ML strategies for multivariate and and multi-step ahead TS forecasting of mobility data - ISF 2020

Presentation given in the framework of the 40th International Symposium of Forecasters (ISF2020) - Track Machine Learning II. Abstract: Most multivariate forecasting methods in the literature are restricted to vector time series of low dimension, linear methods and short horizons. Big data revolution is instead shifting the focus to problems (e.g. issued from the IoT technology) characterized by very large dimension, nonlinearity and long forecasting horizons. We propose DFML, a machine learning version of the Dynamic Factor Model (DFM), a successful forecasting methodology well-known in econometrics. The DFML strategy is composed of three main steps: dimensionality reduction, forecasting and dimensionality increase. The dimensionality reduction step allows to determine a limited number of latent components, on which the forecast is performed, and then transformed back to the original dimension. The selection of the nonlinear forecaster, the number of latent components and the multi-step-ahead strategy is automatically performed based on out-of-sample performance. This strategy is applied on data coming form On Board Units (OBUs) related to the mobility of trucks in Belgium, and specifically, the city of Brussels. We tested several combinations of linear (PCA) and non-linear (deep/shallow/recurrent/convolutional autoencoders) dimensionality reduction techniques and both traditional (ARIMA, ES, Holt-Winters, Theta, among others) and machine learning (k-NN, RF, single and multi-output lazy learning) forecasting techniques. In our experiments, DFML outperforms several state-of-the-art methods as well as deep learning predictors (LSTM, CNN and a combination of the two architecture) for different forecasting horizons.