> ## Documentation Index
> Fetch the complete documentation index at: https://nixtla-old-docs.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

> Generate features for downstream models

# Feature engineering

***

<a href="https://github.com/Nixtla/statsforecast/blob/main/python/statsforecast/feature_engineering.py#L22" target="_blank" style={{ float: "right", fontSize: "smaller" }}>source</a>

### mstl\_decomposition

> ```text theme={null}
>  mstl_decomposition
>                      (df:Union[pandas.core.frame.DataFrame,polars.datafram
>                      e.frame.DataFrame], model:statsforecast.models.MSTL,
>                      freq:str, h:int)
> ```

*Decompose the series into trend and seasonal using the MSTL model.*

|             | **Type**  | **Details**                                                           |
| ----------- | --------- | --------------------------------------------------------------------- |
| df          | Union     | DataFrame with columns \[`unique_id`, `ds`, `y`].                     |
| model       | MSTL      | Model to use for the decomposition.                                   |
| freq        | str       | Frequency of the data (pandas alias)                                  |
| h           | int       | Forecast horizon.                                                     |
| **Returns** | **Tuple** | **Original dataframe with the ‘trend’ and ‘seasonal’ columns added.** |

```python theme={null}
import pandas as pd
from fastcore.test import test_fail
from utilsforecast.losses import smape

from statsforecast.models import Naive
from statsforecast.utils import generate_series
```

```python theme={null}
series = generate_series(10, freq='D')
series['unique_id'] = series['unique_id'].astype('int64')
```

```python theme={null}
test_fail(lambda: mstl_decomposition(series, Naive(), 'D', 14), contains='must be an MSTL instance')
```

```python theme={null}
horizon = 14
model = MSTL(season_length=7)
series = series.sample(frac=1.0)
train_df, X_df = mstl_decomposition(series, model, 'D', horizon)
```

```python theme={null}
series_pl = generate_series(10, freq='D', engine='polars')
series_pl = series_pl.with_columns(unique_id=pl.col('unique_id').cast(pl.Int64))
train_df_pl, X_df_pl = mstl_decomposition(series_pl, model, '1d', horizon)
```

```python theme={null}
pd.testing.assert_series_equal(
    train_df.groupby('unique_id')['ds'].max() + pd.offsets.Day(),
    X_df.groupby('unique_id')['ds'].min()
)
assert X_df.shape[0] == train_df['unique_id'].nunique() * horizon
pd.testing.assert_frame_equal(train_df, train_df_pl.to_pandas())
pd.testing.assert_frame_equal(X_df, X_df_pl.to_pandas())
with_estimate = train_df_pl.with_columns(estimate=pl.col('trend') + pl.col('seasonal'))
assert smape(with_estimate, models=['estimate'])['estimate'].mean() < 0.1
```

```python theme={null}
model = MSTL(season_length=[7, 28])
train_df, X_df = mstl_decomposition(series, model, 'D', horizon)
assert train_df.columns.intersection(X_df.columns).tolist() == ['unique_id', 'ds', 'trend', 'seasonal7', 'seasonal28']
```