Reprogramming Large Language Models for Time Series Analysis

143 concepts89 papers2 proposals

Why would a model trained on billions of words of natural language be effective at predicting the next number in an electrical grid's load sequence or a stock's price movement?

Traditionally, time series forecasting has been treated as a purely mathematical regression problem, handled by specialized numerical architectures. However, as noted by Xiyuan Zhang et al. in Large Language Models for Time Series: A Survey, Large Language Models (LLMs) are fundamentally universal sequence pattern engines. The core challenge of adapting them to temporal data is bridging the "modality gap" between continuous numerical sequences and discrete textual tokens.

By framing time series analysis as a sequence-to-sequence translation or alignment task, researchers have unlocked the ability to leverage the deep relational reasoning and zero-shot capabilities of language models for quantitative data.

The Textualization Interface: Numbers as Narratives

The most direct way to feed time series data into an LLM is to translate continuous numbers into natural language text. Instead of passing a raw vector of floats, the data is wrapped in structured semantic templates.

For instance, Hao Xue et al. in Leveraging Language Foundation Models for Human Mobility Forecasting converted numerical visitor counts into descriptive prompts containing dates, days of the week, and location types. By doing this, the model does not just see abstract values; it connects the numbers to semantic concepts like "weekend traffic" or "holiday closures."

Raw Data: [120, 135, 450] 
   ↓
Textualized Prompt: "On Monday, the visitor count was 120. On Tuesday, it was 135. On Saturday, it rose to 450."

This narrative format allows the model to apply its pre-trained pattern recognition capabilities to quantitative sequences. As demonstrated by Hao Xue et al. in Utilizing Language Models for Energy Load Forecasting, presenting data as text enables language models to capture cyclical and trend-heavy patterns effectively, facilitating zero-shot transferability across entirely different environments without further training.

To make raw data digestible for language models, researchers often abstract high-level features from the underlying numbers:

Cross-Modal Reprogramming: Embedding Alignment

While textualization provides a clean interface, it can be verbose and computationally expensive. An alternative paradigm, known as cross-modal reprogramming, bypasses explicit text conversion by mapping numerical sequences directly into the LLM's existing token embedding space.

As pioneered by Ming Jin et al. in Time-LLM: Time Series Forecasting by Reprogramming Large Language Models, this approach breaks a continuous time series into localized "patches." A lightweight, trainable reprogramming layer then maps these numerical patches directly onto text-like token embeddings, while the original weights of the LLM remain completely frozen.

<div style="text-align: center;"> <p><em>Continuous Time Series → Patching → Reprogramming Layer → Frozen LLM Backbone → Forecast</em></p> </div>

The frozen language model treats these numerical patches as if they were words in a sentence, utilizing its internal attention layers to perform relational reasoning over the temporal data. As evaluated by Zhe Li et al. in FoundTS: Comprehensive and Unified Benchmarking of Foundation Models for Time Series Forecasting, cross-modal reprogramming exploits the model's internal relational reasoning to deliver strong zero-shot generalization across diverse temporal characteristics.

Direct Numeric Strings vs. Dedicated Encoders

Surprisingly, sometimes the simplest representation is the most powerful. Tong Guan et al. in TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Language Models discovered that representing time series data simply as sequences of numeric text strings (e.g., "1.23, 1.45, 1.67") can outperform dedicated neural encoders on complex reasoning tasks. Because specialized encoders suffer from a scarcity of large-scale time series datasets, relying on the robust, pre-trained textual reasoning faculties of the LLM to parse raw digit strings provides a stronger inductive bias in data-scarce environments.

Fusing Statistical Models with Linguistic Reasoning

Rather than forcing LLMs to perform precise arithmetic, a powerful hybrid approach uses the language model to generate qualitative insights that guide traditional numerical models.

Chengsen Wang et al. in Beyond Statistical Analysis: Multimodal Framework for Time Series Forecasting with LLM-Driven Temporal Pattern designed a framework where the LLM generates "textual temporal patterns"—descriptive summaries of future trends—which are then fed into a specialized numerical forecasting model via a memory network. This separation of concerns allows the LLM to handle semantic context and qualitative shifts, while the numerical model handles the exact arithmetic.

This synergy is particularly valuable when combining time series with external, heterogeneous knowledge:

Why the Paradigm Works: Flexibility and Interpretability

Traditional time series models are notoriously rigid; a model trained on a fixed set of inputs cannot easily adapt to new variables. LLMs change this by providing a universal, dimension-agnostic interface.

Harry Amad et al. in Continuously Updating Digital Twins using Large Language Models demonstrated that by representing variables as text, new parameters or real-world changes can be introduced seamlessly at inference time via prompt updates, bypassing the constraints of fixed-input neural networks.

Furthermore, this paradigm unlocks native explainability. The authors of ReasonCast demonstrated that treating forecasting as a task-fused process of understanding and generation allows models to achieve lower error rates while explaining why a trend occurs.

To bring these capabilities to edge devices, Matthieu Boileau et al. in Towards Interpretable Time Series Foundation Models successfully distilled this time-series reasoning capability from massive multimodal models into compact language models as small as 0.5 billion parameters, proving that the structural pattern recognition of language models can be made both lightweight and highly interpretable.

Key Foundational Work

PromptCast: A New Prompt-Based Learning Paradigm for Time Series Forecasting (2022) — contributed 7 concept elements to this direction.
Leveraging Language Foundation Models for Human Mobility Forecasting (2022) — contributed 3 concept elements to this direction.
Prompt-Based Time Series Forecasting: A New Task and Dataset (2022) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_natural_language_data_encoding_445cf066,
  title = {natural language data encoding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/natural-language-data-encoding/rdr_445cf066934f404b941ebee135e36f70},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_natural_language_data_encoding_445cf066_v1,
  title = {natural language data encoding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/natural-language-data-encoding/rdr_445cf066934f404b941ebee135e36f70/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake