From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space
arXiv:2603.12664v1 Announce Type: new Abstract: Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percen
arXiv:2603.12664v1 Announce Type: new Abstract: Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.
Executive Summary
The article proposes a novel approach, TESS, to bridge the modality gap between textual information and time-series forecasting. TESS introduces a Temporal Evolution Semantic Space to translate textual semantics into numerical cues, achieving up to 29% reduction in forecasting error. This approach addresses the limitations of existing methods, which struggle to effectively fuse textual and quantitative signals. The proposed method leverages an LLM with structured prompting and confidence-aware gating to extract interpretable temporal primitives from text.
Key Points
- ▸ TESS introduces a Temporal Evolution Semantic Space to bridge the modality gap
- ▸ The approach achieves up to 29% reduction in forecasting error compared to state-of-the-art baselines
- ▸ TESS leverages an LLM with structured prompting and confidence-aware gating to extract temporal primitives from text
Merits
Effective Modality Fusion
TESS successfully addresses the modality gap, enabling effective fusion of textual and quantitative signals for improved forecasting accuracy
Interpretable Temporal Primitives
The extracted temporal primitives are numerically grounded and interpretable, providing valuable insights into the underlying temporal dynamics
Demerits
Limited Generalizability
The approach may not generalize well to other domains or datasets, requiring further evaluation and testing to establish its robustness
Dependence on LLM Quality
The performance of TESS relies on the quality of the LLM, which may be a limiting factor if the LLM is not sufficiently advanced or well-trained
Expert Commentary
The proposed TESS approach represents a significant advancement in bridging the modality gap between textual information and time-series forecasting. By introducing a Temporal Evolution Semantic Space, TESS enables effective fusion of textual and quantitative signals, leading to improved forecasting accuracy. However, further evaluation and testing are necessary to establish the robustness and generalizability of this approach. The use of LLMs with structured prompting and confidence-aware gating is a notable strength, but also raises concerns about the dependence on LLM quality. Overall, TESS has the potential to contribute substantially to the field of time-series forecasting and multimodal fusion.
Recommendations
- ✓ Further evaluation and testing of TESS on diverse datasets and domains to establish its robustness and generalizability
- ✓ Investigation of alternative LLM architectures and training methods to mitigate the dependence on LLM quality