Research map / Time series / Foundation models for time series
TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TimeXL, a multi-modal time series prediction framework combining a prototype-based explainable encoder with prediction, reflection, and refinement LLM agents in a closed loop, reporting gains in F1/AUC on four weather, finance, and healthcare datasets.
Problem
Multi-modal time series prediction methods fuse textual context for accuracy but offer little explicit reasoning about why or how contextual text affects outcomes, and they cannot easily explain predictions to users in high-stakes domains.
Approach
Couples a multi-modal prototype-based (case-based reasoning) encoder that yields predictions and prototype explanations with three GPT-4o-driven agents: a prediction LLM reasons over encoder explanations, a reflection LLM critiques predictions against ground truth, and a refinement LLM rewrites the text, which triggers encoder retraining in an iterative loop; final predictions fuse encoder and LLM
Key results
- TimeXL reaches AUC 0.808 on Weather versus 0.742 for TimeCAP, the reported up-to-8.9% AUC gain, and F1 0.696 versus 0.668.
- On Healthcare Test-Positive, TimeXL achieves F1 0.987 and AUC 0.996, versus 0.954 and 0.983 for TimeCAP (the strongest baseline, per Table 1 text).
- In the regression demo on Finance, TimeXL reaches RMSE 4.161 and MAE 2.844, improving over TimeCAP (4.456 and 3.088).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| integrate time-series and text modalities in the encoder instead of using a time-series-only encoder (ablation) | TimeXL encoder · Weather, Finance, Healthcare Test-Positive, Healthcare Mortality | F1: 0.602 -> 0.674; 0.585 -> 0.619; 0.889 -> 0.934; 0.861 -> 0.937 | Helped |
| integrate time-series and text modalities in the encoder instead of using a text-only encoder (ablation) | TimeXL encoder · Weather, Finance, Healthcare Test-Positive, Healthcare Mortality | F1: 0.567 -> 0.674; 0.472 -> 0.619; 0.871 -> 0.934; 0.840 -> 0.937 | Helped |
| add prototype explanations to text-only LLM prediction (ablation) | Prediction LLM · Weather, Finance, Healthcare Test-Positive, Healthcare Mortality | F1: 0.645 -> 0.667; 0.496 -> 0.544; 0.974 -> 0.987; 0.901 -> 0.952 | Helped |
| fuse encoder and prediction LLM outputs rather than selecting the best component (ablation) | TimeXL · Weather, Finance, Healthcare Test-Positive, Healthcare Mortality | F1: 0.674 -> 0.696; 0.619 -> 0.631; 0.987 -> 0.987; 0.952 -> 0.956 | Mixed |
| use GPT-4o rather than GPT-4o-mini (ablation) | TimeXL · Healthcare Test-Positive | F1: 0.932 -> 0.987 | Helped |
| use GPT-4o rather than Gemini-2.0-Flash (ablation) | TimeXL · Healthcare Test-Positive | F1: 0.937 -> 0.987 | Helped |
| increase time-series prototype length from 1 to 5 (ablation) | Multi-modal encoder · Finance | AUC: 0.595±0.002 -> 0.769±0.018 | Helped |
| vary prototype length across tested settings (ablation) | Multi-modal encoder · Weather | AUC: 0.748±0.011; 0.754±0.008; 0.746±0.007; 0.743±0.007 | No clear effect |
| use regression TimeXL instead of TimeCAP | TimeXL · Finance regression | RMSE: 4.456 -> 4.161 | Helped |
Surprising findings
- Text-only LLM prediction with GPT-4o is competitive with several time series baselines on healthcare, and the framework's gains depend strongly on LLM capability (GPT-4o outperforms GPT-4o-mini and Gemini-2.0-Flash).
- Iteration gains saturate after one or two refinement rounds.
Limitations
- Relies on proprietary LLM APIs (GPT-4o), with costly iteration and dependence on base LLM capability.
- Evaluation of explanation quality is only indirect (proxy via performance gains with more explanations); LLM-as-judge and human evaluation are left as suggestions.
- Gains reported are on four datasets with classification focus; the regression experiment is a single finance demonstration.
Experimental setup
Datasets: Weather (New York, Kaggle historical hourly weather with summaries) Finance (raw material prices with 14 indices and news, extended from prior work) Healthcare Test-Positive (influenza, from TimeCAP) Healthcare Mortality (influenza and pneumonia, from TimeCAP)
Benchmarks: Weather Finance Healthcare Test-Positive Healthcare Mortality
Metrics: F1 AUROC RMSE MAE MAPE
Baselines: DLinear Autoformer Crossformer TimesNet iTransformer TSMixer TimeMixer FreTS PatchTST LLMTime PromptCast OFA
Builds on: TimeCAP Time-LLM ProtoPNet Sentence-BERT
Closest papers by meaning
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.91)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.90)
- TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI 2025 · similarity 0.90)
- Rethinking Post-Training Recipes for Multimodal Time-Series Forecasting (NeurIPS 2026 · similarity 0.89)
- TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment (AAAI 2025 · similarity 0.89)
- AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.