Time series model shares weights to cut parameters by nearly half
Fracast-0: Fractal Weight Sharing for a Time Series Foundation Model with Only 85K Parameters
Artificial Intelligence
Summary
Predicting future values in time series data often requires models with many parameters, especially when capturing patterns at different time scales. The authors present Fracast-0, a model that reuses the same part across scales by exploiting repeating patterns over time. This approach lets the model stay small—only 85,000 parameters—while still performing well across various forecasting tasks without extra tuning. Their tests show it competes with larger models but with fewer parameters, making it a promising way to create smaller, broadly useful time series models.
What this means in practice
- •For forecasting engineers: Create compact, accurate time series forecast models that handle multiple temporal scales without retraining per dataset.
- •For iot device developers: Deploy efficient forecasting models with fewer parameters on low-resource edge devices monitoring seasonal patterns.
Authors
Tianxiang Zhan, Huanyao Zhang, Yuanpeng He
Abstract
Time series foundation models must preserve multi-domain breadth, probabilistic output, and multiple temporal scales, but parameter count grows when each scale receives a separate representation. We introduce Fracast-0, a probabilistic forecasting foundation model that exploits temporal self-similarity to reuse one operator across scales. A parameter-free detector extracts significant seasonal structure. The encoder applies a shared local block along a geometric dilation ladder with scale conditioning, while the decoder combines context-gathered states with an explicit seasonal future state and reuses a second block along another ladder before emitting nine quantiles. Pretraining across six corpora preserves multi-domain breadth within 85,001 parameters. On 97 GIFT-Eval configurations without per-dataset fine-tuning, Fracast-0 is the smallest of 28 evaluated checkpoints and remains non-dominated in the aggregate parameter-accuracy plane with MASE 0.808 and WQL 0.564. It uses 42.0% fewer parameters than TinyCast, whose MASE and WQL are 4.2% and 3.3% lower. These results support cross-scale weight reuse as a practical route to further time series foundation model compression.