Table 2. Forecasting performance of different models on the training and test set.

The methods with best evaluation metric are in bold.

Models	MSE	MAE	MAPE	MSE	MAE	MAPE
	Training set (1980–2013)			Test set (2014–2018)
ARIMA (2,2,4)	83.10	5.70	15.73	1267.95	32.10	19.31
GM (1,1)	115.55	7.80	25.57	711.69	20.39	12.40
	One-step-ahead prediction			Test set (2014–2018)
Rolling GM	219.36	10.20	16.77	453.34	18.08	11.88
LSTM	120.99	8.11	12.99	2786.00	35.88	22.33
GM-LSTM	123.23	6.69	12.43	290.65	14.17	10.24