SynForecast artifact

Synthetic time series constitute a substantial share of the training corpora of several recent forecasting foundation models. Existing studies report benefits at small data budgets and modest synthetic proportions, but each evaluates individual combinations of model and generator. Studying these benefits systematically requires generators that support controlled configuration, perturbation, reproducibility and mechanism-level ablation. We review 11 existing generators and find gaps in their combined support for these capabilities. In this paper, we present SynForecast, an open-source Python library that generates synthetic time series from 32 generators, 19 of which implement named mechanisms with a published source in econometrics, quantitative finance, statistical physics and point-process theory. Its design addresses four requirements of forecasting research: (i) output in long format, so that generated data reach a forecasting model without conversion, (ii) perturbations as overlays orthogonal to the generating process, so that a perturbation can be varied while the process is held fixed, (iii) randomness drawn before parallel dispatch, so that output does not depend on the number of workers, and (iv) mechanisms grouped into named families that can be removed one at a time. We evaluate SynForecast on M4 at five frequencies, using a 164K-parameter neural network and a gradient-boosted ensemble. We find that a classical baseline attains the lowest error in nine out of ten frequency-history combinations. Among the learned models, a synthetic prior outperforms the same network trained from scratch at a fixed budget by up to 34% when only a few dozen observations are available. However, it loses to the network trained from scratch at full history at every frequency, and none of the three rescalings of history we pre-specified predicts where the sign changes. At the hourly frequency, where the prior fails at every history, we find that the synthetic corpus lacks daily cycles, and adding them reduces zero-shot error by 43 to 63% for both models. Fine-tuning the pretrained model outperforms training from scratch by up to about 15%, but at short histories a fixed 300-step fine-tune is worse than the pretrained model used as is. Finally, removing the classical family of mechanisms costs about 2% for both models, and three further families have effects of about 1% that replicate on the 20 replicates we ran, but not on the 10 we had pre-specified.