Yes, basically it just uses one model to train all different countries’ data. I understand your concern. However, it is still better than random/fixed effects /pooled OLS/ or other methodologies. I compared those models and LSTM based on their RMSEs. Even though this is one single model but it returns high accuracy. Then we still can tune LSTM parameters and try to make this model performs better for those different countries’ data. This is a possible way to analyze and predict panel data. But it does not mean that the inputs must be panel data. Looking at one problem in different perspectives is necessary for exploring new stuff.