src/agri_predict/features.py

import pandas as pd
import numpy as np


def _ensure_datetime_index(df):
    df = df.copy()
    if not isinstance(df.index, pd.DatetimeIndex):
        df = df.set_index('Reported Date')
        df.index = pd.to_datetime(df.index)
    return df


def create_forecasting_features(df):
    df = _ensure_datetime_index(df)
    target_map = df['Modal Price (Rs./Quintal)'].to_dict()

    df['dayofweek'] = df.index.dayofweek
    df['quarter'] = df.index.quarter
    df['month'] = df.index.month
    df['year'] = df.index.year
    df['dayofyear'] = df.index.dayofyear
    df['weekofyear'] = df.index.isocalendar().week

    df['lag14'] = (df.index - pd.Timedelta(days=14)).map(target_map)
    df['lag28'] = (df.index - pd.Timedelta(days=28)).map(target_map)
    df['lag56'] = (df.index - pd.Timedelta(days=56)).map(target_map)
    df['lag_3months'] = (df.index - pd.DateOffset(months=3)).map(target_map)
    df['lag_6months'] = (df.index - pd.DateOffset(months=6)).map(target_map)
    for window in [7, 14, 28]:
        df[f'rolling_mean_{window}'] = df['Modal Price (Rs./Quintal)'].rolling(window=window, min_periods=1).mean()
        df[f'rolling_std_{window}'] = df['Modal Price (Rs./Quintal)'].rolling(window=window, min_periods=1).std()

    df['ema7'] = df['Modal Price (Rs./Quintal)'].ewm(span=7, adjust=False).mean()
    df['ema14'] = df['Modal Price (Rs./Quintal)'].ewm(span=14, adjust=False).mean()
    df['monthly_avg'] = df.groupby('month')['Modal Price (Rs./Quintal)'].transform('mean')
    df['weekly_avg'] = df.groupby('weekofyear')['Modal Price (Rs./Quintal)'].transform('mean')
    df['dayofweek_avg'] = df.groupby('dayofweek')['Modal Price (Rs./Quintal)'].transform('mean')

    df['fourier_sin_365'] = np.sin(2 * np.pi * df.index.dayofyear / 365)
    df['fourier_cos_365'] = np.cos(2 * np.pi * df.index.dayofyear / 365)
    df['fourier_sin_14'] = np.sin(2 * np.pi * df.index.dayofyear / 14)
    df['fourier_cos_14'] = np.cos(2 * np.pi * df.index.dayofyear / 14)

    df['recent_min_14'] = (df.index - pd.Timedelta(days=14)).map(target_map).min()
    df['recent_max_14'] = (df.index - pd.Timedelta(days=14)).map(target_map).max()
    df['recent_range_14'] = df['recent_max_14'] - df['recent_min_14']

    df['yearly_avg'] = df.groupby('year')['Modal Price (Rs./Quintal)'].transform('mean')
    df['cumulative_mean'] = df['Modal Price (Rs./Quintal)'].expanding().mean()

    return df.reset_index()


def create_forecasting_features_1m(df):
    df = _ensure_datetime_index(df)
    target_map = df['Modal Price (Rs./Quintal)'].to_dict()

    df['dayofweek'] = df.index.dayofweek
    df['quarter'] = df.index.quarter
    df['month'] = df.index.month
    df['year'] = df.index.year
    df['dayofyear'] = df.index.dayofyear
    df['weekofyear'] = df.index.isocalendar().week

    df['lag_30'] = (df.index - pd.Timedelta(days=30)).map(target_map)
    df['lag_60'] = (df.index - pd.Timedelta(days=60)).map(target_map)
    df['lag_90'] = (df.index - pd.Timedelta(days=90)).map(target_map)
    df['lag_6months'] = (df.index - pd.DateOffset(months=6)).map(target_map)
    df['lag_12months'] = (df.index - pd.DateOffset(months=12)).map(target_map)

    for window in [30, 60, 90]:
        df[f'rolling_mean_{window}'] = df['Modal Price (Rs./Quintal)'].rolling(window=window, min_periods=1).mean()
        df[f'rolling_std_{window}'] = df['Modal Price (Rs./Quintal)'].rolling(window=window, min_periods=1).std()

    df['ema_30'] = df['Modal Price (Rs./Quintal)'].ewm(span=30, adjust=False).mean()
    df['ema_60'] = df['Modal Price (Rs./Quintal)'].ewm(span=60, adjust=False).mean()

    df['monthly_avg'] = df.groupby('month')['Modal Price (Rs./Quintal)'].transform('mean')
    df['weekly_avg'] = df.groupby('weekofyear')['Modal Price (Rs./Quintal)'].transform('mean')
    df['dayofweek_avg'] = df.groupby('dayofweek')['Modal Price (Rs./Quintal)'].transform('mean')

    df['fourier_sin_365'] = np.sin(2 * np.pi * df.index.dayofyear / 365)
    df['fourier_cos_365'] = np.cos(2 * np.pi * df.index.dayofyear / 365)
    df['fourier_sin_30'] = np.sin(2 * np.pi * df.index.dayofyear / 30)
    df['fourier_cos_30'] = np.cos(2 * np.pi * df.index.dayofyear / 30)

    df['recent_min_30'] = (df.index - pd.Timedelta(days=30)).map(target_map).min()
    df['recent_max_30'] = (df.index - pd.Timedelta(days=30)).map(target_map).max()
    df['recent_range_30'] = df['recent_max_30'] - df['recent_min_30']

    df['yearly_avg'] = df.groupby('year')['Modal Price (Rs./Quintal)'].transform('mean')
    df['cumulative_mean'] = df['Modal Price (Rs./Quintal)'].expanding().mean()

    return df.reset_index()


def create_forecasting_features_3m(df):
    df = _ensure_datetime_index(df)
    target_map = df['Modal Price (Rs./Quintal)'].to_dict()

    df['dayofweek'] = df.index.dayofweek
    df['quarter'] = df.index.quarter
    df['month'] = df.index.month
    df['year'] = df.index.year
    df['dayofyear'] = df.index.dayofyear
    df['weekofyear'] = df.index.isocalendar().week

    df['lag_3months'] = (df.index - pd.DateOffset(months=3)).map(target_map)
    df['lag_6months'] = (df.index - pd.DateOffset(months=6)).map(target_map)
    df['lag_9months'] = (df.index - pd.DateOffset(months=9)).map(target_map)
    df['lag_12months'] = (df.index - pd.DateOffset(months=12)).map(target_map)

    for window in [90, 180, 270, 365]:
        df[f'rolling_mean_{window}'] = df['Modal Price (Rs./Quintal)'].rolling(window=window, min_periods=1).mean()
        df[f'rolling_std_{window}'] = df['Modal Price (Rs./Quintal)'].rolling(window=window, min_periods=1).std()

    df['ema90'] = df['Modal Price (Rs./Quintal)'].ewm(span=90, adjust=False).mean()
    df['ema180'] = df['Modal Price (Rs./Quintal)'].ewm(span=180, adjust=False).mean()
    df['monthly_avg'] = df.groupby('month')['Modal Price (Rs./Quintal)'].transform('mean')
    df['weekly_avg'] = df.groupby('weekofyear')['Modal Price (Rs./Quintal)'].transform('mean')
    df['dayofweek_avg'] = df.groupby('dayofweek')['Modal Price (Rs./Quintal)'].transform('mean')

    df['fourier_sin_90'] = np.sin(2 * np.pi * df.index.dayofyear / 90)
    df['fourier_cos_90'] = np.cos(2 * np.pi * df.index.dayofyear / 90)
    df['fourier_sin_30'] = np.sin(2 * np.pi * df.index.dayofyear / 30)
    df['fourier_cos_30'] = np.cos(2 * np.pi * df.index.dayofyear / 30)

    df['recent_min_90'] = (df.index - pd.Timedelta(days=90)).map(target_map).min()
    df['recent_max_90'] = (df.index - pd.Timedelta(days=90)).map(target_map).max()
    df['recent_range_90'] = df['recent_max_90'] - df['recent_min_90']

    df['yearly_avg'] = df.groupby('year')['Modal Price (Rs./Quintal)'].transform('mean')
    df['cumulative_mean'] = df['Modal Price (Rs./Quintal)'].expanding().mean()

    return df.reset_index()