lecrapaud 0.18.7__py3-none-any.whl → 0.22.6__py3-none-any.whl

lecrapaud/feature_preprocessing.py

@@ -0,0 +1,1202 @@
+import pandas as pd
+import numpy as np
+import joblib
+import os
+
+from sklearn.compose import ColumnTransformer
+from sklearn.decomposition import PCA
+from sklearn.impute import SimpleImputer
+from sklearn.preprocessing import StandardScaler
+from sklearn.pipeline import Pipeline
+from category_encoders import BinaryEncoder, CountEncoder
+from sklearn.preprocessing import OneHotEncoder, OrdinalEncoder
+from sklearn.model_selection import train_test_split
+
+from lecrapaud.integrations.openai_integration import (
+    truncate_text,
+    get_openai_embeddings,
+)
+from lecrapaud.feature_selection import get_features_by_types
+from lecrapaud.utils import logger
+from lecrapaud.db import Target, Feature, Experiment
+from lecrapaud.config import PYTHON_ENV
+from lecrapaud.feature_engineering import convert_object_columns_that_are_numeric
+from lecrapaud.mixins import LeCrapaudTransformerMixin
+
+
+class FeaturePreprocessor(LeCrapaudTransformerMixin):
+
+    def __init__(
+        self,
+        experiment=None,
+        **kwargs,
+    ):
+        # The mixin will automatically set all experiment.context parameters as attributes
+        super().__init__(experiment=experiment, **kwargs)
+
+        # Set defaults for attributes not automatically set by mixin
+        if not hasattr(self, "time_series"):
+            self.time_series = False
+        if not hasattr(self, "date_column"):
+            self.date_column = None
+        if not hasattr(self, "group_column"):
+            self.group_column = None
+        if not hasattr(self, "val_size"):
+            self.val_size = 0.2
+        if not hasattr(self, "test_size"):
+            self.test_size = 0.2
+        if not hasattr(self, "target_numbers"):
+            self.target_numbers = []
+        if not hasattr(self, "target_clf"):
+            self.target_clf = []
+
+        # Handle list parameters with uppercase conversion
+        if not hasattr(self, "columns_pca"):
+            self.columns_pca = []
+        else:
+            self.columns_pca = [col.upper() for col in self.columns_pca]
+        if not hasattr(self, "pca_temporal"):
+            self.pca_temporal = []
+        if not hasattr(self, "pca_cross_sectional"):
+            self.pca_cross_sectional = []
+        if not hasattr(self, "columns_onehot"):
+            self.columns_onehot = []
+        else:
+            self.columns_onehot = [col.upper() for col in self.columns_onehot]
+        if not hasattr(self, "columns_binary"):
+            self.columns_binary = []
+        else:
+            self.columns_binary = [col.upper() for col in self.columns_binary]
+        if not hasattr(self, "columns_ordinal"):
+            self.columns_ordinal = []
+        else:
+            self.columns_ordinal = [col.upper() for col in self.columns_ordinal]
+        if not hasattr(self, "columns_frequency"):
+            self.columns_frequency = []
+        else:
+            self.columns_frequency = [col.upper() for col in self.columns_frequency]
+
+        # Set experiment-related paths if experiment is available
+        if self.experiment:
+            self.experiment_dir = self.experiment.path
+            self.experiment_id = self.experiment.id
+            self.data_dir = f"{self.experiment_dir}/data"
+            self.preprocessing_dir = f"{self.experiment_dir}/preprocessing"
+
+    def fit(self, X, y=None):
+        """
+        Fit the preprocessor (learns PCA components, encoders, etc.).
+
+        Args:
+            X (pd.DataFrame): Input data
+            y: Target values (ignored)
+
+        Returns:
+            self: Returns self for chaining
+        """
+        X, y = self._validate_data(X, y)
+
+        # Store data and make columns uppercase
+        data = X.copy()
+        data.columns = data.columns.str.upper()
+
+        joblib.dump(
+            list(data.columns),
+            f"{self.preprocessing_dir}/all_features_before_encoding.pkl",
+        )
+
+        # Fit PCA components
+        data, self.pcas_ = self.add_pca_features(data)
+        data, self.pcas_cross_sectional_ = self.add_pca_feature_cross_sectional(data)
+        data, self.pcas_temporal_ = self.add_pca_feature_temporal(data)
+
+        # Fit encoding transformer
+        data, self.transformer_ = self.encode_categorical_features(data)
+
+        # Save fitted transformers if experiment is available
+        if self.experiment:
+            joblib.dump(self.pcas_, f"{self.preprocessing_dir}/pcas.pkl")
+            joblib.dump(
+                self.pcas_cross_sectional_,
+                f"{self.preprocessing_dir}/pcas_cross_sectional.pkl",
+            )
+            joblib.dump(
+                self.pcas_temporal_, f"{self.preprocessing_dir}/pcas_temporal.pkl"
+            )
+            joblib.dump(
+                self.transformer_, f"{self.preprocessing_dir}/column_transformer.pkl"
+            )
+
+            # Save features and summary
+            joblib.dump(
+                list(data.columns),
+                f"{self.preprocessing_dir}/all_features_before_selection.pkl",
+            )
+
+            if PYTHON_ENV == "Development":
+                joblib.dump(X, f"{self.data_dir}/full.pkl")
+
+            summary = summarize_dataframe(data)
+            summary.to_csv(f"{self.experiment_dir}/feature_summary.csv", index=False)
+
+        self._set_fitted()
+        return self
+
+    def transform(self, X):
+        """
+        Transform the input data using fitted components.
+
+        Args:
+            X (pd.DataFrame): Input data
+
+        Returns:
+            pd.DataFrame: Transformed data
+        """
+        # Allow loading persisted artifacts even in a fresh instance
+        if not getattr(self, "is_fitted_", False) and self.experiment:
+            if os.path.exists(f"{self.preprocessing_dir}/column_transformer.pkl"):
+                self.is_fitted_ = True
+
+        self._check_is_fitted()
+        X, _ = self._validate_data(X, reset=False)
+
+        # Transform data
+        data = X.copy()
+        data.columns = data.columns.str.upper()
+
+        # Load fitted components if not already in memory
+        if not hasattr(self, "pcas_") and self.experiment:
+            if os.path.exists(f"{self.preprocessing_dir}/pcas.pkl"):
+                self.pcas_ = joblib.load(f"{self.preprocessing_dir}/pcas.pkl")
+
+        if not hasattr(self, "pcas_cross_sectional_") and self.experiment:
+            if os.path.exists(f"{self.preprocessing_dir}/pcas_cross_sectional.pkl"):
+                self.pcas_cross_sectional_ = joblib.load(
+                    f"{self.preprocessing_dir}/pcas_cross_sectional.pkl"
+                )
+
+        if not hasattr(self, "pcas_temporal_") and self.experiment:
+            if os.path.exists(f"{self.preprocessing_dir}/pcas_temporal.pkl"):
+                self.pcas_temporal_ = joblib.load(
+                    f"{self.preprocessing_dir}/pcas_temporal.pkl"
+                )
+
+        if not hasattr(self, "transformer_") and self.experiment:
+            if os.path.exists(f"{self.preprocessing_dir}/column_transformer.pkl"):
+                self.transformer_ = joblib.load(
+                    f"{self.preprocessing_dir}/column_transformer.pkl"
+                )
+
+        # Apply PCA transformations using fitted components
+        if hasattr(self, "pcas_"):
+            data, _ = self.add_pca_features(data, pcas=self.pcas_)
+        if hasattr(self, "pcas_cross_sectional_"):
+            data, _ = self.add_pca_feature_cross_sectional(
+                data, pcas=self.pcas_cross_sectional_
+            )
+        if hasattr(self, "pcas_temporal_"):
+            data, _ = self.add_pca_feature_temporal(data, pcas=self.pcas_temporal_)
+
+        # Apply encoding using fitted transformer
+        if hasattr(self, "transformer_"):
+            data, _ = self.encode_categorical_features(
+                data, transformer=self.transformer_
+            )
+
+        return data
+
+    # embedding and pca
+    def add_pca_features(
+        self, df: pd.DataFrame, n_components: int = 5, pcas=None
+    ) -> tuple[pd.DataFrame, dict]:
+        """
+        Adds PCA components as new columns to a DataFrame from a column containing numpy arrays.
+        NEED TRAIN/TEST SPLIT BEFORE APPLYING - LIKE ENCODING CATEGORICAL VARIABLES
+
+        Parameters:
+            df (pd.DataFrame): Input DataFrame
+            column (str): Name of the column containing np.ndarray
+            n_components (int): Number of PCA components to keep
+
+        Returns:
+            pd.DataFrame: DataFrame with new PCA columns added
+        """
+        columns: list[str] = self.columns_pca
+
+        pcas_dict = {}
+        for column in columns:
+            # Convert text to embeddings if necessary
+            if not isinstance(df[column].iloc[0], (np.ndarray, list)):
+                sentences = df[column].astype(str).tolist()
+                logger.info(
+                    f"Total sentences to embed for column {column}: {len(sentences)}"
+                )
+
+                # Truncate each sentence
+                truncate_sentences = [truncate_text(sentence) for sentence in sentences]
+
+                # embedding
+                embedding_matrix = get_openai_embeddings(truncate_sentences)
+            else:
+                logger.info(f"Column {column} is already embeddings")
+                # Stack the vectors into a 2D array
+                embedding_matrix = np.vstack(df[column].values)
+
+            # Apply PCA
+            if pcas:
+                pca = pcas[column]
+                pca_features = pca.transform(embedding_matrix)
+            else:
+                pca = PCA(n_components=n_components)
+                pca_features = pca.fit_transform(embedding_matrix)
+
+            # Add PCA columns
+            for i in range(n_components):
+                df[f"{column}_pca_{i+1}"] = pca_features[:, i]
+
+            # Drop the original column
+            df.drop(column, axis=1, inplace=True)
+            pcas_dict.update({column: pca})
+
+        return df, pcas_dict
+
+    def add_pca_feature_cross_sectional_old(
+        self,
+        df: pd.DataFrame,
+        *,
+        n_components: int = 5,
+        pcas: dict[str, Pipeline] | None = None,  # si fourni: transform only
+        impute_strategy: str = "median",
+        standardize: bool = True,
+    ) -> tuple[pd.DataFrame, dict[str, Pipeline]]:
+        """
+        Construit un pivot (index=index_col, columns=columns_col, values=value_col),
+        fit (ou réutilise) un Pipeline Imputer(+Scaler)+PCA, puis merge les scores
+        (par index_col) dans df. Renvoie (df_avec_features, pipe).
+        """
+
+        pcas_dict = {}
+        index_saved = df.index
+
+        for pca_cross_sectional in self.pca_cross_sectional:
+            name, index_col, columns_col, value_col = (
+                pca_cross_sectional[k] for k in ("name", "index", "columns", "value")
+            )
+            prefix = f"CS_PC_{name}"
+
+            pivot = df.pivot_table(
+                index=index_col, columns=columns_col, values=value_col
+            ).sort_index()
+
+            # Pipeline à réutiliser entre train et test
+            if pcas is None:
+                steps = [("imputer", SimpleImputer(strategy=impute_strategy))]
+                if standardize:
+                    steps.append(
+                        ("scaler", StandardScaler(with_mean=True, with_std=True))
+                    )
+                pca = PCA(n_components=n_components, random_state=0)
+                steps.append(("pca", pca))
+                pipe = Pipeline(steps)
+                pipe.fit(pivot)  # <- fit sur TRAIN uniquement
+            else:
+                pipe = pcas[name]  # <- TEST : on réutilise le pipe existant
+
+            scores = pipe.transform(pivot)  # shape: (n_index, n_components)
+            cols = [f"{prefix}_{i}" for i in range(n_components)]
+            scores_df = pd.DataFrame(scores, index=pivot.index, columns=cols)
+
+            df = df.merge(scores_df.reset_index(), on=index_col, how="left")
+            df.index = index_saved
+            pcas_dict.update({name: pipe})
+
+        return df, pcas_dict
+
+    def add_pca_feature_cross_sectional(
+        self,
+        df: pd.DataFrame,
+        *,
+        n_components: int = 5,
+        pcas: dict[str, Pipeline] | None = None,  # si fourni: transform only
+        impute_strategy: str = "median",
+        standardize: bool = True,
+        lookback_days: int = 365,  # nombre de jours à regarder en arrière pour le fit
+        refresh_frequency: int = 90,  # refresh la PCA tous les X jours
+    ) -> tuple[pd.DataFrame, dict[str, Pipeline]]:
+        """
+        Construit un pivot (index=index_col, columns=columns_col, values=value_col),
+        fit (ou réutilise) un Pipeline Imputer(+Scaler)+PCA, puis merge les scores
+        (par index_col) dans df. Renvoie (df_avec_features, pipe).
+
+        Pour les séries temporelles : fit la PCA uniquement sur les données passées
+        pour éviter le leakage, avec refresh périodique.
+
+        Gère le cas des données panel où on a plusieurs séries temporelles
+        (ex: plusieurs stocks avec les mêmes dates).
+        """
+
+        pcas_dict = {}
+        index_saved = df.index
+
+        for pca_cross_sectional in self.pca_cross_sectional:
+            name, index_col, columns_col, value_col = (
+                pca_cross_sectional[k] for k in ("name", "index", "columns", "value")
+            )
+            prefix = f"CS_PC_{name}"
+
+            # Vérifier si c'est une série temporelle avec index = date
+            # Les dates sont déjà en ordinal après cyclic_encode_date
+            is_time_series = self.time_series and index_col == self.date_column
+
+            if is_time_series:
+                # Cas spécial : PCA cross-sectional sur des données de panel time series
+                # Par exemple : PCA sur les returns de tous les stocks à chaque date
+                # pour capturer le régime de marché
+
+                all_scores = []
+
+                # Les dates sont déjà en ordinal
+                unique_dates = sorted(df[index_col].unique())
+
+                # Pour l'inference, utiliser la PCA fournie
+                if pcas is not None:
+                    pipe = pcas[name]
+                    pivot = df.pivot_table(
+                        index=index_col, columns=columns_col, values=value_col
+                    ).sort_index()
+                    scores = pipe.transform(pivot)
+                    cols = [f"{prefix}_{i}" for i in range(n_components)]
+                    scores_df = pd.DataFrame(scores, index=pivot.index, columns=cols)
+                else:
+                    # Training : fit PCA de manière expanding avec refresh périodique
+                    pipe = None
+                    last_fit_date = None
+
+                    for i, current_date_ordinal in enumerate(unique_dates):
+                        # Convertir l'ordinal en date pour les calculs de temps
+                        current_date = pd.Timestamp.fromordinal(
+                            int(current_date_ordinal)
+                        )
+
+                        # Déterminer si on doit refitter la PCA
+                        should_refit = pipe is None or (  # Première fois
+                            last_fit_date is not None
+                            and (current_date - last_fit_date).days >= refresh_frequency
+                        )
+
+                        if (
+                            should_refit and i > 30
+                        ):  # Attendre au moins 30 jours de données
+                            # Prendre les données des 'lookback_days' derniers jours
+                            lookback_start_date = current_date - pd.Timedelta(
+                                days=lookback_days
+                            )
+                            lookback_start_ordinal = pd.Timestamp.toordinal(
+                                lookback_start_date
+                            )
+
+                            # Masque pour les dates passées uniquement (éviter le leakage)
+                            mask_fit = (df[index_col] >= lookback_start_ordinal) & (
+                                df[index_col] < current_date_ordinal
+                            )
+                            df_fit = df[mask_fit]
+
+                            if len(df_fit) > 0:
+                                # Créer le pivot pour la période de lookback
+                                pivot_fit = df_fit.pivot_table(
+                                    index=index_col,
+                                    columns=columns_col,
+                                    values=value_col,
+                                ).sort_index()
+
+                                # Vérifier qu'on a assez de dates et de colonnes
+                                if (
+                                    len(pivot_fit) >= n_components
+                                    and pivot_fit.shape[1] >= n_components
+                                ):
+                                    # Créer nouveau pipeline
+                                    steps = [
+                                        (
+                                            "imputer",
+                                            SimpleImputer(strategy=impute_strategy),
+                                        )
+                                    ]
+                                    if standardize:
+                                        steps.append(
+                                            (
+                                                "scaler",
+                                                StandardScaler(
+                                                    with_mean=True, with_std=True
+                                                ),
+                                            )
+                                        )
+                                    pca = PCA(n_components=n_components, random_state=0)
+                                    steps.append(("pca", pca))
+                                    pipe = Pipeline(steps)
+                                    pipe.fit(pivot_fit)
+                                    last_fit_date = current_date
+
+                                    logger.debug(
+                                        f"PCA {name} refitted at date {current_date.strftime('%Y-%m-%d')} "
+                                        f"using {len(pivot_fit)} dates and {pivot_fit.shape[1]} columns"
+                                    )
+
+                        # Transform pour la date courante uniquement
+                        if pipe is not None:
+                            df_current = df[df[index_col] == current_date_ordinal]
+                            if len(df_current) > 0:
+                                pivot_current = df_current.pivot_table(
+                                    index=index_col,
+                                    columns=columns_col,
+                                    values=value_col,
+                                )
+                                try:
+                                    scores_current = pipe.transform(pivot_current)
+                                    scores_dict = {
+                                        index_col: [current_date_ordinal],
+                                        **{
+                                            f"{prefix}_{j}": [scores_current[0, j]]
+                                            for j in range(n_components)
+                                        },
+                                    }
+                                    all_scores.append(pd.DataFrame(scores_dict))
+                                except Exception as e:
+                                    # En cas d'erreur (ex: nouvelles colonnes), créer des valeurs manquantes
+                                    logger.debug(
+                                        f"PCA transform error at date {current_date}: {str(e)}"
+                                    )
+                                    scores_dict = {
+                                        index_col: [current_date_ordinal],
+                                        **{
+                                            f"{prefix}_{j}": [np.nan]
+                                            for j in range(n_components)
+                                        },
+                                    }
+                                    all_scores.append(pd.DataFrame(scores_dict))
+                        else:
+                            # Pas encore de PCA fittée, créer des NaN
+                            scores_dict = {
+                                index_col: [current_date_ordinal],
+                                **{
+                                    f"{prefix}_{j}": [np.nan]
+                                    for j in range(n_components)
+                                },
+                            }
+                            all_scores.append(pd.DataFrame(scores_dict))
+
+                    # Combiner tous les scores
+                    if all_scores:
+                        scores_df = pd.concat(all_scores, ignore_index=True)
+                    else:
+                        # Créer un DataFrame vide avec les bonnes colonnes
+                        cols = [f"{prefix}_{i}" for i in range(n_components)]
+                        scores_df = pd.DataFrame(columns=[index_col] + cols)
+
+                # Merger les scores
+                df = df.merge(scores_df, on=index_col, how="left")
+                df.index = index_saved
+
+                # Forward fill puis 0 pour éviter les NaN
+                pca_cols = [col for col in df.columns if col.startswith(prefix)]
+                df[pca_cols] = df[pca_cols].fillna(method="ffill").fillna(0)
+
+                pcas_dict.update({name: pipe})
+
+            else:
+                # Approche classique (non time series ou index != date)
+                pivot = df.pivot_table(
+                    index=index_col, columns=columns_col, values=value_col
+                ).sort_index()
+
+                # Pipeline à réutiliser entre train et test
+                if pcas is None:
+                    steps = [("imputer", SimpleImputer(strategy=impute_strategy))]
+                    if standardize:
+                        steps.append(
+                            ("scaler", StandardScaler(with_mean=True, with_std=True))
+                        )
+                    pca = PCA(n_components=n_components, random_state=0)
+                    steps.append(("pca", pca))
+                    pipe = Pipeline(steps)
+                    pipe.fit(pivot)  # <- fit sur TRAIN uniquement
+                else:
+                    pipe = pcas[name]  # <- TEST : on réutilise le pipe existant
+
+                scores = pipe.transform(pivot)  # shape: (n_index, n_components)
+                cols = [f"{prefix}_{i}" for i in range(n_components)]
+                scores_df = pd.DataFrame(scores, index=pivot.index, columns=cols)
+
+                df = df.merge(scores_df.reset_index(), on=index_col, how="left")
+                df.index = index_saved
+                pcas_dict.update({name: pipe})
+
+        return df, pcas_dict
+
+    # ----------------- 2) PCA TEMPORELLE (liste de colonnes lags) ----------------
+    def add_pca_feature_temporal_old(
+        self,
+        df: pd.DataFrame,
+        *,
+        n_components: int = 5,
+        pcas: dict[str, Pipeline] | None = None,  # si fourni: transform only
+        impute_strategy: (
+            str | None
+        ) = None,  # None = on exige toutes les colonnes présentes
+        standardize: bool = True,
+    ) -> tuple[pd.DataFrame, dict[str, Pipeline]]:
+        """
+        Applique une PCA sur une matrice (rows = lignes df, cols = lags).
+        Fit le Pipeline sur TRAIN si pcas=None; sinon, utilise pcas et fait transform.
+        Ajoute les colonnes f"{prefix}_{i}" dans df. Renvoie (df, pipe).
+        """
+        pcas_dict = {}
+
+        for pca_temporal in self.pca_temporal:
+            name, cols = (pca_temporal[k] for k in ("name", "columns"))
+            prefix = f"TMP_PC_{name}"
+
+            # Masque des lignes utilisables
+            if impute_strategy is None:
+                mask = (
+                    df[cols].notna().all(axis=1)
+                )  # on n'impute pas → lignes complètes
+                X_fit = df.loc[mask, cols]
+            else:
+                mask = df[cols].notna().any(axis=1)  # on imputera → au moins une valeur
+                X_fit = df.loc[mask, cols]
+
+            # Pipeline
+            if pcas is None:
+                steps = []
+                if impute_strategy is not None:
+                    steps.append(("imputer", SimpleImputer(strategy=impute_strategy)))
+                if standardize:
+                    steps.append(
+                        ("scaler", StandardScaler(with_mean=True, with_std=True))
+                    )
+                pca = PCA(n_components=n_components, random_state=0)
+                steps.append(("pca", pca))
+                pipe = Pipeline(steps)
+                if not X_fit.empty:
+                    pipe.fit(X_fit)  # <- fit sur TRAIN uniquement
+            else:
+                pipe = pcas[name]  # <- TEST
+
+            # Transform uniquement sur lignes valides (mask)
+            if not df.loc[mask, cols].empty:
+                Z = pipe.transform(df.loc[mask, cols])
+                for i in range(n_components):
+                    df.loc[mask, f"{prefix}_{i}"] = Z[:, i]
+            else:
+                # crée les colonnes vides si aucune ligne valide (cohérence de schéma)
+                for i in range(n_components):
+                    df[f"{prefix}_{i}"] = pd.NA
+
+            pcas_dict.update({name: pipe})
+
+        return df, pcas_dict
+
+    def add_pca_feature_temporal(
+        self,
+        df: pd.DataFrame,
+        *,
+        n_components: int = 5,
+        pcas: dict[str, Pipeline] | None = None,
+        impute_strategy: str = "median",
+        standardize: bool = True,
+        lookback_days: int = 365,
+        refresh_frequency: int = 90,
+    ) -> tuple[pd.DataFrame, dict[str, Pipeline]]:
+        """
+        PCA temporelle pour time series avec support panel data.
+        Crée automatiquement les colonnes de lags et évite le look-ahead bias.
+
+        Format pca_temporal simplifié:
+        [{"name": "LAST_20_RET", "column": "RET", "lags": 20}]
+        """
+        pcas_dict = {}
+
+        for pca_config in self.pca_temporal:
+            # Support both old and new format
+            if "columns" in pca_config:
+                # Old format: use existing columns
+                name = pca_config["name"]
+                lag_columns = pca_config["columns"]
+                base_column = None
+                num_lags = len(lag_columns)
+            else:
+                # New format: create lag columns
+                name = pca_config["name"]
+                base_column = pca_config["column"].upper()
+                num_lags = pca_config.get("lags", 20)
+
+                # Create lag columns if they don't exist
+                if self.group_column:
+                    # Panel data: create lags by group
+                    for lag in range(1, num_lags + 1):
+                        lag_col = f"{base_column}_-{lag}"
+                        if lag_col not in df.columns:
+                            df[lag_col] = df.groupby(self.group_column)[
+                                base_column
+                            ].shift(lag)
+                else:
+                    # Simple time series
+                    for lag in range(1, num_lags + 1):
+                        lag_col = f"{base_column}_-{lag}"
+                        if lag_col not in df.columns:
+                            df[lag_col] = df[base_column].shift(lag)
+
+                lag_columns = [f"{base_column}_-{i}" for i in range(1, num_lags + 1)]
+
+            prefix = f"TMP_PC_{name}"
+
+            # For time series: avoid look-ahead bias
+            if self.time_series and self.date_column:
+                all_scores = []
+                unique_dates = sorted(df[self.date_column].unique())
+
+                if pcas is not None:
+                    # transform: use provided PCA
+                    pipe = pcas[name]
+
+                    # Apply to all data at once
+                    mask = df[lag_columns].notna().all(axis=1)
+                    if mask.any():
+                        X_transform = df.loc[mask, lag_columns]
+                        scores = pipe.transform(X_transform)
+
+                        for i in range(n_components):
+                            df.loc[mask, f"{prefix}_{i}"] = scores[:, i]
+
+                    # Fill NaN with forward fill then 0
+                    pca_cols = [f"{prefix}_{i}" for i in range(n_components)]
+                    df[pca_cols] = df[pca_cols].fillna(method="ffill").fillna(0)
+
+                else:
+                    # Training: expanding window with periodic refresh
+                    pipe = None
+                    last_fit_date = None
+
+                    for current_date_ordinal in unique_dates:
+                        current_date = pd.Timestamp.fromordinal(
+                            int(current_date_ordinal)
+                        )
+
+                        # Determine if we should refit
+                        should_refit = pipe is None or (
+                            last_fit_date is not None
+                            and (current_date - last_fit_date).days >= refresh_frequency
+                        )
+
+                        if (
+                            should_refit
+                            and len(df[df[self.date_column] < current_date_ordinal])
+                            > num_lags * 2
+                        ):
+                            # Get historical data for fitting
+                            lookback_start = current_date - pd.Timedelta(
+                                days=lookback_days
+                            )
+                            lookback_start_ordinal = pd.Timestamp.toordinal(
+                                lookback_start
+                            )
+
+                            mask_fit = (
+                                (df[self.date_column] >= lookback_start_ordinal)
+                                & (df[self.date_column] < current_date_ordinal)
+                                & df[lag_columns].notna().all(axis=1)
+                            )
+
+                            if mask_fit.sum() >= n_components:
+                                X_fit = df.loc[mask_fit, lag_columns]
+
+                                # Create pipeline
+                                steps = []
+                                if impute_strategy is not None:
+                                    steps.append(
+                                        (
+                                            "imputer",
+                                            SimpleImputer(strategy=impute_strategy),
+                                        )
+                                    )
+                                if standardize:
+                                    steps.append(("scaler", StandardScaler()))
+                                steps.append(
+                                    (
+                                        "pca",
+                                        PCA(n_components=n_components, random_state=0),
+                                    )
+                                )
+
+                                pipe = Pipeline(steps)
+                                pipe.fit(X_fit)
+                                last_fit_date = current_date
+
+                                logger.debug(
+                                    f"Temporal PCA {name} refitted at {current_date.strftime('%Y-%m-%d')} "
+                                    f"using {len(X_fit)} samples"
+                                )
+
+                        # Transform current date data
+                        if pipe is not None:
+                            mask_current = (
+                                df[self.date_column] == current_date_ordinal
+                            ) & df[lag_columns].notna().all(axis=1)
+
+                            if mask_current.any():
+                                X_current = df.loc[mask_current, lag_columns]
+                                scores = pipe.transform(X_current)
+
+                                for i in range(n_components):
+                                    df.loc[mask_current, f"{prefix}_{i}"] = scores[:, i]
+
+                    # Fill NaN with forward fill then 0
+                    pca_cols = [f"{prefix}_{i}" for i in range(n_components)]
+                    for col in pca_cols:
+                        if col not in df.columns:
+                            df[col] = 0
+                    df[pca_cols] = df[pca_cols].fillna(method="ffill").fillna(0)
+
+                pcas_dict[name] = pipe
+
+            else:
+                # Non time-series: use original approach
+                mask = df[lag_columns].notna().all(axis=1)
+
+                if pcas is None and mask.any():
+                    X_fit = df.loc[mask, lag_columns]
+
+                    steps = []
+                    if impute_strategy is not None:
+                        steps.append(
+                            ("imputer", SimpleImputer(strategy=impute_strategy))
+                        )
+                    if standardize:
+                        steps.append(("scaler", StandardScaler()))
+                    steps.append(
+                        ("pca", PCA(n_components=n_components, random_state=0))
+                    )
+
+                    pipe = Pipeline(steps)
+                    pipe.fit(X_fit)
+                    pcas_dict[name] = pipe
+                elif pcas is not None:
+                    pipe = pcas[name]
+                    pcas_dict[name] = pipe
+                else:
+                    continue
+
+                if mask.any():
+                    X_transform = df.loc[mask, lag_columns]
+                    scores = pipe.transform(X_transform)
+
+                    for i in range(n_components):
+                        df.loc[mask, f"{prefix}_{i}"] = scores[:, i]
+
+                # Fill missing values
+                pca_cols = [f"{prefix}_{i}" for i in range(n_components)]
+                for col in pca_cols:
+                    if col not in df.columns:
+                        df[col] = 0
+                df[pca_cols] = df[pca_cols].fillna(0)
+
+        return df, pcas_dict
+
+    # encoding categorical features
+    def encode_categorical_features(
+        self,
+        df: pd.DataFrame,
+        transformer: ColumnTransformer | None = None,
+    ) -> tuple[pd.DataFrame, ColumnTransformer]:
+        """
+        Encodes categorical columns using one-hot, binary, ordinal, and frequency encoding.
+
+        Parameters:
+            df (pd.DataFrame): Input DataFrame
+            columns_onehot (list[str]) Creates one binary column per category forLow-cardinality categorical features
+            columns_binary (list[str]) Converts categories into binary and splits bits across columns for Mid-to-high cardinality (e.g., 10–100 unique values)
+            columns_ordinal (list[str]) Assigns integer ranks to categories When order matters (e.g., low < medium < high)
+            columns_frequency (list[str]) Replaces each category with its frequency count, normalized to proportion. High-cardinality features with meaning in frequency
+            transformer (ColumnTransformer, optional): if provided, applies transform only
+
+        Returns:
+            tuple: (transformed DataFrame, ColumnTransformer)
+        """
+        columns_onehot: list[str] = self.columns_onehot
+        columns_binary: list[str] = self.columns_binary
+        columns_ordinal: list[str] = self.columns_ordinal
+        columns_frequency: list[str] = self.columns_frequency
+
+        X = df.loc[:, ~df.columns.str.contains("^TARGET_")]
+        y = df.loc[:, df.columns.str.contains("^TARGET_")]
+        save_in_db = False
+
+        all_columns = (
+            columns_onehot + columns_binary + columns_ordinal + columns_frequency
+        )
+
+        if transformer:
+            transformed = transformer.transform(X)
+        else:
+            transformer = ColumnTransformer(
+                transformers=[
+                    (
+                        "onehot",
+                        OneHotEncoder(handle_unknown="ignore", sparse_output=False),
+                        columns_onehot,
+                    ),
+                    (
+                        "ordinal",
+                        OrdinalEncoder(
+                            handle_unknown="use_encoded_value", unknown_value=-1
+                        ),
+                        columns_ordinal,
+                    ),
+                    ("binary", BinaryEncoder(handle_unknown="value"), columns_binary),
+                    ("freq", CountEncoder(normalize=True), columns_frequency),
+                ],
+                remainder="passthrough",
+            )
+            transformed = transformer.fit_transform(X)
+            save_in_db = True
+
+        # Build output column names
+        column_names = []
+
+        if columns_onehot:
+            column_names.extend(
+                transformer.named_transformers_["onehot"]
+                .get_feature_names_out(columns_onehot)
+                .tolist()
+            )
+
+        if columns_ordinal:
+            column_names.extend(columns_ordinal)
+
+        if columns_binary:
+            column_names.extend(
+                transformer.named_transformers_["binary"]
+                .get_feature_names_out(columns_binary)
+                .tolist()
+            )
+
+        if columns_frequency:
+            column_names.extend(columns_frequency)
+
+        # Add passthrough (non-encoded) columns
+        passthrough_columns = [col for col in X.columns if col not in all_columns]
+        column_names.extend(passthrough_columns)
+
+        X_transformed = pd.DataFrame(transformed, columns=column_names, index=df.index)
+
+        # Try to convert columns to best possible dtypes
+        X_transformed = X_transformed.convert_dtypes()
+
+        # Insert features in db
+        if save_in_db:
+            # Get feature types from transformed data
+            categorical_features, numerical_features = get_features_by_types(
+                X_transformed
+            )
+
+            # Get column names from DataFrames
+            cat_feature_names = categorical_features.columns.tolist()
+            num_feature_names = numerical_features.columns.tolist()
+
+            # Combine all feature names and their types
+            all_feature_names = cat_feature_names + num_feature_names
+            all_feature_types = ["categorical"] * len(cat_feature_names) + [
+                "numerical"
+            ] * len(num_feature_names)
+
+            # Upsert features in bulk if we have any features
+            if all_feature_names:
+                Feature.bulk_upsert(
+                    name=all_feature_names,
+                    type=all_feature_types,
+                )
+
+            # Upsert targets in bulk
+            target_names = y.columns.tolist()
+            target_types = [
+                (
+                    "classification"
+                    if int(target.split("_")[1]) in self.target_clf
+                    else "regression"
+                )
+                for target in target_names
+            ]
+
+            Target.bulk_upsert(name=target_names, type=target_types)
+
+            # Get all the upserted objects
+            targets = Target.filter(name__in=target_names)
+
+            # Update experiment with targets
+            experiment = Experiment.get(self.experiment_id)
+            if experiment:
+                experiment.targets = targets
+                experiment.save()
+
+        return pd.concat([X_transformed, y], axis=1), transformer
+
+
+# utils
+def summarize_dataframe(
+    df: pd.DataFrame, sample_categorical_threshold: int = 15
+) -> pd.DataFrame:
+    summary = []
+
+    def is_hashable_series(series: pd.Series) -> bool:
+        try:
+            _ = series.dropna().unique()
+            return True
+        except TypeError:
+            return False
+
+    df = convert_object_columns_that_are_numeric(df)
+    df = df.convert_dtypes()
+
+    for col in df.columns:
+        total_missing = df[col].isna().sum()
+        col_data = df[col].dropna()
+        dtype = col_data.dtype
+
+        if col_data.empty:
+            summary.append(
+                {
+                    "Column": col,
+                    "Dtype": dtype,
+                    "Type": "unknown",
+                    "Detail": "No non-null values",
+                    "Missing": total_missing,
+                }
+            )
+            continue
+
+        # Case 1: Numeric columns
+        if pd.api.types.is_numeric_dtype(col_data):
+            unique_vals = col_data.nunique()
+
+            if set(col_data.unique()).issubset({0, 1}):
+                col_type = "binary-categorical"
+                detail = "0/1 values only"
+            elif (
+                pd.api.types.is_integer_dtype(col_data)
+                and unique_vals <= sample_categorical_threshold
+            ):
+                col_type = "multi-categorical"
+                top_vals = col_data.value_counts().head(10)
+                detail = ", ".join(f"{k} ({v})" for k, v in top_vals.items())
+            else:
+                col_type = "numeric"
+                q = col_data.quantile([0, 0.25, 0.5, 0.75, 1])
+                detail = (
+                    f"Min: {q.iloc[0]:.2f}, Q1: {q.iloc[1]:.2f}, Median: {q.iloc[2]:.2f}, "
+                    f"Q3: {q.iloc[3]:.2f}, Max: {q.iloc[4]:.2f}"
+                )
+
+        # Case 2: Object or other hashable columns
+        elif is_hashable_series(col_data):
+            unique_vals = col_data.nunique()
+            if unique_vals <= sample_categorical_threshold:
+                col_type = "object-categorical"
+                top_vals = col_data.value_counts().head(10)
+                detail = ", ".join(f"{k} ({v})" for k, v in top_vals.items())
+            else:
+                col_type = "high-cardinality-categorical"
+                detail = f"{unique_vals} unique values"
+
+        # Case 3: Unusable columns
+        else:
+            col_type = "non-hashable"
+            detail = f"Non-hashable type: {type(col_data.iloc[0])}"
+
+        summary.append(
+            {
+                "Column": col,
+                "Dtype": dtype,
+                "Type": col_type,
+                "Detail": detail,
+                "Missing": total_missing,
+            }
+        )
+
+    return pd.DataFrame(summary)
+
+
+# Utility functions for data splitting
+# ===================================
+
+
+def split_data(
+    data,
+    experiment=None,
+    time_series=None,
+    date_column=None,
+    group_column=None,
+    val_size=None,
+    test_size=None,
+    target_numbers=None,
+    target_clf=None,
+    experiment_id=None,
+):
+    """
+    Utility function to split data into train, validation, and test sets.
+
+    Args:
+        data (pd.DataFrame): Input data to split
+        experiment: LeCrapaud experiment instance (preferred - extracts all params automatically)
+        time_series (bool): Whether to use time series splitting (overrides experiment)
+        date_column (str): Date column for time series splitting (overrides experiment)
+        group_column (str): Group column for time series splitting (overrides experiment)
+        val_size (float): Validation set size (0.0-1.0) (overrides experiment)
+        test_size (float): Test set size (0.0-1.0) (overrides experiment)
+        target_numbers (list): List of target numbers for stratification (overrides experiment)
+        target_clf (list): List of classification target numbers (overrides experiment)
+        experiment_id (int): Optional experiment ID to update sizes in database
+
+    Returns:
+        tuple: (train, val, test) DataFrames
+    """
+    # Extract parameters from experiment if provided
+    if experiment is not None:
+        # Check if it's a BaseExperiment or just the experiment database object
+        if hasattr(experiment, "context") and experiment.context:
+            # It's a database experiment object with context
+            context = experiment.context
+            if time_series is None:
+                time_series = context.get("time_series", False)
+            if date_column is None:
+                date_column = context.get("date_column")
+            if group_column is None:
+                group_column = context.get("group_column")
+            if val_size is None:
+                val_size = context.get("val_size", 0.2)
+            if test_size is None:
+                test_size = context.get("test_size", 0.2)
+            if target_numbers is None:
+                target_numbers = context.get("target_numbers", [])
+            if target_clf is None:
+                target_clf = context.get("target_clf", [])
+            if experiment_id is None:
+                experiment_id = experiment.id
+
+    # Set defaults if still None
+    if time_series is None:
+        time_series = False
+    if val_size is None:
+        val_size = 0.2
+    if test_size is None:
+        test_size = 0.2
+    if target_numbers is None:
+        target_numbers = []
+    if target_clf is None:
+        target_clf = []
+
+    dates = {}
+    if time_series:
+        (train, val, test), dates = _split_time_series(
+            data, date_column, group_column, val_size, test_size
+        )
+    else:
+        # Use first target for stratification if it's a classification target
+        stratify_col = None
+        if target_numbers and target_clf and target_numbers[0] in target_clf:
+            stratify_col = f"TARGET_{target_numbers[0]}"
+        train, val, test = _split_standard(data, val_size, test_size, stratify_col)
+
+    # Update experiment with sizes if experiment_id provided
+    if experiment_id:
+        Experiment.update(
+            id=experiment_id,
+            train_size=len(train),
+            val_size=len(val),
+            test_size=len(test),
+            **dates,
+        )
+
+    return train, val, test
+
+
+def _split_time_series(data, date_column, group_column, val_size, test_size):
+    """Time series splitting preserving temporal order."""
+    if not date_column:
+        raise ValueError("Please specify a date_column for time series")
+
+    df = data.copy()
+    if group_column:
+        df.sort_values([date_column, group_column], inplace=True)
+    else:
+        df.sort_values(date_column, inplace=True)
+
+    dates = df[date_column].unique()
+
+    val_first_id = int(len(dates) * (1 - val_size - test_size)) + 1
+    test_first_id = int(len(dates) * (1 - test_size)) + 1
+
+    train = df[df[date_column].isin(dates[:val_first_id])]
+    val = df[df[date_column].isin(dates[val_first_id:test_first_id])]
+    test = df[df[date_column].isin(dates[test_first_id:])]
+
+    dates = {}
+    for name, data in zip(["train", "val", "test"], [train, val, test]):
+        dates[f"{name}_start_date"] = (
+            data[date_column].map(pd.Timestamp.fromordinal).iat[0]
+        )
+        dates[f"{name}_end_date"] = (
+            data[date_column].map(pd.Timestamp.fromordinal).iat[-1]
+        )
+
+        logger.info(
+            f"{data.shape} {name} data from {dates[f'{name}_start_date'].strftime('%d/%m/%Y')} to {dates[f'{name}_end_date'].strftime('%d/%m/%Y')}"
+        )
+
+    return (
+        train.reset_index(drop=True),
+        val.reset_index(drop=True),
+        test.reset_index(drop=True),
+    ), dates
+
+
+def _split_standard(data, val_size, test_size, stratify_col=None, random_state=42):
+    """Standard random splitting with optional stratification."""
+    from sklearn.model_selection import train_test_split
+
+    df = data.copy()
+
+    stratify_vals = (
+        df[stratify_col] if stratify_col and stratify_col in df.columns else None
+    )
+
+    # First split: train + (val + test)
+    train, temp = train_test_split(
+        df,
+        test_size=val_size + test_size,
+        random_state=random_state,
+        stratify=stratify_vals,
+    )
+
+    # Adjust stratify target for val/test split
+    stratify_temp = (
+        temp[stratify_col] if stratify_col and stratify_col in df.columns else None
+    )
+
+    # Compute val and test sizes relative to temp
+    val_ratio = val_size / (val_size + test_size)
+
+    val, test = train_test_split(
+        temp,
+        test_size=1 - val_ratio,
+        random_state=random_state,
+        stratify=stratify_temp,
+    )
+
+    for name, data in zip(["train", "val", "test"], [train, val, test]):
+        logger.info(f"{data.shape} {name} data")
+
+    return (
+        train.reset_index(drop=True),
+        val.reset_index(drop=True),
+        test.reset_index(drop=True),
+    )