Neural network data preprocessing: is it necessary for time series forecasting?

Neural networks (NNs) are a commonly used method to solve the time series-forecasting problem. NNs have some advantages compared with traditional forecasting models, such as auto regressive moving average or auto regressive integrated moving average. NNs do not need to have any statistical assumption like normal distribution. However, data preprocessing, normalization, trend adjusting, seasonal adjusting, or both differencing can introduce better results in some studies. In this study, we have tried to investigate whether data preprocessing methods are useful for time series data, which contains trend, seasonality or unit root. For this purpose, we collected the real time series data belonging to monthly or quarterly observations and used nonlinear autoregressive (NAR) and multilayer perceptron (MLP) models. Although we obtained significant differences between data preprocessing methods, the structure of MLP with differenced variable produced the worst results