AI and ML for Coders in PyTorch

Chapter 10. Creating ML Models to Predict Sequences

Chapter 9 introduced sequence data and the attributes of a time series, including seasonality, trend, autocorrelation, and noise. You created a synthetic series to use for predictions, and you explored how to do basic statistical forecasting.

Over the next couple of chapters, you’ll learn how to use ML for forecasting. But before you start creating models, you need to understand how to structure the time series data for training predictive models by creating what we’ll call a windowed dataset.

To understand why you need to do this, consider the time series you created in Chapter 9. You can see a plot of it in Figure 10-1.

If at any point, you want to predict a value at time t, you’ll want to predict it as a function of the values preceding time t. For example, say you want to predict the value of the time series at time step 1,200 as a function of the 30 values preceding it. In this case, the values from time steps 1,170 to 1,199 would determine the value at time step 1,200 (see Figure 10-2).

Figure 10-1. Synthetic time series

Figure 10-2. Previous values impacting prediction

Now, this begins to look familiar: you can consider the values from 1,170 to 1,199 to be your features and the value at 1,200 to be your target label. If ...