Awesome Time Series Forecasting/Prediction Papers

This repository contains a reading list of papers (400+ papers !!!) on Time Series Forecasting/Prediction (TSF) and Spatio-Temporal Forecasting/Prediction (STF). These papers are mainly categorized according to the type of model. This repository is still being continuously improved. In addition to papers that have been accepted by top conferences or journals, the repository also includes the latest papers from arXiv. If you have found any relevant papers that need to be included in this repository, please feel free to submit a pull request (PR) or open an issue. If you find this repository useful, please give it a 🌟.

Each paper may apply to one or several types of forecasting, including univariate time series forecasting, multivariate time series forecasting, and spatio-temporal forecasting, which are also marked in the Type column. If covariates and exogenous variables are not considered, univariate time series forecasting involves predicting the future of one variable with the history of this variable, while multivariate time series forecasting involves predicting the future of C variables with the history of C variables. Note that repeating univariate forecasting multiple times can also achieve the goal of multivariate forecasting, which is called channel-independent. However, univariate forecasting methods cannot extract relationships between variables, so the basis for distinguishing between univariate and multivariate forecasting methods is whether the method involves interaction between variables. Besides, in the era of deep learning, many univariate models can be easily modified to directly process multiple variables for multivariate forecasting. And multivariate models generally can be directly used for univariate forecasting. Here we classify solely based on the model's description in the original paper. Spatio-temporal forecasting is often used in traffic and weather forecasting, and it adds a spatial dimension compared to univariate and multivariate forecasting. In spatio-temporal forecasting, if each measurement point has only one variable, it is equivalent to multivariate forecasting. Therefore, the distinction between spatio-temporal forecasting and multivariate forecasting is not clear. Spatio-temporal models can usually be directly applied to multivariate forecasting, and multivariate models can also be used for spatio-temporal forecasting with minor modifications. Here we also classify solely based on the model's description in the original paper.

• univariate time series forecasting: , where is the history length, is the prediction horizon length.
• multivariate time series forecasting: , where is the number of variables (channels).
• spatio-temporal forecasting: , where is the spatial dimension (number of measurement points). However, some spatio-temporal models set the output channel to 1, and even the input channel to 1, which is actually equivalent to multivariate time series forecasting.
• irregular time series: observation/sampling times are irregular.

News.

🚩 2023/11/1: I have marked some recommended papers with 🌟 (Just my personal preference 😉).

🚩 2023/11/1: I have added a new category : models specifically designed for irregular time series.

🚩 2023/11/1: I also recommend you to check out some other GitHub repositories about awesome time series papers: time-series-transformers-review, awesome-AI-for-time-series-papers, time-series-papers, deep-learning-time-series.

🚩 2023/11/3: There are some popular toolkits or code libraries that integrate many time series models: PyPOTS, Time-Series-Library, Prophet, Darts, Kats, tsai, GluonTS, PyTorchForecasting, tslearn, AutoGluon, flow-forecast, PyFlux.

🚩 2023/12/28: Since the topic of LLM(Large Language Model)+TS(Time Series) has been popular recently, I have introduced a category (LLM) to include related papers. This is distinguished from the Pretrain category. Pretrain mainly contains papers which design agent tasks (contrastive or generative) suitable for time series, and only use large-scale time series data for pre-training.

🚩 2024/4/1: Some researchers have introduced the recently popular Mamba model into the field of time series forecasting, which can be found in the SSM (State Space Model) table.

🚩 2024/6/7: I will mark some hot papers with 🔥 (Papers with over 100 citations).

🚩 2024/9/10: I am preparing to open a new GitHub repository to collect papers related to Video Spatio-Temporal Forecasting (VSTF). The mapping function for VSTF is , where and are the height and width of each frame. Compared to spatio-temporal forecasting mentioned before, it replaces with . This setup is commonly used in video prediction and weather forecasting. Stay tuned!

🚩 2024/10/23: I have introduced a new table (Multimodal) to include papers that utilize multimodal data (such as relevant text) to assist in forecasting and a new table (KAN) to include papers that utilize Kolmogorov–Arnold Networks.

🚩 2024/12/30: Christoph Bergmeir raised insightful questions about the benchmarks in the field of time series forecasting during his talk at NIPS 2024. This critique is highly valuable and well worth watching. I strongly recommend watching this talk before embarking on time series research.

🚩 2025/06/02: I have divided the papers in the Pretrain & Representation Table into two groups: Representation Learning and Foundation Models. The former focuses on designing pretrain tasks (such as contrastive learning and masked modeling), while the latter typically provides time series foundation models pre-trained on large-scale time series datasets.

<details><summary><h2 style="display: inline;">Survey & Benchmark.</h2></summary>

Date|Method|Conference|Paper Title and Paper Interpretation (In Chinese)|Code -----|----|-----|-----|----- 15-11-23|Multi-step|ACOMP 2015|Comparison of Strategies for Multi-step-Ahead Prediction of Time Series Using Neural Network|None 19-06-11|STData🔥 | TKDE 2020|Deep learning for spatio-temporal data mining: A survey|None 19-06-20|DL🔥 | SENSJ 2019|A Review of Deep Learning Models for Time Series Prediction|None 20-09-27|DL🔥 |Arxiv 2020|Time Series Forecasting With Deep Learning: A Survey|None 22-02-15|Transformer🔥 |IJCAI 2023|Transformers in Time Series: A Survey|PaperList 23-03-25|STGNN🔥 |TKDE 2023|Spatio-Temporal Graph Neural Networks for Predictive Learning in Urban Computing: A Survey|None 23-05-01|Diffusion|Arxiv 2023|Diffusion Models for Time Series Applications: A Survey|None 23-06-14|LargeST|NIPS 2023|LargeST: A Benchmark Dataset for Large-Scale Traffic Forecasting|largest 23-06-16|SSL🔥|TPAMI 2024|Self-Supervised Learning for Time Series Analysis: Taxonomy, Progress, and Prospects|None 23-06-20|OpenSTL|NIPS 2023|OpenSTL: A Comprehensive Benchmark of Spatio-Temporal Predictive Learning|[Benchmark](https