Improvement in Prediction Performance Using Predictive Error Compensated Neural Networks

This study presents an improvement in the Predictive Error Compensated Wavelet Neural Network (PECNET) model intending to increase the accuracy of next-day predictions. The model employs multiple neural networks trained with the closing prices from various time scales in the time-domain input. Each subsequent separately trained cascade network is used for error estimation to minimize the prediction errors of the primary network without relying on iteration. Cascading predictive error compensation not only reduces the error rate but also helps alleviate the overfitting problem. The BIST30 index was used as the primary prediction. The results show significant improvements in prediction accuracy by applying different frequencies of data. Additionally, using indices from Far Eastern stock markets such as the N225, HSI, and XJO, findings show significant improvements in forecasting accuracy. Furthermore, our study analyzes different training mechanisms highlighting the effectiveness of using online training strategies. Overall, the results show an improvement in forecasting and financial analytics, intending to support investment decision-making processes.