himura467/tsfmx
Multimodal extension of Google's TimesFM for time series forecasting with text
What's novel
Multimodal extension of Google's TimesFM for time series forecasting with text
Code Analysis
19 files read · 4 roundsA research framework that injects text embeddings into the internal representation space of time series foundation models (Chronos-2, TimesFM 2.5) via a lightweight linear fusion layer, with a rigorous ablation methodology to measure whether the model actually reads the text.
Strengths
Genuinely thoughtful research methodology — the 9-mode ablation framework (including oracle ablations that invert the question), the text-mean centering to combat embedding anisotropy, and the text dropout to prevent fusion collapse show deep understanding of the failure modes. The TSFM adapters are real, non-trivial integrations that correctly handle each model's specific tokenization, normalization, and output projection.
Weaknesses
Zero tests despite 114 source files is a significant gap for a project claiming Production/Stable status. The fusion mechanism itself is architecturally simple (1-3 linear layers), and the sample-by-sample preprocessing loop will be a bottleneck on large datasets.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
705
Contributors
3
Files
60
Active weeks
30
Repository
Language
Python
Stars
2
Forks
0
License
MIT