If a predictive model using empirical risk minimization consistently underperforms on unseen data, what might be the underlying cause?

When a model trained with empirical risk minimization works well on the training set but fails on new data, it is usually overfitting: the model has learned noise and idiosyncrasies of the training sample rather than the true underlying pattern. This happens when the model is too flexible for the amount of data or the data are not representative of the real population. For example, a polynomial regression of degree ten fitted to ten points will fit those points perfectly but will predict wildly for any other input. The mismatch shows that the empirical risk is low while the true risk on unseen data is high. To fix this, one can use simpler models, add regularization, or gather more representative data.