...
This may happen because the model is not unique and a given model realization may not be the best predictor.
Estimating the accuracy on Validation dataset which was not a part pf a training provides a better assessment of the model predictability.
In order to assess predictability of the model it should bee validated on the data outside the Training dataset, which is called Validation dataset.
If the If the model discrepancy on Validation dataset is close to the model discrepancy on Training dataset one can say that a given model has a good predictability within the Source Dataset range.
Splitting the Source Dataset into Training dataset and Validation dataset can be done in different ways.
...