Exercises
Now we're at the end of the chapter, why not try some of the following exercises? You'll find guides on how to complete them all throughout this chapter:
A good trick is to use LSTMs on top of one-dimensional convolution, as one-dimensional convolution can go over large sequences while using fewer parameters. Try to implement an architecture that first uses a few convolutional and pooling layers and then a few LSTM layers. Try it out on the web traffic dataset. Then try adding (recurrent) dropout. Can you beat the LSTM model?
Add uncertainty to your web traffic forecasts. To do this, remember to run your model with dropout turned on at inference time. You will obtain multiple forecasts for one time step. Think about what this would mean in the context of trading and stock prices.
Visit the Kaggle datasets page and search for time series data. Make a forecasting model. This involves feature engineering with autocorrelation and Fourier transformation, picking the right model from the...
