Generally when it comes to assessing the performance of an ANN the most reliable approach is using a test set. However, as you further progress into the future there will be no more test data to use, as your model will provide the estimated values. In this regard, if you want to retrain your ANN using new incoming data you cannot test your adjusted model anymore. On what terms should you choose between two trained models without a test set? Validation error, loss etc. ? I look forward to any suggestions.