Mechanics of forecasting¶

This notebook uncovers the details of forecasting by pipelines in ETNA library. We are going to explain how pipelines are dealing with dataset, transforms and models to make a prediction.

Table of contents

• Loading dataset
• Forecasting
  • Context-free models
  • Context-required models
  • ML models
• Summary

1. Loading dataset ¶

Let's load and look at the dataset

2. Forecasting ¶

Now let's dive deeper into forecasting without pipelines. We are going to use only TSDataset, transforms and models.

3.1 Context-free models ¶

Let's start by using the ProphetModel, because it doesn't require any transformations and doesn't need any context.

Fitting the model is very easy

To make a forecast we should create a dataset with future data by using make_future method. We are currently interested in only future_steps parameter, it determines how many timestamps should be created after the end of the history.

As a result we would have a dataset with future_steps timestamps.

Now we are ready to make a forecast

We should note that forecast_ts isn't a new dataset, it is the same object as future_ts, but filled with predicted values

Now let's look at a metric and plot the prediction.

3.2 Context-required models ¶

First of all, let's clarify that context is. The context is a history data the precedes the forecasting horizon.

And now let's expand our scheme to models that require some history context for forecasting. The example is NaiveModel, because it needs to know the value lag steps ago.

The fitting doesn't change

The models has context_size attribute that in this particular case is equal to lag

Future generation now needs a new parameter: tail_steps, it determines how many timestamps should be created before the end of the history.

The result will contain future_steps + tail_step timestamps.

Forecasting is slightly changed too. We need to pass prediction_size parameter that determines how many timestamps we want to see in our result.

The forecast_ts and future_ts are still the same object

The result of forecasting

3.3 ML models ¶

Now we are going to expand our scheme even further by using transformations.

Let's define the transformations

Fitting the models requires the transformations to be applied to the dataset

As you can see, there are several changes made by the transforms:

• Added date_flag_day_number_in_week column;
• Added lag_14, ..., lag_19 columns;
• Added segment_code column;
• Logarithm applied to target column.

Now we are ready to fit our model

In this case preparing future doesn't require dealing with the context, all the necessary information is in the features. But we have to deal with transformations by passing them into make_future method.

Making a forecast

The forecasted values are too small because we forecasted the target after the logarithm transformation. To get the predictions in original domain we should apply inverse transformation to the predicted values.

The result of forecasting

Summary ¶

As we can see, pipelines do a lot of work under the hood.

Training:

• Applying transformations
• Fitting the model

Forecasting:

• Generating future dataset with applied transformations
• Forecasting with the fitted model
• Inverse transformation