TempusBench: An Evaluation Framework for Time-Series Forecasting

TempusBench addresses a critical gap in the rapidly expanding Time-Series Foundation Model (TSFM) space: the lack of a unified, rigorous evaluation standard. Currently, models like Google's TimesFM, Amazon's Chronos, and Lag-Llama often report performance on disparate datasets with inconsistent preprocessing. While the project is very new (1 day old, 0 stars), the 13 forks suggest significant immediate interest, likely from the academic community following its arXiv release. The defensibility is low (3) because the value of a benchmark lies entirely in its social adoption (becoming a 'de facto standard') rather than technical complexity. If researchers do not adopt it for their papers, the code itself offers no moat. It faces competition from existing libraries like GluonTS or Darts, which have established evaluation utilities, though TempusBench specifically targets 'foundation' models which often require zero-shot or few-shot evaluation protocols. Frontier risk is medium; while OpenAI and Google focus on building the models, they have a vested interest in the benchmarks used to market them. There is a high risk of market consolidation, as the community typically converges on one or two standard benchmarks (similar to GLUE for NLP or ImageNet for CV). The displacement horizon is 1-2 years, as the fast-moving nature of TS-AI means benchmarks must be updated constantly to include new 'unseen' datasets to prevent data leakage from training sets.