Research in artificial neural networks (NN) has recently refocused on their ability to forecast seasonal time series patterns without prior deseasonalisation. While research over the past decade presented contradicting results on the benefit of deseasonalising seasonal time series, recent evaluations on artificial and empirical time series by Zhang et al. and Hill et al. suggest that NN provide only suboptimal results if the time series are not deseasonalised first. While this seems implausible considering the NN ability of universal approximation, it also contrasts experimental results within our research group. Considering the relevance of predicting seasonal and other autoregressive patterns directly from unpreprocessed time series for automatic forecasting procedures, we seek to explore this widening gap in research practices, focussing on purely seasonal time series patterns as opposed to various forms of trend-seasonal patterns used in other evaluations. We evaluate different preprocessing schemes, NN architectures and external explanatory variables to encode seasonality. Experimental predictions are computed for three artificial seasonal time series used in a recent evaluation by Zhang and Qi. Our experiments suggest that NN are capable of predicting seasonal patterns, indication the need for an extended analysis.