Assessing forecast gains from ‘deep learning’ over time-series methodologies

There is a plethora of standard time series techniques for time series forecasting including ARIMA, ARIMAX, Spectral Analysis and Decomposition. A requirement for the application of these techniques is some degree of correlation in the series (eg the AR terms) and past effects from innovations. These properties imply that each observation is partially predictable from previous observations, from previous random spikes, or from both. An obvious assumption made is that the correlations inherent in the data set have been adequately modeled. Thus after a model has been built, any leftover variations (residuals) are considered i.i.d, independent and normally distributed with mean zero and constant variance over time. There is no further information from the residuals that can be used in the model. Implicit in these techniques is the notion that existing patterns in the time series will continue into the future. These standard techniques work well for short-term prediction, but do not prove to be effective in capturing the characteristics of data in longer period. ARIMA for instance gives more importance to immediate data points in the test set and tries to perform well for them but as we get far we see a larger variance in the predicted output. Due to the dynamic nature of the time series data often these assumptions are not met when there is non-linear autocorrelation in the series. Non-linearities in the data can be efficiently addressed with Deep Learning Techniques. Time series data are often subject to sequence dependence problem, which Deep Learning Techniques such as RNN can resolve as they are adaptive in nature. Other variants of Deep Learning such as LSTM (Long Short Term Memory) and GRU (Gated Recurrent Units) which can easily be trained based on long-term period to pick up the true dynamics of series and achieve better modeling and forecast results. Investors and real estate analysts are increasingly coming across of Deep Learning methods for market analysis and portfolio construction. We investigate potential forecast gains arising from the adoption of these models over conventional time-series models. We make use of the monthly data-series at the city level in the Europe produced by RCA. We are interested both in directional and point forecasts. The forecast evaluation takes place over different time horizons and with the application of conventional forecast assessment metrics including Diebold Mariano.