MAP4TS: A Multi-Aspect Prompting Framework for Time-Series Forecasting with Large Language Models

TL;DR

MAP4TS introduces a Multi-Aspect Prompting Framework that explicitly integrates classical time-series analysis into large language model forecasting by designing four prompts (Global Domain, Local Domain, Statistical, Temporal). The architecture fuses textual prompts with raw series via an Encoding Module and a Cross-Modality Alignment Module, using GPT-2 as the forecasting backbone to produce predictions. Across eight diverse datasets, MAP4TS achieves state-of-the-art performance, with ablations showing the value of each prompt and the cross-attention fusion for modality alignment. The work demonstrates improved stability and generalization, suggesting that knowledge-grounded, prompt-conditioned LLMs can deliver robust, interpretable time-series intelligence.

Abstract

Recent advances have investigated the use of pretrained large language models (LLMs) for time-series forecasting by aligning numerical inputs with LLM embedding spaces. However, existing multimodal approaches often overlook the distinct statistical properties and temporal dependencies that are fundamental to time-series data. To bridge this gap, we propose MAP4TS, a novel Multi-Aspect Prompting Framework that explicitly incorporates classical time-series analysis into the prompt design. Our framework introduces four specialized prompt components: a Global Domain Prompt that conveys dataset-level context, a Local Domain Prompt that encodes recent trends and series-specific behaviors, and a pair of Statistical and Temporal Prompts that embed handcrafted insights derived from autocorrelation (ACF), partial autocorrelation (PACF), and Fourier analysis. Multi-Aspect Prompts are combined with raw time-series embeddings and passed through a cross-modality alignment module to produce unified representations, which are then processed by an LLM and projected for final forecasting. Extensive experiments across eight diverse datasets show that MAP4TS consistently outperforms state-of-the-art LLM-based methods. Our ablation studies further reveal that prompt-aware designs significantly enhance performance stability and that GPT-2 backbones, when paired with structured prompts, outperform larger models like LLaMA in long-term forecasting tasks.

Figures (7)

• Figure 1: Comparison of MAP4TS (red) and state-of-the-art time-series LLMs (TimeLLM jin2023time, TimesNet wu2022timesnet, CALF liu2025calf). Ours exhibits superior performance across benchmarks.
• Figure 2: The overall architecture and procedure of MAP4TS and Examples of four-aspect prompts for the ETTh1 dataset: Global Domain, Local Domain, Statistical, and Temporal. Each prompt reflects a specific perspective over the time-series data. More examples are in Figure \ref{['fig:environment-prompts']} and Figure \ref{['fig:climate-prompts']}.
• Figure 3: The Environment dataset uses four-aspect prompts. The Global prompt offers dataset context and purpose, while the Local prompt analyzes short-term dynamics by summarizing 12-week windows segmented into 7-day patches. The Statistical and Temporal prompts provide numerical data and analysis for structural and periodic interpretation.
• Figure 4: The Climate dataset uses four-aspect prompts. The Global prompt offers dataset context and purpose, while the Local prompt analyzes short-term dynamics by summarizing 12-week windows segmented into 1-week patches. The Statistical and Temporal prompts provide numerical data and analysis for structural and periodic interpretation.
• Figure 5: Attention map samples from ETTh1 and Traffic. "Global", "Local", "Statistical", "Temporal" indicates respective prompt type and "TS" represents time-series embedding.