NanoForecast v0.5: Competitive Time Series Forecasting Through Training Pipeline Optimization
A 6.5M-parameter forecaster reportedly matched or beat much larger baselines after training-pipeline fixes, without changing the model architecture.
NanoForecast v0.5 retrains the earlier v0.3 setup with corrected loss handling, tensor alignment, and broader augmentation, cutting MASE from 3.030 to 1.704 under the paper’s fixed protocol. It beats TimesFM on the three ETT datasets and exchange-rate data, while TimesFM remains ahead on electricity and traffic. The paper says training takes about 12 hours on a single NVIDIA T4 and inference runs without a GPU on an Apple M4 CPU. Code, checkpoints, and evaluation tooling are released under Apache 2.0. HF Daily Papers' note
NanoForecast v0.5 retrains the earlier v0.3 setup with corrected loss handling, tensor alignment, and broader augmentation, cutting MASE from 3.030 to 1.704 under the paper’s fixed protocol. It beats TimesFM on the three ETT datasets and exchange-rate data, while TimesFM remains ahead on electricity and traffic. The paper says training takes about 12 hours on a single NVIDIA T4 and inference runs without a GPU on an Apple M4 CPU. Code, checkpoints, and evaluation tooling are released under Apache 2.0. HF Daily Papers' note
score 4