TFB

Why we included this project

Anyone who has to pick a forecasting method for real deployments quickly hits the trust problem: papers report results on datasets and hyperparameters chosen to flatter their own model, so genuine apples-to-apples comparison is rare. TFB tries to fix that by running dozens of deep learning and statistical baselines under one controlled protocol across 27 multivariate datasets, with tuned hyperparameters and reproduction scripts committed for every combination. The repo also computes time-series characteristics such as trend, seasonality, stationarity, and correlation, which helps explain why one method wins on a particular signal. A companion leaderboard (OpenTS-Bench) keeps current results up to date, so it is a handy reference point when weighing established transformer-style forecasters against newer lightweight baselines. If you evaluate models against real-world data, this is a credible shortcut to reproducible numbers.