Timer-S1: A Billion-Scale Time Series Foundation Model with Serial Scaling

ABSTRACT

While text and image foundation models have scaled massively, time-series forecasting has lagged behind due to architectural constraints. Timer-S1 introduces a billion-parameter foundation model specifically optimized for serialized numerical data. Through a novel serial scaling methodology, the model exhibits unprecedented zero-shot forecasting capabilities across diverse domains like finance, weather, and sensor telemetry.