TimeSeries is an Arduino library for time series forecasting.
The library provides support for well-known statistical forecasting models, such as autoregressive and moving average models. Recursive multi-step forecasting is supported.
- Dependency-free: the library is completely dependency-free;
- Dynamic-memory-free: every model provided by the library is implemented statically using templates and circular buffers;
- Compatibility: as it makes no use of any board-specific instructions/features, this library should be compatible with almost every Arduino board;
- Support for multi-step forecasting: the library supports recursive multi-step forecasting. Although only the recursive method is implemented directly, direct methods could of course be used by instantiating multiple models. See Sahoo et al. [1] for more details on these multi-step forecasting methods.
TimeSeries currently supports the following models:
| Model | Class | Description | One-step forecasting | Multi-step forecasting |
|---|---|---|---|---|
| AR(p) | AR | An autoregressive model of order p. | Fully tested. | Fully tested. |
| MA(q) | MA | A moving average model of order q. | Fully tested. | Fully tested. |
For more information on these models, see [2, 3, 4].
When forecasting using an nth order model and a forecasting horizon of h, the time and memory complexities of the models are as follows:
| Model | Time complexity | Memory complexity |
|---|---|---|
| AR | O(nh) | O(2n) |
| MA | O(nh) | O(3n) |
Note: these values were simply derived from the code, and are thus purely indicative. Computational experiments will be conducted in the future.
TimeSeries is available for installation via the official Arduino library index. You can therefore download it using the Arduino IDE. Alternatively, the library is available for download directly from Github.
The official documentation is available at https://miguelloureiro98.github.io/TimeSeries/.
Check out the following examples for some guidance on how to use the library:
These features are currently not supported, but may be included in future versions of the library:
- Non-stationary classical time series models (e.g. ARIMA, SARIMA, ARCH);
- Classical time series models with exogenous variables (e.g. ARMAX, SARIMAX);
- Classical control-oriented models (e.g. CARIMA);
- Multivariate forecasting models (e.g. VAR, VARIMA);
- Time series deep learning models (e.g. RNN, LSTM, GRU);
- Confidence interval estimation;
- Prediction accuracy estimation (depending on the prediction horizon);
- Support for directional forecasting;
- Support for interval forecasting;
- Support for online training (for some models, at least);
- Complementary Python libraries (so the library can be used to deploy models trained in Python).
Note: some of these might actually result in separate libraries.
[1] Debashis Sahoo, Naveksha Sood, Usha Rani, George Abraham, Varun Dutt, and AD Dileep. Comparative analysis of multi-step time-series forecasting for network load dataset. In 2020 11th International Conference on Computing, Communication and Networking Technologies (ICCCNT), pages 1–7. IEEE, 2020.
[2] George E.P. Box, Gwilym M. Jenkins, Gregory C. Reinsel, and Greta M. Ljung. Time series analysis: forecasting and control. John Wiley & Sons, 2015.
[3] William W.S. Wei. Time Series Analysis: Univariate and Multivariate Methods. Addison-Wesley, 2006.
[4] Aileen Nielsen. Practical time series analysis: Prediction with statistics and machine learning. O’Reilly Media, 2019.
This library is licenced under the Apache 2.0 licence.
TimeSeries was developed and is currently maintained by Miguel Loureiro, a mechanical engineer who specialises in control systems, machine learning, and optimisation.