71 Sales, Compressed Interest Rates, and Neural Network Predictions

This work investigates the hypothesis that compressed interest rates are as efficient a predictor of aggregate sales as uncompressed interest rates. It will be shown that this efficiency can be yielded with no more than 5% of the interest rates' spectral or scaling data. The compressions were done using the discrete cosine transform (2.3% of spectral data) and the Daubechies wavelets db4 (4.3% of scaling data). The forecasting experiments were performed using dynamic focused time-lagged feedforward neural networks. The results of the relative sales forecasts were compared and contrasted and the relative performances of the models were examined to determine the different compression techniques' effects on the predictions. Based on visual examination and the performance statistics correlation coefficients, root mean square errors, mean absolute errors, and Theil inequality coefficients, the results were deemed generally inconclusive and require further investigation at compression closer to 5% of the spectral interest rates data.