Foundation Model Forecasts: Form and Function
Not on the map. Why?
Abstract (arXiv)
Time-series foundation models (TSFMs) achieve strong forecast accuracy, yet accuracy alone does not determine practical value. The form of a forecast -- point, quantile, parametric, or trajectory ensemble -- fundamentally constrains which operational tasks it can support. We survey recent TSFMs and find that two-thirds produce only point or parametric forecasts, while many operational tasks require trajectory ensembles that preserve temporal dependence. We establish when forecast types can be converted and when they cannot: trajectory ensembles convert to simpler forms via marginalization without additional assumptions, but the reverse requires imposing temporal dependence through copulas or conformal methods. We prove that marginals cannot determine path-dependent event probabilities -- infinitely many joint distributions share identical marginals but yield different answers to operational questions. We map six fundamental forecasting tasks to minimal sufficient forecast types and provide a task-aligned evaluation framework. Our analysis clarifies when forecast type, not accuracy, differentiates practical utility.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 57 of its 97 references that are in the map. First posted on arXiv 22 Oct 2025.
Builds on
57 of 97 references.
References · 57
- 2022A Time Series is Worth 64 Words: Long-term Forecasting with Transformerscited by 2,039
- 2022Are Transformers Effective for Time Series Forecasting?cited by 1,424
- 2022TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysiscited by 1,058
- 2017DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networkscited by 1,013
- 2023iTransformer: Inverted Transformers Are Effective for Time Series Forecastingcited by 931
- 2021A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantificationcited by 792
- 2023Time-LLM: Time Series Forecasting by Reprogramming Large Language Modelscited by 700
- 2023A decoder-only foundation model for time-series forecastingcited by 692
- 2023One Fits All:Power General Time Series Analysis by Pretrained LMcited by 632
- 2024Unified Training of Universal Time Series Forecasting Transformerscited by 614
- 2021CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputationcited by 582
- 2023Large Language Models Are Zero-Shot Time Series Forecasterscited by 500
38 more
- 2024MOMENT: A Family of Open Time-series Foundation Modelscited by 482
- 2018Accurate Uncertainties for Deep Learning Using Calibrated Regressioncited by 422
- 2024Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Expertscited by 287
- 2024Foundation Models for Time Series Analysis: A Tutorial and Surveycited by 239
- 2017A Multi-Horizon Quantile Recurrent Forecastercited by 211
- 2025Chronos-2: From Univariate to Universal Forecastingcited by 194
- 2024Diffusion-TS: Interpretable Diffusion for General Time Series Generationcited by 184
- 2025Sundial: A Family of Highly Capable Time Series Foundation Modelscited by 180
- 2023TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecastingcited by 174
- 2023TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Seriescited by 169
- 2023UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecastingcited by 144
- 2023TSMixer: Lightweight MLP-Mixer Model for Multivariate Time Series Forecastingcited by 139
- 2024Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Seriescited by 127
- 2025TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learningcited by 120
- 2024Large Language Models for Time Series: A Surveycited by 114
- 2023Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecastingcited by 113
- 2023Large Models for Time Series and Spatio-Temporal Data: A Survey and Outlookcited by 110
- 2023ForecastPFN: Synthetically-Trained Zero-Shot Forecastingcited by 104
- 2024Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Expertscited by 95
- 2024UniTS: A Unified Multi-Task Time Series Modelcited by 94
- 2024Timer-XL: Long-Context Transformers for Unified Time Series Forecastingcited by 87
- 2024Empowering Time Series Analysis with Large Language Models: A Surveycited by 60
- 2025From Tables to Time: Extending TabPFN-v2 to Time Series Forecastingcited by 60
- 2020Conformal prediction for time seriescited by 53
- 2024Large Language Models for Forecasting and Anomaly Detection: A Systematic Literature Reviewcited by 53
- 2024Toto: Time Series Optimized Transformer for Observabilitycited by 48
- 2024Conformal prediction for multi-dimensional time series by ellipsoidal setscited by 37
- 2024In-Context Fine-Tuning for Time-Series Foundation Modelscited by 35
- 2024TimeDiT: General-purpose Diffusion Transformers for Time Series Foundation Modelcited by 33
- 2022SwinVRNN: A Data-Driven Ensemble Forecasting Model via Learned Distribution Perturbationcited by 32
- 2023TACTiS-2: Better, Faster, Simpler Attentional Copulas for Multivariate Time Seriescited by 26
- 2013Uncertainty Quantification in Complex Simulation Models Using Ensemble Copula Couplingcited by 22
- 2024Mamba4Cast: Efficient Zero-Shot Time Series Forecasting with State Space Modelscited by 17
- 2025TSPulse: Tiny Pre-Trained Models with Disentangled Representations for Rapid Time-Series Analysiscited by 12
- 2025Zero-Shot Time Series Forecasting with Covariates via In-Context Learningcited by 12
- 2024Only the Curve Shape Matters: Training Foundation Models for Zero-Shot Multivariate Time Series Forecasting through Next Curve Shape Predictioncited by 11
- 2024Fourier Head: Helping Large Language Models Learn Complex Probability Distributionscited by 10
- 2025FinCast: A Foundation Model for Financial Time-Series Forecastingcited by 10
50 most cited of 57 All 57 (JSON)
Built on by
Not cited yet.