Control-based forecasting improves accuracy for changing time series

CTRL: Control-Based Time Series Forecasting with LLM-Guided Residual Learning

Machine LearningComputation and Language

Summary

Time series forecasting involves predicting future data points based on past observations, which is important in many areas like weather or finance. The authors show that large language models (LLMs), which are good at reasoning, have so far been used only as simple predictors or direct forecasters that struggle when conditions change. They introduce CTRL, which separates reasoning from prediction by letting a fixed forecast model make a base guess and then using LLMs as controllers to understand and adjust errors in the forecast’s trend, seasonality, and irregular parts. This method helps make predictions more stable and accurate when the data’s behavior shifts over time.

What this means in practice

  • For business data analysts: Improve forecasting stability and accuracy for sales or demand data that change over time by using adaptive semantic corrections.
  • For energy grid operators: Enhance load and consumption forecasts under shifting temporal patterns by applying interpretable LLM-guided adjustments to baseline predictions.

Authors

Minkyoung Kim, Daeun Ji, Yohan Lee, Beomsoo Kim, Beakcheol Jang

Abstract

Time series forecasting underpins critical decision-making across diverse domains. While large language models (LLMs) offer promising reasoning capabilities, existing LLM-based time series forecasting approaches either reduce them to numerical predictors that bypass their strengths, or allow direct forecast generation that destabilizes predictions in non-stationary settings. We introduce CTRL, a framework that decouples semantic reasoning from quantitative prediction. A frozen backbone generates base forecasts, while specialized LLM agents function as controllers that analyze backbone prediction errors through decomposed trend, seasonal, and irregular components, grounding reasoning in interpretable temporal structure. Each agent outputs compact control signals that a lightweight residual decoder translates into forecast corrections. CTRL incorporates label-free test-time adaptation that detects distribution shift from input statistics alone and readapts control signals with only 3-24 LLM calls via caching. CTRL is explicitly designed to improve robustness under non-stationary temporal dynamics and distribution shift, while remaining competitive on highly stationary time series where adaptive correction provides limited additional benefit.
View PDFOpen arXiv