Contents
How does the loss function in LSTM work?
The concept here is that if the direction matches between the true price and the predicted price for the day, we keep the loss as squared difference. If it doesn’t match, then we multiply the squared difference by alpha (1000). Finally, a customized loss function is completed.
Are there any limitations to the LSTM model?
LSTM model or any other recurrent neural network model is always a black box — trading strategy can only be based on price movement without any reasons to support, and the strategies are hard to extend to portfolio allocation The limitations (1) and (3) are hard to solve without any more resources.
Is the MSE based on the LSTM model?
MSE mainly focuses on the difference between real price and predicted price without considering whether the predicted direction is correct or not. This characteristic would create huge troubles if we apply trading strategies like put / call options based on the prediction from LSTM model.
How to understand LSTM model for regression on timeseries?
I’m building a lstm model for regression on timeseries. To verify my implementation of the model and understand keras, I’m using a toyproblem to make sure I understand what’s going on. Problem is I do not understand what’s going on here.
Is the LSTM model good enough to do well?
Yes, it is desirable if we simply judge the model by looking at mean squared error (MSE). But is it good enough to do well and help us earn big money in real world trading? Sorry to say, the answer is always NO. In this article, we would give a try to customize the loss function to make our LSTM model more applicable in real world.
How to train LSTM units in Adam optimizer?
A simple architecture of LSTM units trained using Adam optimizer and Mean Squared Loss function for 25 epochs. Note that instead of using model.fit (), we use model.fit_generator () because we have created a data generator.
How to test a prediction with LSTM data?
We can test the model on testing data and see if the prediction and the actual values overlap. Rather than computing loss between predicted and actual values, we can plot it. From the graph, we can see that prediction and the actual value (ground truth) somewhat overlap. But if you zoom in, the fit is not perfect.