Table 3.

The comparison results of in-sample simulating and out-of-sample predicted performances for the three models.

Models	Simulated performance			Predicted performance
	MAE	MAPE	RMSE	MAE	MAPE	RMSE
In-sample observations from June 2008 to December 2017				6 out-of-sample predictions
SARIMA	30610.689	0.223	44383.404	40101.500	0.387	51150.446
NAR	19652.784	0.222	29740.835	21004.106	0.361	27364.566
LSTM	21548.164	0.203	27716.239	19505.800	0.221	25820.167
Reduced percentages (%)
LSTM vs. SARIMA	29.606	8.969	37.553	51.359	42.894	49.521
LSTM vs. NAR	−6.192	8.520	4.562	3.736	36.176	3.019
In-sample observations from June 2008 to June 2017				12 out-of-sample predictions
SARIMA	30268.335	0.275	43330.873	41542.355	0.366	49978.944
NAR	20678.368	0.272	31998.251	61130.494	0.373	105803.807
LSTM	20584.538	0.192	26364.913	32451.557	0.262	41678.916
Reduced percentages (%)
LSTM vs. SARIMA	31.993	30.182	39.154	21.883	28.415	16.607
LSTM vs. NAR	0.310	29.091	13.001	69.035	30.328	128.304
In-sample observations from June 2008 to December 2016				18 out-of-sample predictions
SARIMA	30136.495	0.211	43456.811	43434.889	0.479	53852.232
NAR	19957.815	0.186	29345.366	73291.494	0.629	88387.215
LSTM	19931.051	0.176	25155.016	39729.413	0.351	50951.064
Reduced percentages (%)
LSTM vs. SARIMA	33.864	16.588	42.115	8.531	26.722	5.387
LSTM vs. NAR	0.089	4.739	9.643	77.270	58.038	69.516