Time series / Benchmark

Long-horizon forecasting (ETT, Electricity, Traffic, Weather, Exchange, ILI)

The classic long-term multivariate forecasting datasets popularised by Informer (ETT) and Autoformer (ETTh1, ETTh2, ETTm1, ETTm2, Electricity, Traffic, Weather, Exchange-rate, ILI), with prediction lengths 96/192/336/720 (shorter for ILI, e.g. Autoformer's input-36-predict-60 setting) and fixed chronological train/val/test splits.

Homepage

Test split: ETT (ETTh1, ETTh2, ETTm1, ETTm2): Informer: 'The train/val/test is 12/4/4 months.' (https://arxiv.org/pdf/2012.07436v3, Section 4 Datasets); Autoformer code (data_provider/data_loader.py, Dataset_ETT_hour): 'border2s = [12 * 30 * 24, 12 * 30 * 24 + 4 * 30 * 24, 12 * 30 * 24 + 8 * 30 * 24]' (x4 for the 15-minute ETTm files), i.e. train = first 12x30 days, val = next 4x30 days, test = next 4x30 days; data after day 600 is unused. Electricity, Traffic, Weather, Exchange-rate, ILI: Autoformer: 'We follow standard protocol and split all datasets into training, validation and test set in chronological order by the ratio of 6:2:2 for the ETT dataset and 7:1:2 for the other datasets.' (https://arxiv.org/pdf/2106.13008v5, Section 4); code (Dataset_Custom): 'num_train = int(len(df_raw) * 0.7)' and 'num_test = int(len(df_raw) * 0.2)', so test = the last 20% of time steps. Informer's original ECL split was 'The train/val/test is 15/3/4 months.'; this benchmark follows the Autoformer protocol.