keras-metrics | DEPRECATED since Keras | Machine Learning library

I have an LSTM Encoder-Decoder model that I have developed in order to classify price movements based on the Jump-Diffusion model (binary classification problem essentially).

My model is split 75/25 between training and validation.

My issue is that after applying class imbalance techniques such as SMOTE, my model's predictive accuracy is very high across both training and validation (could be overfitting still). But, when it comes to precision, recall and f1 score my training model performs well again but on the validation side my precision and recall have declined significantly. This obviously leads to a lower f1 score on validation side.

Does anyone know why the validation accuracy would be high but the precision and recall have both declined significantly? Is this an issue with the way my model is calculating precision and recall on the validation side, or just my model is overfitting leading to lower validation results?

See image below for summary of results of model, I can also provide the notebook as well if needed.

Edit: Including Relevant Code