Synthetic Data Hurts Most Time Series Models, Helps Some

An extensive empirical investigation into the use of synthetic data augmentation for forecasting time series has shown that its impact is highly dependent on the model architecture. Out of 4,218 experiments, channel-mixing models, including TimesNet and iTransformer, performed better in most cases, while channel-independent models like DLinear and PatchTST experienced consistent declines in performance. In scenarios with limited resources, TimesNet, when trained on just 10% of Weather data using synthetic augmentation, outperformed the full-data baseline in 4 out of 16 combinations of sparsity datasets. Nevertheless, across all architectures, augmentation negatively affected performance in 67% of trials. The Seasonal-Trend generator was the only method that consistently improved results across the benchmarks tested. This research, available on arXiv, evaluates five architectures, four synthetic signals, and seven datasets.