Flowtsfm cuts parameters and improves time series forecasting accuracy
FlowTSFM: Turning Encoder Depth into Quantile Transport
Machine Learning
Summary
Forecasting future values in time series data usually uses many layers of a Transformer model, each learned separately. The authors propose FlowTSFM, which repeatedly applies the same small Transformer block and guides intermediate steps toward the final forecast using a special loss. This approach uses fewer parameters, keeps prediction quality close to bigger models, and organizes the intermediate steps better. FlowTSFM also introduces a new way to measure how well intermediate predictions move consistently toward the final answer.
What this means in practice
- •For time series engineers: Build forecasting models that use fewer parameters while maintaining accuracy by leveraging recurrent Transformer blocks with shared weights.
- •For machine learning platform teams: Integrate efficient time series foundation models that reduce resource usage and improve interpretability of intermediate forecast states.
Authors
Bahaeddine Abdessalem, Shifeng Xie, Zehao Xiao, Youssef Attia El Hili, Ambroise Odonnat, Jianfeng Zhang, Lujia Pan, Keli Zhang, Malik Tiomoko
Abstract
Encoder-based time series foundation models (TSFMs) typically rely on deep stacks of independently parameterized Transformer layers, where only the final forecast is supervised and intermediate representations have no explicit predictive role. We introduce FlowTSFM, an encoder architecture that interprets depth as a recurrent transport process: a single Transformer block is iteratively applied with shared parameters, while a quantile-flow objective supervises intermediate states along a prescribed trajectory from a prior distribution toward the final forecast. The objective combines pinball forecasting loss with path-level position matching. With only 38.8M parameters, FlowTSFM achieves competitive performance on GIFT-Eval and TIME, remaining within 1.8-4.6% MASE of stronger baselines while using approximately $3\times$ fewer parameters than a 12-layer Chronos-2 model (119.5M). Beyond accuracy, we introduce CosMean, a scale-free diagnostic measuring whether recurrent updates consistently align toward the final prediction. Under a matched intermediate-state probing protocol, FlowTSFM achieves a CosMean score of 0.919 compared with 0.350 for Chronos-2, suggesting that recurrent parameter sharing combined with path supervision is associated with substantially more structured predictive trajectories at a favorable accuracy-efficiency trade-off.