Experimental evaluation of baselines for forecasting social media timeseries

EPJ Data Science

Table 5 Forecasting performance for total activity predicted across topics in each domain and platform. Baselines were set to forecast one week of activity at hour granularity (168 hours). To compute each metric, we aggregated the predicted activity across all topics and compare with the total number of activities in ground truth. Thus, the performance results are implicitly weighted by the number of activities in each topic. Bold entries are for the best aggregate metric

Domain	Platform	Metric	ARIMA	Hawkes	Shifted	Hawkes + ARIMA
Vz19	Twitter	APE	35.54	74.61	67.88	55.07
		RMSE	20,216.11	24,015.25	24,312.27	21,898.09
		SMAPE	81.25	93.59	98.17	76.38
		DTW	0.53	0.67	0.83	0.57
		Skewness	1.37	0.05	1.13	0.90
		Volatility	17,562.68	17,836.05	15,222.80	18,309.18
	YouTube	APE	40.70	81.07	52.39	60.88
		RMSE	128.06	173.08	148.72	148.04
		SMAPE	45.13	119.75	73.88	66.81
		DTW	0.55	0.40	0.68	0.49
		Skewness	2.71	1.65	1.45	2.52
		Volatility	92.85	104.28	77.50	106.69
CPEC	Twitter	APE	270.56	75.97	0.64	97.29
		RMSE	2082.84	767.49	1055.06	1029.92
		SMAPE	123.11	108.25	84.60	98.15
		DTW	0.92	0.45	0.74	0.91
		Skewness	3.04	0.29	0.75	3.11
		Volatility	133.45	447.03	266.56	208.28
	YouTube	APE	46.11	61.14	5.70	29.53
		RMSE	1.54	1.57	2.13	1.52
		SMAPE	103.58	113.69	116.64	114.56
		DTW	0.73	0.61	0.63	0.74
		Skewness	2.75	0.83	1.50	0.93
		Volatility	0.74	0.47	0.33	0.62
BRIA	Twitter	APE	33.30	39.34	43.84	36.32
		RMSE	430.54	733.79	731.99	533.80
		SMAPE	60.34	69.60	54.31	61.53
		DTW	0.48	0.45	0.44	0.45
		Skewness	3.09	0.13	0.08	0.40
		Volatility	294.21	142.47	213.30	120.38
	YouTube	APE	53.99	78.71	10.27	11.03
		RMSE	2.80	3.42	3.55	2.64
		SMAPE	127.34	116.49	117.78	113.58
		DTW	0.87	0.44	0.66	0.49
		Skewness	3.05	1.43	1.76	1.44
		Volatility	2.04	0.17	0.14	0.97