Skforecast

0.22.0 verified Mon Apr 27 auth: no python

Skforecast is a Python library for time series forecasting using statistical and machine learning models. It works with any estimator compatible with the scikit-learn API, including LightGBM, XGBoost, CatBoost, Keras, and many others. Current version is 0.22.0, with a release cadence of roughly quarterly.

pip install skforecast

Common errors

error ImportError: cannot import name 'grid_search_forecaster' from 'skforecast.model_selection' ↓

cause In v0.22.0, the model_selection module was refactored; `grid_search_forecaster` function was removed.

fix

Use from skforecast.model_selection import BacktestGridSearch instead. See docs: https://skforecast.org/0.22.0/api/model_selection

error ValueError: The length of the values (1) does not match the length of the index (100) ↓

cause When using `ForecasterAutoregCustom`, the custom `create_lags` function returns an array with wrong shape (e.g., one column instead of multiple lags).

fix

Ensure create_lags returns a NumPy array of shape (n_samples, n_lags).

error AttributeError: 'ForecasterAutoreg' object has no attribute 'predict_interval' ↓

cause `predict_interval` was removed in v0.22.0; interval prediction is now done via `predict` with `interval=True`.

fix

Use forecaster.predict(steps=5, interval=True).

Warnings

breaking In v0.22.0, model_selection module was reorganized. Functions like `grid_search_forecaster` and `random_search_forecaster` are removed. Use `BacktestGridSearch` and `BacktestRandomSearch` instead. ↓

fix Replace `from skforecast.model_selection import grid_search_forecaster` with `from skforecast.model_selection import BacktestGridSearch` and adjust usage.

breaking In v0.21.0, `TimeSeriesDifferentiator` moved from `skforecast.utils` to `skforecast.preprocessing`. Old import will raise ImportError. ↓

fix Change import to `from skforecast.preprocessing import TimeSeriesDifferentiator`.

deprecated `ForecasterAutoregDirect` is deprecated in favor of `ForecasterAutoregMultiOutput` since v0.12.0. ↓

fix Use `ForecasterAutoregMultiOutput` with steps parameter.

gotcha When using `ForecasterAutoregCustom`, the custom `create_lags` function must return a numpy array of shape (len(y), len(lags)). Wrong shape leads to silent errors. ↓

fix Ensure function returns array with correct dimensions; test with a small dataset first.

gotcha `ForecasterAutoreg.fit()` expects `y` as a pandas Series. Passing a numpy array or DataFrame with multiple columns will raise ValueError. ↓

fix Always convert to pd.Series: `y = pd.Series(y_array)`.

deprecated `ForecasterAutoreg` parameter `selector_step` is deprecated since v0.18.0. Use `feature_selection` argument in `fit()` instead. ↓

fix Remove `selector_step` and pass `feature_selection=True` to `fit()`.

Imports

ForecasterAutoreg

wrong

from skforecast.ForecasterAutoreg import ...

correct

from skforecast.recursive import ForecasterAutoreg

Module structure changed in v0.9.0; old import path no longer works.

ForecasterAutoregDirect

wrong

from skforecast.ForecasterAutoregDirect import ...

correct

from skforecast.direct import ForecasterAutoregDirect

Direct forecaster moved to separate module.

ForecasterAutoregCustom

from skforecast.recursive import ForecasterAutoregCustom

ForecasterAutoregMultiOutput

from skforecast.recursive import ForecasterAutoregMultiOutput

Introduced in v0.12.0.

ForecasterSARIMAX

from skforecast.sarimax import ForecasterSARIMAX

BacktestGridSearch

wrong

from skforecast.model_selection_grid_search import ...

correct

from skforecast.model_selection import BacktestGridSearch

Renamed in v0.22.0; old import removed.

BacktestRandomSearch

from skforecast.model_selection import BacktestRandomSearch

TimeSeriesDifferentiator

wrong

from skforecast.utils import TimeSeriesDifferentiator

correct

from skforecast.preprocessing import TimeSeriesDifferentiator

Moved to preprocessing module in v0.21.0.

Quickstart

Basic usage: create a ForecasterAutoreg with a random forest regressor and 10 lags, fit on a univariate series, and predict 5 steps ahead.

import numpy as np
import pandas as pd
from skforecast.recursive import ForecasterAutoreg
from sklearn.ensemble import RandomForestRegressor

# Example data
y = pd.Series(np.random.rand(100), index=pd.date_range(start='2020-01-01', periods=100, freq='D'))

# Create forecaster
forecaster = ForecasterAutoreg(
    regressor=RandomForestRegressor(random_state=123),
    lags=10
)

# Fit
forecaster.fit(y=y)

# Predict
predictions = forecaster.predict(steps=5)
print(predictions.head())