Mode decomposition method integrating mode reconstruction, feature extraction, and ELM for tourist arrival forecasting

Abstract A novel hybrid learning process based on the decompose-ensemble principle is proposed in this paper, integrating the NSRX learning structure with extreme learning machine (ELM) as an efficient predictor. While training the proposed model, the self-adaptive decomposition method of empirical mode decomposition (EMD) is first used to divide a training set of tourist arrival series into several relatively regular sub-series. Then, these decomposed sub-series are reconstructed into three components of high, moderate, and low frequency based on the balance of reconstructed components’ relative stationarity and the fluctuation patterns between components and the original data series. Next, extracted features and forecasting results for the three components, obtained via ELM, are combined with d -lags historical data from the undecomposed training set; this set serves as the training sample input to train the hybrid model for enhanced tourist arrival prediction. For illustration and verification purposes, the proposed learning paradigm is applied to predict Hong Kong's monthly inbound tourist arrivals from 14 source markets from January 2007 to December 2018. Empirical results demonstrate that the proposed novel ensemble-learning paradigm outperforms all benchmark models, including five popular single models and five ensemble models, in terms of prediction accuracy. These findings suggest that the proposed model shows promise in forecasting complicated time series demonstrating high volatility and irregularity.