Recurrent Neural Networks for Forecasting Social Processes

We compare performance of RNN and LSTM models over shorter time series that arise often in economics or social sciences. Our goal is to determine to what extent some modern developments and increase in computing power can improve their forecasts. Our experiments indicate that in general, these neural networks do not outperform the standard statistical benchmarks. In the same time, the performance of RNNs is good enough to justify their usage as forecasting method.