Hybrid BPNN-Weighted Grey-CLSP Forecasting

Conventional GM(1, 1|α) prediction always produces the huge singleton residual error around the turning point region of a time series and this phenomenon is called overshooting. A novel forecasting technique using a hybrid BPNN-weighted Grey-CLSP (BWGC) prediction that employs a back-propagation neural net (BPNN) to automatically adjust a linear combination of GM(1, 1|α) prediction and the cumulated 3-point least squared linear prediction (C3LPS) is presented herein to resolve this overshooting problem. This is because utilizing an underestimated output from C3LPS to offset an overshoot predicted output from the grey prediction will dramatically reduce the big residual error. This model exhibits a smoothing effect on the forecast to yield better an accuracy for the non-periodic short-term prediction. A three-layer BPNN with a structure of 5 × 14 × 2 multilayer-perceptron is used to tune the weights for both models. This approach was verified to be not only suitable for a stochastic type prediction (international stock price indices forecasting) but also for an inertia type prediction (forecasting the path of a typhoon).