我觉得可以的啊，我有遇到过，模型检验都通过了，就是误差很大的，很头疼

好! There is no general answer for the question. Let's put in this way, if a current model have an out-sample error 15% and yours 7%. You make a big improvement. The less error, the better. If having your model is better than NO model, you still make a contribution.

The objective of forecast model is quite different from that of policy evaluation. For the forecast model the out-sample accuracy is the most important. For policy evaluation the unbised estomator/treatment effect is number one important.