PES: Probabilistic Exponential Smoothing for Time Series Forecasting

Antonio Cifonelli, Stephane Canu

Keywords: time series forecast, demand forecast, probabilistic forecast, recurrent neural network, exponential smoothing, automatic differentiation

Abstract: Time series forecasting is a common task in many industries. It helps organizations in setting goals, making plans and taking decisions. Probabilistic forecasting, in addition, summarizes the confidence over future quantities, a useful property when targeting uncertainty. This paper proposes PES - Probabilistic Exponential Smoothing -, a hybrid model for univariate time series forecasting. The contribution is two-fold: we introduce a RNN-like cell incorporating a simple exponential smoothing operator; building on this new cell we develop an intelligible and data-efficient model. The proposed solution shows several desirable characteristics; being easy to implement and fast to train, it can accommodate multiple seasonality and learn them via cross-learning. It can produce intervals as well as point-forecasts and its structure could be a valuable time series decomposition scheme. We test the PES model over a demand forecasting task on a well-known, publicly available, dataset. Finally we show that the results obtained compare favorably to the state-of-the-art.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Probabilistic Methods (eg, variational inference, causal inference, Gaussian processes)