Table 19:

Results on multimodal datasets in the finance domain. U: unimodal models, M: multimodal fusion paradigms, O: optimization objectives, T: training structures. MulT struggles on these datasets even though it performs strongly on similar multimodal time-series datasets in the affective computing domain. Other methods also show high variance across different data partitions.

	Dataset	Stocks-F&B	Stocks-Health	Stock-Tech
	Metric	MSE ↓	MSE ↓	MSE ↓
	Mean	2.140	0.575	0.140
U	ARIMA Unimodal	2.199 1.856±0.093	0.620 0.541±0.010	0.152 0.125±0.004
M	EF-LSTM LF-LSTM EF-Transformer LF-Transformer MulT [156]	1.835±0.098 1.893±0.106 2.144±0.014 2.155±0.023 2.053±0.022	0.526±0.017 0.541±0.018 0.573±0.006 0.573±0.006 0.555±0.005	0.121±0.003 0.120±0.008 0.143±0.003 0.143±0.004 0.135±0.003
T	GradBlend [167]	1.820±0.138	0.537±0.011	0.138±0.030