...
This may happen because the model is not unique and a given realization model realization may not be the best across the all data pointspredictor.
Estimating the accuracy on Validation dataset which was not a part pf a training provides a better assessment of the model predictability.
...