Event Details

The Centre for Marketing Analytics and Forecasting (CMAF) is organising a forum at 12noon on 13th of February in George Fox LT3. This time Sasan Barak (PhD student of CMAF) will present on the topic of Deep Learning for Forecasting Model Selection

Abstract: Deep Learning has achieved breakthrough accuracy in classification tasks of image, speech and general pattern recognition. As a result, the underlying algorithms of deep neural networks (DNN) have seen a resurgence of interest across disciplines,including time series forecasting. However, these applications of DNNs see them applied as forecasting algorithms, similar to conventional neural network algorithms, using autoregressive input vectors, and thus far removed from their original domains of classifying image data. However, in forecasting model selection such applications of image recognition exist. Traditionally, expert based forecasting model specification utilises time series graphs, seasonal(year-on-year) plots, autocorrelation functions and spectral analysis charts in order to identify the existence and type of seasonality, trend, outliers and structural breaks, serving as model selection filters to narrow down the choice of potentially useful models. While visual data exploration allows accurate forecasting model selection, it does not facilitate large-scale automation of model selection over many individual time series. In this paper, we propose a novel use of deep learning in time series image recognition for model selection. We train deep neural networks on an image of a time series for a multi-class classification of its patterns of level, trend, seasonality, and combination of these components, thus select between Exponential Smoothing (ETS) base learners.