Transforming exogenous features
Compute transformations on your exogenous features for MLForecast
The MLForecast class allows you to compute lag transformations on your target, however, sometimes you want to also compute transformations on your dynamic exogenous features. This guide shows you how to accomplish that.
Data setup
from mlforecast.utils import generate_series, generate_prices_for_series
series = generate_series(10, equal_ends=True)
prices = generate_prices_for_series(series)
prices.head(2)
ds | unique_id | price | |
---|---|---|---|
0 | 2000-10-05 | 0 | 0.548814 |
1 | 2000-10-06 | 0 | 0.715189 |
Suppose that you have some series along with their prices for each id
and date and you want to compute forecasts for the next 7 days. Since
the price is a dynamic feature you have to provide the future values
through X_df
in
MLForecast.predict
.
If you want to use not only the price but the lag7 of the price and the
expanding mean of the lag1 for example, you can compute them before
training, merge them with your series and then provide the future values
through X_df
. Consider the following example.
Computing the transformations
from mlforecast.lag_transforms import ExpandingMean
from mlforecast.feature_engineering import transform_exog
transformed_prices = transform_exog(prices, lags=[7], lag_transforms={1: [ExpandingMean()]})
transformed_prices.head(10)
ds | unique_id | price | price_lag7 | price_expanding_mean_lag1 | |
---|---|---|---|---|---|
0 | 2000-10-05 | 0 | 0.548814 | NaN | NaN |
1 | 2000-10-06 | 0 | 0.715189 | NaN | 0.548814 |
2 | 2000-10-07 | 0 | 0.602763 | NaN | 0.632001 |
3 | 2000-10-08 | 0 | 0.544883 | NaN | 0.622255 |
4 | 2000-10-09 | 0 | 0.423655 | NaN | 0.602912 |
5 | 2000-10-10 | 0 | 0.645894 | NaN | 0.567061 |
6 | 2000-10-11 | 0 | 0.437587 | NaN | 0.580200 |
7 | 2000-10-12 | 0 | 0.891773 | 0.548814 | 0.559827 |
8 | 2000-10-13 | 0 | 0.963663 | 0.715189 | 0.601320 |
9 | 2000-10-14 | 0 | 0.383442 | 0.602763 | 0.641580 |
You can now merge this with your original series
series_with_prices = series.merge(transformed_prices, on=['unique_id', 'ds'])
series_with_prices.head(10)
unique_id | ds | y | price | price_lag7 | price_expanding_mean_lag1 | |
---|---|---|---|---|---|---|
0 | 0 | 2000-10-05 | 0.322947 | 0.548814 | NaN | NaN |
1 | 0 | 2000-10-06 | 1.218794 | 0.715189 | NaN | 0.548814 |
2 | 0 | 2000-10-07 | 2.445887 | 0.602763 | NaN | 0.632001 |
3 | 0 | 2000-10-08 | 3.481831 | 0.544883 | NaN | 0.622255 |
4 | 0 | 2000-10-09 | 4.191721 | 0.423655 | NaN | 0.602912 |
5 | 0 | 2000-10-10 | 5.395863 | 0.645894 | NaN | 0.567061 |
6 | 0 | 2000-10-11 | 6.264447 | 0.437587 | NaN | 0.580200 |
7 | 0 | 2000-10-12 | 0.284022 | 0.891773 | 0.548814 | 0.559827 |
8 | 0 | 2000-10-13 | 1.462798 | 0.963663 | 0.715189 | 0.601320 |
9 | 0 | 2000-10-14 | 2.035518 | 0.383442 | 0.602763 | 0.641580 |
You can then define your forecast object. Note that you can still compute lag features based on the target as you normally would.
from sklearn.linear_model import LinearRegression
from mlforecast import MLForecast
fcst = MLForecast(
models=[LinearRegression()],
freq='D',
lags=[1],
date_features=['dayofweek'],
)
fcst.preprocess(series_with_prices, dropna=True).head()
unique_id | ds | y | price | price_lag7 | price_expanding_mean_lag1 | lag1 | dayofweek | |
---|---|---|---|---|---|---|---|---|
1 | 0 | 2000-10-06 | 1.218794 | 0.715189 | NaN | 0.548814 | 0.322947 | 4 |
2 | 0 | 2000-10-07 | 2.445887 | 0.602763 | NaN | 0.632001 | 1.218794 | 5 |
3 | 0 | 2000-10-08 | 3.481831 | 0.544883 | NaN | 0.622255 | 2.445887 | 6 |
4 | 0 | 2000-10-09 | 4.191721 | 0.423655 | NaN | 0.602912 | 3.481831 | 0 |
5 | 0 | 2000-10-10 | 5.395863 | 0.645894 | NaN | 0.567061 | 4.191721 | 1 |
It’s important to note that the dropna
argument only considers the
null values generated by the lag features based on the target. If you
want to drop all rows containing null values you have to do that in your
original series.
series_with_prices2 = series_with_prices.dropna()
fcst.preprocess(series_with_prices2, dropna=True, static_features=[]).head()
unique_id | ds | y | price | price_lag7 | price_expanding_mean_lag1 | lag1 | dayofweek | |
---|---|---|---|---|---|---|---|---|
8 | 0 | 2000-10-13 | 1.462798 | 0.963663 | 0.715189 | 0.601320 | 0.284022 | 4 |
9 | 0 | 2000-10-14 | 2.035518 | 0.383442 | 0.602763 | 0.641580 | 1.462798 | 5 |
10 | 0 | 2000-10-15 | 3.043565 | 0.791725 | 0.544883 | 0.615766 | 2.035518 | 6 |
11 | 0 | 2000-10-16 | 4.010109 | 0.528895 | 0.423655 | 0.631763 | 3.043565 | 0 |
12 | 0 | 2000-10-17 | 5.416310 | 0.568045 | 0.645894 | 0.623190 | 4.010109 | 1 |
You can now train the model.
fcst.fit(series_with_prices2, static_features=[])
MLForecast(models=[LinearRegression], freq=D, lag_features=['lag1'], date_features=['dayofweek'], num_threads=1)
And predict using the prices. Note that you can provide the dataframe with the full history and mlforecast will filter the required dates for the forecasting horizon.
fcst.predict(1, X_df=transformed_prices).head()
unique_id | ds | LinearRegression | |
---|---|---|---|
0 | 0 | 2001-05-15 | 3.803967 |
1 | 1 | 2001-05-15 | 3.512489 |
2 | 2 | 2001-05-15 | 3.170019 |
3 | 3 | 2001-05-15 | 4.307121 |
4 | 4 | 2001-05-15 | 3.018758 |
In this example we have prices for the next 7 days, if you try to forecast a longer horizon you’ll get an error.
from fastcore.test import test_fail
test_fail(lambda: fcst.predict(8, X_df=transformed_prices), contains='Found missing inputs in X_df')