Decreasing training loss but random fluctuating validation loss?

[Liwei Dai]

I built a simple LSTM model to predict futures prices, the inputs are the open, high, low, close, volume, open interest, and some other technical indicators. 

The target variable is if I enter into the trade tomorrow, and set up a specific trailing stop loss exit strategy, the profit and loss rate for this specific trade.

I have left over some details.

But in general, the training loss is decreasing, but the validation loss seems random. I suppose that the model didn't learn anything general to the out-of-sample

data. What should be done in this case? Like some diagnostics? I think there are many aspects that could be checked out here.

1) the input features, what should be included, what should be excluded? Does two features offset the effect of each other?

    But how do I know which features are better by diagnostics? Try them many times?

2) The model structure, maybe add one more layer? But it seems that the model is overfitting, which suggests to decrease the model complexity.

     Change some activation functions?

3) Increase dropout rate? To what extent?

4) Modify my loss function?

How to improve the model from this point?

Thanks, please give me some advice.