We propose to show that compressed sales data obtained from less that 5% of sales discrete cosine transform coefficients are a statistically accurate predictor of inventories' future behavior. The focused time-lagged feedforward network (TLFN) and recurrent neural networks (RNN) models were used to carry out the experiment. Their results were compared to each other as well as to similar work, using uncompressed sales, previously done by the authors. The prediction accuracy was substantiated with root mean square error (RMSE) and correlation coefficient statistical tests that showed that compressed sales were a sufficient and statistically accurate predictor of aggregate inventories' future behavior. In the prediction, TLFN models performed relatively better than the RNN models and the best out-of-sample predictions were obtained using windowed forecasting. The results obtained with compressed sales as the predictor of inventories were better than those obtained in the authors' previous work wherein uncompressed sales were the predictor.
展开▼
