Research map / Time series / Foundation models for time series
From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection
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.
Fine-tunes Llama 2 (7B) with LoRA on time series paired with LLM-agent-filtered news and supplementary info, using an evaluation agent to iteratively refine news selection, and reports gains over numeric-only and classic forecasters on electricity, exchange, traffic, and bitcoin data.
Problem
Standard time series forecasters ignore external social events and news, so they struggle with sudden disruptions; incorporating unstructured news is hard because most news is irrelevant and the relevance between news and series is not known a priori.
Approach
LLM-based reasoning agents select and categorize news for forecasting, and an evaluation agent reviews prediction errors to find overlooked news and refine the selection logic over iterations; the selected news is then paired with time series in text prompts to fine-tune an LLM forecaster that predicts digit tokens.
Key results
- Filtered-news prompt reduces electricity RMSE from 337.10 (numeric only) to 280.39 and MAPE to 5.15%, and beats GPT4TS (RMSE 377.62) and PatchTST (365.41).
- On Bitcoin the method reaches MAPE 4.95% versus 5.08% for GPT4TS and 5.13% for PatchTST; on Exchange it reaches MAPE 0.65% vs 0.68% for iTransformer.
- Unfiltered news hurts performance (electricity RMSE 407.86 vs 336.41 without news), and evaluation-agent iterations improve electricity RMSE from 313.89 to 280.39 over four selections.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add agent-filtered news to the textual prompt instead of using no news (ablation) | Llama 2 time-series forecaster · Electricity | RMSE: 336.41 -> 280.39 | Helped |
| Add agent-filtered news to the textual prompt instead of using no news (ablation) | Llama 2 time-series forecaster · Exchange | RMSE: 7.41 -> 6.46 | Helped |
| Add agent-filtered news to the textual prompt instead of using no news (ablation) | Llama 2 time-series forecaster · Traffic | RMSE: 4.44 -> 4.22 | Helped |
| Add agent-filtered news to the textual prompt instead of using no news (ablation) | Llama 2 time-series forecaster · Bitcoin | RMSE: 3.87 -> 3.67 | Helped |
| Use unfiltered news instead of agent-filtered news (ablation) | Llama 2 time-series forecaster · Electricity | RMSE: 280.39 -> 407.86 | Hurt |
| Use unfiltered news instead of agent-filtered news (ablation) | Llama 2 time-series forecaster · Exchange | RMSE: 6.46 -> 8.28 | Hurt |
| Use unfiltered news instead of agent-filtered news (ablation) | Llama 2 time-series forecaster · Traffic | RMSE: 4.22 -> 4.89 | Hurt |
| Use unfiltered news instead of agent-filtered news (ablation) | Llama 2 time-series forecaster · Bitcoin | RMSE: 3.67 -> 4.02 | Hurt |
| Add all supplementary information to filtered news (ablation) | Llama 2 time-series forecaster · Electricity | RMSE: 330.79 -> 280.39 | Helped |
| Add textual descriptions instead of using only numeric tokens (ablation) | Llama 2 time-series forecaster · Electricity | RMSE: 337.10 -> 336.41 | Helped |
| Add textual descriptions instead of using only numeric tokens (ablation) | Llama 2 time-series forecaster · Exchange | RMSE: 7.80 -> 7.41 | Helped |
| Add the reflection-based news-selection update from the initial selection to the second selection (ablation) | Llama 2 time-series forecaster · Electricity | RMSE: 313.89 -> 287.35 | Helped |
Surprising findings
- Unfiltered news degrades forecasts relative to no news, showing that irrelevant news acts as noise.
- Traffic gains from news are modest, with the authors attributing this to coarse, regional news lacking local road detail.
Limitations
- Gains are modest or absent in domains with weak human/market influence such as traffic and meteorological data.
- Limited by the maximum token length of pre-trained LLMs, causing truncation for long series and multiple sequences.
- Relies on news quality and possible misinformation; the authors note need for verification mechanisms.
Experimental setup
Datasets: AEMO half-hourly electricity demand Exchange Rates API daily exchange rates Traffic (California hourly volume) Bitcoin daily price GDELT news.com.au Yahoo Finance OpenWeatherMap
Benchmarks: Electricity (AEMO) Exchange Traffic Bitcoin Illness
Metrics: RMSE MSE MAE MAPE
Baselines: Informer Autoformer FEDformer Pyraformer PatchTST iTransformer FiLM TimesNet DLinear GPT4TS
Builds on: LoRA chain-of-thought prompting Gruver et al. LLM time series forecasting GPT4TS
Closest papers by meaning
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.89)
- Approaching Human-Level Forecasting with Language Models (NeurIPS 2024 · similarity 0.89)
- TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI 2025 · similarity 0.88)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.88)
- AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · similarity 0.88)
- Are Language Models Actually Useful for Time Series Forecasting? (NeurIPS 2024 · similarity 0.87)
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.