Table 2.

Comparison of different training/fine-tuning strategies of our method (ablation study on fine-tuning strategy). Performance is evaluated using macro F₁ score, precision, and recall. We find that first fine-tuning both branches (forecasting and score prediction) then additionally fine-tuning the score prediction branch yields best results.

Pre-trained	Fine-tune strategy	F1	Pre	Rec
Yes	Both branches then class branch	0.76	0.79	0.75
Yes	Both branches	0.72	0.75	0.71
Yes	Class branch	0.66	0.72	0.63
No		0.60	0.64	0.58