Skip to main content
SynForecast produces the standard Nixtla long format: unique_id identifies a series, ds identifies time, and y contains the target. Materialized SynForecast data can therefore be passed directly to NeuralForecast, MLForecast, or StatsForecast. The executable guides compare three training regimes against the same temporal holdout:
  • Observed only: the baseline that synthetic workflows must beat.
  • Observed + synthetic: SynAugment is fitted to the training split and adds statistically matched histories.
  • Synthetic only: the forecasting model is fitted without observed target values. The exact transfer mechanism depends on the model family.

​
Integration guides

  • NeuralForecast: zero-shot synthetic pretraining and synthetic pretraining followed by fine-tuning.
  • MLForecast: global training with augmentation and zero-shot transfer through predict(new_df=...).
  • StatsForecast: local-model behavior and a synthetic-history ensemble instead of transferable pretraining.
All three use the classic Box–Jenkins airline passenger series, reserve its last 12 months before generation, report holdout MAE, and save their forecast plots. They are workflow demonstrations rather than evidence that synthetic data always improves accuracy.

​
Materialize synthetic data

Fit augmentation parameters on training observations only. Using validation or test values would leak future information into the generated training set. Synthetic data also inherits its generators’ assumptions, so retain an observed-only baseline and evaluate multiple temporal folds and random seeds. These examples materialize data before model fitting. On-the-fly training-data generation would require a streaming data-source interface in the forecasting library and is not part of the current integration.