Monash Time Series Forecasting Repository

Our Aim

---

Our aim is to introduce the first comprehensive time series forecasting repository containing datasets of related time series to facilitate the evaluation of global forecasting models. All datasets are intended to use only for research purpose. Our repository contains 30 datasets including both publicly available time series datasets (in different formats) and datasets curated by us. Many datasets have different versions based on the frequency and the inclusion of missing values, making the total number of dataset variations to 58. Furthermore, it includes both real-world and competition time series datasets covering varied domains.

We have also characterised each dataset and executed several baseline methods with them.

We recommend you to read our paper for a detailed discussion of the datasets, their original sources, feature analysis and baseline evaluation. If you use our work, please cite the paper "Rakshitha Godahewa, Christoph Bergmeir, Geoffrey I. Webb, Rob J. Hyndman, Pablo Montero-Manso, Monash Time Series Forecasting Archive".

            @InProceedings{godahewa2021monash,
              author = "Godahewa, Rakshitha and Bergmeir, Christoph and Webb, Geoffrey I. and Hyndman, Rob J. and Montero-Manso, Pablo",
              title = "Monash Time Series Forecasting Archive",
              booktitle = "Neural Information Processing Systems Track on Datasets and Benchmarks",
              year = "2021"
            }
          

Datasets

---

The following table shows a list of time series datasets that are currently available in our archive. The datasets are available in .tsf format which is a new format we propose to store time series data pioneered by sktime .ts format. The wrappers to load data into R and Python environments are available in our github repository.

Dataset	Domain	No: of Series	Min. Length	Max. Length	Competition	Multivariate	Download	Source
M1	Multiple	1001	15	150	Yes	No	Yearly Quarterly Monthly	Makridakis et al., 1982
M3	Multiple	3003	20	144	Yes	No	Yearly Quarterly Monthly Other	Makridakis and Hibon, 2000
M4	Multiple	100000	19	9933	Yes	No	Yearly Quarterly Monthly Weekly Daily Hourly	Makridakis et al., 2020
Tourism	Tourism	1311	11	333	Yes	No	Yearly Quarterly Monthly	Athanasopoulos et al., 2011
CIF 2016	Banking	72	34	120	Yes	No	Monthly	Stepnicka and Burda, 2017
London Smart Meters	Energy	5560	288	39648	No	No	W Missing W/O Missing	Jean-Michel, 2019
Aus. Electricity Demand	Energy	5	230736	232272	No	No	Half Hourly	Curated by us
Wind Farms	Energy	339	6345	527040	No	No	W Missing W/O Missing	Curated by us
Dominick	Sales	115704	28	393	No	No	Weekly	James M. Kilts Center, 2020
Bitcoin	Economic	18	2659	4581	No	No	W Missing W/O Missing	Curated by us
Pedestrian Counts	Transport	66	576	96424	No	No	Hourly	City of Melbourne, 2020
Vehicle Trips	Transport	329	70	243	No	No	W Missing W/O Missing	fivethirtyeight, 2015
KDD Cup 2018	Nature	270	9504	10920	Yes	No	W Missing W/O Missing	KDD Cup, 2018
Weather	Nature	3010	1332	65981	No	No	Daily	Sparks et al., 2020
NN5	Banking	111	791	791	Yes	Yes	Daily W Missing Daily W/O Missing Weekly	Ben Taieb et al., 2012
Web Traffic	Web	145063	803	803	Yes	Yes	Daily W Missing Daily W/O Missing Weekly	Google, 2017
Solar	Energy	137	52560	52560	No	Yes	10 Minutes Weekly	Solar, 2020
Electricity	Energy	321	26304	26304	No	Yes	Hourly Weekly	UCI, 2020
Car Parts	Sales	2674	51	51	No	Yes	W Missing W/O Missing	Hyndman, 2015
FRED-MD	Economic	107	728	728	No	Yes	Monthly	McCracken and Ng, 2016
San Francisco Traffic	Transport	862	17544	17544	No	Yes	Hourly Weekly	Caltrans, 2020
Rideshare	Transport	2304	541	541	No	Yes	W Missing W/O Missing	Curated by us
Hospital	Health	767	84	84	No	Yes	Monthly	Hyndman, 2015
COVID Deaths	Nature	266	212	212	No	Yes	Daily	Johns Hopkins University, 2020
Temperature Rain	Nature	32072	725	725	No	Yes	W Missing W/O Missing	Curated by us
Sunspot	Nature	1	73931	73931	No	No	W Missing W/O Missing	Sunspot, 2015
Saugeen River Flow	Nature	1	23741	23741	No	No	Daily	McLeod and Gweon, 2013
US Births	Nature	1	7305	7305	No	No	Daily	Pruim et al., 2020
Solar Power	Energy	1	7397222	7397222	No	No	4 Seconds	Curated by us
Wind Power	Energy	1	7397147	7397147	No	No	4 Seconds	Curated by us

Results

---

In our paper, we have evaluated the performance of 13 baseline forecasting methods across the datasets in our repository. The baseline methods include 6 traditional univariate forecasting models: Simple Exponential Smoothing (SES), Theta (Assimakopoulos and Nikolopoulos, 2000), Exponential Smoothing (ETS, Hyndman, 2008), Auto-Regressive Integrated Moving Average (ARIMA, Box and Jenkins, 1990), Trigonometric Box-Cox ARMA Trend Seasonal (TBATS, Livera et al., 2011) and Dynamic Harmonic Regression ARIMA (DHR-ARIMA, Hyndman and Athanasopoulos, 2021), and 7 global forecasting models: Pooled Regression (PR, Trapero et al., 2015), CatBoost (Prokhorenkova et al., 2018), Feed-Forward Neural Network (FFNN, Goodfellow et al., 2016), DeepAR (Salinas et al., 2020), N-BEATS (Oreshkin et al., 2019), WaveNet (Borovykh et al., 2017) and Transformer (Vaswani et al., 2017). Later, we have executed Prophet model as an additional baseline across all datasets in our repository. Furthermore, we have shown the results we obtained from the Informer model (Zhou et al, 2021) for some of the datasets. For evaluation, we use 4 error metrics namely the symmetric Mean Absolute Percentage Error (sMAPE), Mean Absolute Scaled Error (MASE, Hyndman and Koehler, 2006), Mean Absolute Error (MAE, Sammut and Webb, 2010), and Root Mean Squared Error (RMSE). The sMAPE error measure was calculated in 2 ways: the original version and the modified version suggested by Suilin (2017). For more details of these baseline methods and error metrics, please refer to our paper.

The results table shows the results of mean MASE of each baseline method across the datasets in our repository. The best model across each dataset is highlighted in boldface. We use 2 versions of ARIMA. The results of the general ARIMA method are reported for yearly, quarterly, monthly, and daily datasets whereas the results of DHR-ARIMA are reported for weekly datasets and multi-seasonal datasets such as 10 minutely, half hourly, and hourly.

The results of all error metrics across all baselines except the Prophet and Informer models are available in the online appendix. The results of all error metrics across the Prophet model are available here. The results of all error metrics across the Informer model, the reasons for selecting the datasets for the Informer model experiments and the considered Informer model configurations are available here.

In November 2024, we have updated the repository with benchmark runs of many more deep-learning methods, namely Autoformer, DLinear, NLinear, NBEATS, N-HITS, TiDE, PatchTST, and TimesFM. A report detailing these new runs is here.

We also expect to run new baselines in the future and the results tables will be updated accordingly. As new forecasting models emerge rapidly, we also provide a simple interface for you to implement other statistical, machine learning and deep learning baselines. Our github repository contains detailed instructions and example code snippets explaining how to integrate new forecasting models to our framework. The results of the newly integrated forecasting models are also evaluated in the same way as our baselines using the same evaluation metrics and thus, the results of new forecasting models and our baselines are directly comparable. After integrating the new forecasting models, you can send us a pull-request on github to officially integrate your implementations to our framework. You are also invited to send us the results of your new forecasting models. If computationally feasible, we expect to re-execute the models and confirm the results. In the future, we expect to maintain two results tables here with the confirmed and unconfirmed results of the forecasting models.

A spreadsheet with all results and also some other error measures is here.

^*The results of the Informer model are only recorded for the datasets with equal-length series. For the datasets with unequal-length series, the Informer model is required to be executed per each series where the execution time is considerably high (for details, see here). Furthermore, the intermittent datasets such as Carparts, Rideshare, Web Traffic, Covid Deaths and Temperature Rain are not considered for the Informer experiments.

About Us

---

Contribute to Our Repository

---

We encourage other researchers to contribute time series datasets or benchmarking results to our repository either by directly uploading the datasets into our repository and/or by contacting us via email. You will then be listed as a contributor, and in the acknowledgements section.

If there are any copyright issues of the datasets, please contact us via email.

Acknowledgement

---

We are very grateful to the Department of Data Science and Artificial Intelligence of Monash University for their sponsorship.

Gareth Davies from Neural Aspect has contributed benchmark runs of many of the deep-learning methods, namely Autoformer, DLinear, NLinear, NBEATS, N-HITS, TiDE, PatchTST, and TimesFM.