Output details

The effects of variable stationarity in a financial time-series on Artificial Neural Networks

<24>There is a wide-spread assumption that de-trending and/or differencing non-stationary time series data prior to applying machine learning for forecasting increases accuracy. Here systematic experiments with hand-crafted (Sect.VI) and genuine stock-market data (Sect.VII) reveal a more complex situation. We show preprocessing should be applied selectively, treating stationarity as a variable property of the data to inform that choice. This novel approach is demonstrated for the first time on Artificial Neural Networks (ANNs). We also reveal how the effects of variable stationarity are influenced by other data properties and parameters of the learning task, thus providing methodological guidance to ANN users