Time-Series Foundation Models
Pretrained models for sequential numeric data — I fine-tuned IBM Granite TSPulse on GPU and benchmarked it against tabular baselines.
Time-series foundation models are pretrained on large collections of sequential numeric data and then adapted to a specific series, the way language models are adapted to a task.
How I use it
For Real-time Data Systems, I fine-tuned IBM's Granite TSPulse (from the granite-tsfm toolkit) on GPU with PyTorch in Google Colab, running hyperparameter grids and evaluating against baselines: a tabular foundation model used in-context and gradient-boosted trees from the scikit-learn ecosystem. A practical finding I documented: the model normalizes each sample internally, which removes absolute level information — so feature design had to put that signal back explicitly.
Why it matters
Foundation models are not automatically better; they have to beat honest baselines on the same data. The same discipline applies to choosing between a hosted LLM and a smaller model in agent systems.