Table 9.

Differences in Fama-MacBeth (1973) regression coefficients for daily bullish sentiment (unique cashtags)

Panel A: Twitter
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Harvard-IV (1)	–	−2.27	-0.39	−1.92	−0.06	−2.15	−1.62	1.67	−0.28
LM (2)		–	1.73	0.16	2.18	−0.25	0.39	3.69	1.75
L1 (3)			–	−1.92	0.30	−2.15	−1.36	1.97	0.10
L2 (4)				–	1.92	−0.45	0.27	3.45	1.94
VADER (5)					–	−2.10	−1.51	1.74	−0.23
Naive Bayes (6)						–	1.01	3.98	2.26
Max. entropy (7)							–	3.17	1.52
Deep-MLSA (8)								–	−1.86
DeepMoji (9)									–

Panel B: StockTwits
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Harvard-IV (1)	–	−0.78	0.65	0.52	0.38	2.95	3.30	3.41	1.08
LM (2)		–	1.38	1.32	1.10	3.70	4.09	4.37	1.80
L1 (3)			–	−0.20	−0.31	2.50	3.01	2.37	0.41
L2 (4)				–	−0.15	2.89	3.54	2.51	0.64
VADER (5)					–	2.56	2.95	2.98	0.72
Naive Bayes (6)						–	0.90	0.28	−2.27
Max. entropy (7)							–	−0.15	−2.84
Deep-MLSA (8)								–	−2.01
DeepMoji (9)									–

Note: The table reports t-statistics for the difference in the average regression coefficients for daily bullish investor sentiment estimated from short messages published on Twitter (Panel A) and StockTwits (Panel B) with unique cashtags. The dependent variable of the cross-sectional regressions is the one-day ahead retail investors’ order imbalance. More precisely, we take the difference between coefficients obtained from methodologies reported in the rows with those reported in columns. The t-statistics are constructed using Newey-West (1987) standard errors. Differences which are statistically significant at the 5% level are highlighted by boldfaced numbers