spotforecast2 0.0.1__py3-none-any.whl

spotforecast2/model_selection/utils_metrics.py

@@ -0,0 +1,103 @@
+"""Metrics calculation utilities for model selection."""
+
+from __future__ import annotations
+import pandas as pd
+
+
+def _calculate_metrics_one_step_ahead(
+    forecaster: object,
+    metrics: list,
+    X_train: pd.DataFrame,
+    y_train: pd.Series | dict[int, pd.Series],
+    X_test: pd.DataFrame,
+    y_test: pd.Series | dict[int, pd.Series],
+) -> list:
+    """
+    Calculate metrics when predictions are one-step-ahead.
+
+    When forecaster is of type `ForecasterDirect`, only the estimator for step 1
+    is used.
+
+    Args:
+        forecaster: Forecaster model.
+        metrics: List of metrics.
+        X_train: Predictor values used to train the model.
+        y_train: Target values related to each row of `X_train`.
+        X_test: Predictor values used to test the model.
+        y_test: Target values related to each row of `X_test`.
+
+    Returns:
+        List with metric values.
+
+    Notes:
+        When using this metric in validation, `y_train` doesn't include
+        the first window_size observations used to create the predictors and/or
+        rolling features.
+
+    Examples:
+        >>> import pandas as pd
+        >>> import numpy as np
+        >>> from sklearn.linear_model import LinearRegression
+        >>> from sklearn.metrics import mean_squared_error
+        >>> from spotforecast2.forecaster.recursive import ForecasterRecursive
+        >>> from spotforecast2.model_selection.utils_metrics import _calculate_metrics_one_step_ahead
+        >>>
+        >>> forecaster = ForecasterRecursive(LinearRegression(), lags=2)
+        >>>
+        >>> # Mock training data (already transformed into predictors/target)
+        >>> X_train = pd.DataFrame(np.array([[1, 2], [2, 3], [3, 4]]), columns=['lag_2', 'lag_1'])
+        >>> y_train = pd.Series(np.array([3, 4, 5]), name='y')
+        >>>
+        >>> # Mock test data
+        >>> X_test = pd.DataFrame(np.array([[4, 5]]), columns=['lag_2', 'lag_1'])
+        >>> y_test = pd.Series(np.array([6]), name='y')
+        >>>
+        >>> metrics = [mean_squared_error]
+        >>> result = _calculate_metrics_one_step_ahead(
+        ...     forecaster=forecaster,
+        ...     metrics=metrics,
+        ...     X_train=X_train,
+        ...     y_train=y_train,
+        ...     X_test=X_test,
+        ...     y_test=y_test
+        ... )
+        >>> print(result)
+        [0.0]
+    """
+
+    if type(forecaster).__name__ == "ForecasterDirect":
+
+        step = 1  # Only the model for step 1 is optimized.
+        X_train, y_train = forecaster.filter_train_X_y_for_step(
+            step=step, X_train=X_train, y_train=y_train
+        )
+        X_test, y_test = forecaster.filter_train_X_y_for_step(
+            step=step, X_train=X_test, y_train=y_test
+        )
+        forecaster.estimators_[step].fit(X_train, y_train)
+        y_pred = forecaster.estimators_[step].predict(X_test)
+
+    else:
+        forecaster.estimator.fit(X_train, y_train)
+        y_pred = forecaster.estimator.predict(X_test)
+
+    y_true = y_test.to_numpy()
+    y_pred = y_pred.ravel()
+    y_train = y_train.to_numpy()
+
+    if forecaster.differentiation is not None:
+        y_true = forecaster.differentiator.inverse_transform_next_window(y_true)
+        y_pred = forecaster.differentiator.inverse_transform_next_window(y_pred)
+        y_train = forecaster.differentiator.inverse_transform_training(y_train)
+
+    if forecaster.transformer_y is not None:
+        y_true = forecaster.transformer_y.inverse_transform(y_true.reshape(-1, 1))
+        y_pred = forecaster.transformer_y.inverse_transform(y_pred.reshape(-1, 1))
+        y_train = forecaster.transformer_y.inverse_transform(y_train.reshape(-1, 1))
+
+    y_true = y_true.ravel()
+    y_pred = y_pred.ravel()
+    y_train = y_train.ravel()
+    metric_values = [m(y_true=y_true, y_pred=y_pred, y_train=y_train) for m in metrics]
+
+    return metric_values