Neural Forecasting

In this joint project with the Axel-Springer-Verlag, we develop a system based on machine learning techniques aimed at predicting the daily sales rates of newspapers. The prediction is individually done for every retail trader. The task comprises the forecast of thousands of time-series differing in many facets like the length of the given history, the average sales level, the noise ratio, seasonal changes, and individual characteristics of the respective retailer. To deal with this task, we use a neural prediction model which is both adaptive to every retail trader and general enough to minimize the individual engineering effort. The model is trained on the daily sales in past and thus adapts to the individualities of each retailer. The prediction of future sales figures simplifies to the evaluation of a mathematical function. The new approach was tested so far on two thousand retail traders of the Bild-Zeitung and results in a significant reduction of the prediction error compared to conventional forecasting algorithms. Our research is aimed at the further improvement of the prediction model, as well as the extension to the prediction of sales rates of magazines.

For financial trading, we developed multiple neural forecasting systems, predicting exchange rates (e.g. Dollar/ D-Mark, Dollar/ Jen), futures and other financial time series.