Multi-Level Text Alignment for Interpretable and Uncertainty-Aware Time Series Forecasting with LLMs

Adapting large language models (LLMs) for time series forecasting introduces unique challenges due to the mismatch between continuous time series tokens and the discrete word tokens LLMs are trained on. While LLMs have shown success on natural language tasks, aligning time series tokens with the semantic space of LLMs while preserving both accuracy and interpretability remains an open problem. Existing reprogramming and prompt-based approaches often generate word embeddings that are misaligned with or semantically irrelevant to the corresponding time series. For example, a segment with an increasing trend may be mapped to words like ”mirror” or ”keyboard”, which makes them hard to interpret. To bridge the gap, we propose MLTA-LLM, a Multi-Level Text Alignment framework for time series forecasting with LLMs. MLTA-LLM decomposes time series into trend, seasonal, and residual components, each of which is aligned with semantic anchor embeddings or text prototype embeddings. This component-specific alignment forces the LLM to process time series data through structured, interpretable representations. In addition, we integrate conformal prediction to provide statistically rigorous prediction intervals for uncertainty quantification. Experiments on multiple benchmark datasets show that MLTA-LLM not only improves forecasting accuracy over state-of-the-art baselines but also yields more interpretable representations and better-calibrated uncertainty estimates.