Generating features
Leverage StatsForecast models to create features
Some models create internal representations of the series that can be
useful for other models to use as inputs. One example is the
MSTL
model, which decomposes the series into trend and seasonal components.
This guide shows you how to use the
mstl_decomposition
function to extract those features for training and then use their
future values for inference.
import os
from functools import partial
import pandas as pd
import statsforecast
from statsforecast import StatsForecast
from statsforecast.feature_engineering import mstl_decomposition
from statsforecast.models import ARIMA, MSTL
from utilsforecast.evaluation import evaluate
from utilsforecast.losses import smape, mase
# this makes it so that the outputs of the predict methods have the id as a column
# instead of as the index
os.environ['NIXTLA_ID_AS_COL'] = '1'
df = pd.read_parquet('https://datasets-nixtla.s3.amazonaws.com/m4-hourly.parquet')
uids = df['unique_id'].unique()[:10]
df = df[df['unique_id'].isin(uids)]
df.head()
unique_id | ds | y | |
---|---|---|---|
0 | H1 | 1 | 605.0 |
1 | H1 | 2 | 586.0 |
2 | H1 | 3 | 586.0 |
3 | H1 | 4 | 559.0 |
4 | H1 | 5 | 511.0 |
Suppose that you want to use an ARIMA model to forecast your series but you want to incorporate the trend and seasonal components from the MSTL model as external regressors. You can define the MSTL model to use and then provide it to the mstl_decomposition function.
freq = 1
season_length = 24
horizon = 2 * season_length
valid = df.groupby('unique_id').tail(horizon)
train = df.drop(valid.index)
model = MSTL(season_length=24)
transformed_df, X_df = mstl_decomposition(train, model=model, freq=freq, h=horizon)
This generates the dataframe that we should use for training (with the trend and seasonal columns added), as well as the dataframe we should use to forecast.
transformed_df.head()
unique_id | ds | y | trend | seasonal | |
---|---|---|---|---|---|
0 | H1 | 1 | 605.0 | 501.350550 | 124.683643 |
1 | H1 | 2 | 586.0 | 506.424549 | 87.115039 |
2 | H1 | 3 | 586.0 | 511.453736 | 79.479564 |
3 | H1 | 4 | 559.0 | 516.434474 | 45.616992 |
4 | H1 | 5 | 511.0 | 521.362991 | -10.940498 |
X_df.head()
unique_id | ds | trend | seasonal | |
---|---|---|---|---|
0 | H1 | 701 | 626.519191 | -22.738998 |
1 | H1 | 702 | 627.423191 | -108.992314 |
2 | H1 | 703 | 628.146391 | -148.675478 |
3 | H1 | 704 | 628.724951 | -183.715284 |
4 | H1 | 705 | 629.187799 | -208.038694 |
We can now train our ARIMA models and compute our forecasts.
sf = StatsForecast(
models=[ARIMA(order=(1, 0, 1), season_length=season_length)],
freq=freq
)
preds = sf.forecast(h=horizon, df=transformed_df, X_df=X_df)
preds.head()
unique_id | ds | ARIMA | |
---|---|---|---|
0 | H1 | 701 | 597.716125 |
1 | H1 | 702 | 513.244507 |
2 | H1 | 703 | 475.042450 |
3 | H1 | 704 | 441.217743 |
4 | H1 | 705 | 417.895813 |
We can now evaluate the performance.
def compute_evaluation(preds):
full = preds.merge(valid, on=['unique_id', 'ds'])
mase24 = partial(mase, seasonality=24)
res = evaluate(full, metrics=[smape, mase24], train_df=train).groupby('metric')['ARIMA'].mean()
res_smape = '{:.1%}'.format(res['smape'])
res_mase = '{:.1f}'.format(res['mase'])
return pd.Series({'mase': res_mase, 'smape': res_smape})
compute_evaluation(preds)
mase 1.0
smape 3.7%
dtype: object
And compare this with just using the series values.
preds_noexog = sf.forecast(h=horizon, df=train)
compute_evaluation(preds_noexog)
mase 2.3
smape 7.7%
dtype: object