Deep learning for time-series
Recent years have seen a proliferation of deep neural networks, with unprecedented improvements across various application domains, in particular images, natural language processing, and sound. The potential advantage of deep learning models is that they can be much more accurate than other types of models, thereby pushing the envelope in domains such as vision, sound, and natural language processing (NLP).
In forecasting, especially demand forecasting, data is often highly erratic, discontinuous, or bursty, which violates the core assumptions of classical techniques, such as Gaussian errors, stationarity, or homoscedasticity, as discussed in Chapter 5, Forecasting of Time-Series. Deep learning techniques applied to forecasting, classification, or regression tasks could overcome many of the challenges faced by classical approaches, and, most importantly, they could provide a way to model non-linear dynamics usually neglected by traditional methods...