If I have understood your question correctly, the first approach would be better:
This is just my thoughts, which stems from intuition: For two forecasted returns of the financial assets, you will always be able to calculate a predicted outperformance value and compare it to the realized version when available. However, forecasting the outperformance metric only (assuming it is some sort of transformation/distance measure between the forecasted returns) you will not be able to back out the values of both forecasted returns (accurately). Therefore, the first method offers you the freedom to use alternative metrics and statistical evaluations on the forecasted returns, giving you a more nuanced picture of your predictions as-well as your statistical models.
In the end, you can do both. Try and observe how much the forecasted outperformance metric (found from a statistical model, ie. your second method) deviates from the realized counterpart as-well as the forecasted outperformance metric calculated from the forecasted returns (your first method). See which method is closer to the realized values. If the difference is marginal, I would opt for the first method.
Nevertheless, if your only goal is to forecast the outperformance metric, then the second method should be fine. If you want to do an extensive analysis on the financial assets, including alternative metrics and statistical evaluations, then the first method should be your choice. I hope this provides some feedback.