Beyond Fit & Predict: Forecasting API for the Foundation Model Era

Franz J Király, Felipe Lourenço Angelim Vieira, Simon Blanke, Benedikt Heidrich, Jigyasu

Keywords: Software Engineering;Technical Debt;Foundation Time Series Models;Forecasting

TL;DR: Add an explicit pretrain phase so TSFMs separate global learning from task-time binding, while predict stays pure—enabling leakage-free backtesting, seamless composition, and optional fine-tuning.

Abstract: The emergence of Time Series Foundation Models (TSFMs) challenges the traditional fit-predict design pattern used in several forecasting libraries. Current frameworks force TSFM workflows -- which distinguish between global pre-training, zero-shot inference, and fine-tuning -- into interfaces originally designed for local, task-specific models. This mismatch necessitates workarounds that compromise evaluation integrity and leak data. We propose a formal API expansion to sktime, also applicable to other frameworks, that introduces a dedicated pretrain phase. This design restores command–query separation, keeps evaluation, ensembling, and deployment model-agnostic, and provides a unified interface for both classical and foundation models.

Track: Industry and Applications Track (max 2 pages)

Submission Number: 21