Pith. sign in

REVIEW 9 cited by

Large Language Models Are Zero-Shot Time Series Forecasters

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2310.07820 v3 pith:43A67ZYO submitted 2023-10-11 cs.LG

Nate Gruver , Marc Finzi , Shikai Qiu , Andrew Gordon Wilson This is my paper
classification cs.LG
keywords seriestimellmsmodelsperformancedatadistributionsfind
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

By encoding time series as a string of numerical digits, we can frame time series forecasting as next-token prediction in text. Developing this approach, we find that large language models (LLMs) such as GPT-3 and LLaMA-2 can surprisingly zero-shot extrapolate time series at a level comparable to or exceeding the performance of purpose-built time series models trained on the downstream tasks. To facilitate this performance, we propose procedures for effectively tokenizing time series data and converting discrete distributions over tokens into highly flexible densities over continuous values. We argue the success of LLMs for time series stems from their ability to naturally represent multimodal distributions, in conjunction with biases for simplicity, and repetition, which align with the salient features in many time series, such as repeated seasonal trends. We also show how LLMs can naturally handle missing data without imputation through non-numerical text, accommodate textual side information, and answer questions to help explain predictions. While we find that increasing model size generally improves performance on time series, we show GPT-4 can perform worse than GPT-3 because of how it tokenizes numbers, and poor uncertainty calibration, which is likely the result of alignment interventions such as RLHF.

Discussion (0). Sign in to comment.

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. OpenAlex reports about 96 citations worldwide. Full citation record

  1. LLM-Guided Task-Semantic Field Factorization for Industrial Process Forecasting

    cs.LG 2026-07 conditional novelty 6.0 of 10

    TSF converts process variable documents into frozen semantic vectors that scale the numerical input window before a time-series backbone, yielding average MAE reductions of 2.9–3.6% across industrial forecasting tasks.

  2. Not Just What, But When: Integrating Irregular Intervals to LLM for Sequential Recommendation

    cs.IR 2025-07 conditional novelty 6.0 of 10

    IntervalLLM integrates irregular time intervals into an LLM recommender via interval embeddings and interval-infused attention, improving next-item Hit Rate@1 on three benchmarks and adding a new interval-perspective ...

  3. Causal Graph Fuzzy LLMs: A First Introduction and Applications in Time Series Forecasting

    cs.LG 2025-07 conditional novelty 5.0 of 10

    CGF-LLM combines fuzzy time series and PCMCI causal graphs into text input for fine-tuned GPT-2, reporting improved one-step-ahead forecast NRMSE and a reduction in token count on four datasets.

  4. Towards Interpretable Time Series Foundation Models

    cs.CL 2025-07 conditional novelty 5.0 of 10

    After fine-tuning on 180 synthetic mean-reverting series annotated by a large multimodal model, small Qwen models can describe trend direction, noise intensity, and extremum location in natural language.

  5. Evaluation of a Foundational Model and Stochastic Models for Forecasting Sporadic or Spiky Production Outages of High-Performance Machine Learning Services

    cs.LG 2025-06 conditional novelty 5.0 of 10

    On seven years of monthly production outage counts from a large ML service, a fine-tuned TimesFM foundation model beats moving-average and autoregressive baselines for total outages, but per root cause the best model varies.

  6. Random Initialization Can't Catch Up: The Advantage of Language Model Transfer for Time Series Forecasting

    cs.CL 2025-06 conditional novelty 5.0 of 10

    Pretrained T5 language weights give a persistent validation-loss advantage over random initialization for low-data time series forecasting, and the advantage does not vanish within the training budget.

  7. When can isotropy help adapt LLMs' next word prediction to numerical domains?

    cs.CL 2025-05 reject novelty 5.0 of 10

    Using a log-linear model and Jacobian analysis, the paper claims isotropy in LLM hidden embeddings stabilizes the softmax partition function and improves time-series forecasting, with illustrative experiments on five ...

  8. Technology-assisted Personalized Yoga for Better Health -- Challenges and Outlook

    cs.HC 2025-08 unverdicted novelty 4.0 of 10

    A vision statement for personalized yoga decision support whose claimed content cannot be verified, because the supplied full text is a different paper on maritime drift prediction.

  9. Towards Foundation Auto-Encoders for Time-Series Anomaly Detection

    cs.LG 2025-07 conditional novelty 4.0 of 10

    A univariate VAE with dilated convolutions is proposed as a simple 'foundation' model for time-series anomaly detection, with preliminary zero-shot experiments on two datasets.

Pith tools