model to these patterns during pretraining.

What's Added

New module: tfmplayground/priors/timeseries/
TemporalXSampler: Generates features with trends, seasonality, AR processes, and random walks
TimeSeriesSCM: Creates (X, y) pairs respecting temporal causality with lag features
ForecastPriorDataset: Batch generation with temporal train/test splits
config.py: Hyperparameters and presets

New scripts:

pretrain_forecasting.py: Training script for time series priors
eval_forecasting.py: Evaluation script comparing trained vs baseline

Tests:

tests/test_timeseries_priors.py: Unit tests for all components

Documentation:

Updated README with usage instructions
Design document in docs/design/

Usage

Train

python pretrain_forecasting.py --epochs 100 --steps 50 --priortype mixed

Evaluate

python eval_forecasting.py --model nanotabpfn_forecasting_weights.pth

[borawhocodess]

the google doc link is not public (requires request to access)