import pandas as pd

def detect_ruptures(df, epsilon=0.05):
    # Colonnes clés pour identifier les comptes
    key_cols = [
        'Agreement - Code',
        'Company - Id',
        'Company - Ultimate Parent Id',
        'Registrar Account - Region',
        'RegistrarAccount - Country',
        'Registrar Account - ID'
    ]
    
    # Travailler sur une copie
    df_temp = df.copy()

    # Colonnes de dates
    df_temp['Centralisation Date'] = pd.to_datetime(df_temp['Centralisation Date'])

    # Dates distinctes
    full_dates = (
        pd.Series(df_temp['Centralisation Date'].unique())
        .sort_values()
        .reset_index(drop=True)
    )

    # Combinaisons comptes × dates
    accounts = df_temp[key_cols].drop_duplicates()
    full_index = accounts.merge(
        pd.DataFrame({'Centralisation Date': full_dates}), 
        how='cross'
    )

    # Agréger les AUM par clé
    agg_cols = key_cols + ['Centralisation Date']
    df_agg = (
        df_temp.groupby(agg_cols)['Value - AUM €']
        .sum()
        .reset_index()
    )

    # Merge sur toutes les combinaisons
    df_full = pd.merge(full_index, df_agg, on=agg_cols, how='left')

    # Remplissage des trous par 0
    df_full['Value - AUM €'] = df_full['Value - AUM €'].fillna(0)

    # Tri
    df_full = df_full.sort_values(key_cols + ['Centralisation Date'])

    # Variation et valeur précédente
    df_full['AUM_diff'] = df_full.groupby(key_cols)['Value - AUM €'].diff().fillna(0)
    df_full['prev_value'] = df_full.groupby(key_cols)['Value - AUM €'].shift(1).fillna(0)

    # Comptes qui perdent tout
    df_zero = df_full[(df_full['AUM_diff'] < 0) & (df_full['Value - AUM €'] == 0)].copy()

    # Comptes qui partent de 0
    df_from_zero = df_full[(df_full['AUM_diff'] > 0) & (df_full['prev_value'] == 0)].copy()

    # Colonnes pour le merge (sans ID)
    merge_cols = [
        'Centralisation Date',
        'Agreement - Code',
        'Company - Id',
        'Company - Ultimate Parent Id',
        'Registrar Account - Region',
        'RegistrarAccount - Country'
    ]

    # Détection des ruptures
    ruptures = pd.merge(df_zero, df_from_zero, on=merge_cols, suffixes=('_old','_new'))

    # Calcul de la différence relative selon epsilon
    ruptures['diff_rel'] = abs(ruptures['AUM_diff_old'] + ruptures['AUM_diff_new']) / (
        (abs(ruptures['AUM_diff_old']) + abs(ruptures['AUM_diff_new'])) / 2
    )

    # Filtrage avec epsilon
    ruptures = ruptures[ruptures['diff_rel'] <= epsilon].drop(columns=['diff_rel'])

    # Colonnes finales
    ruptures_df = ruptures[['Centralisation Date','Registrar Account - ID_old','Registrar Account - ID_new','AUM_diff_new']]
    ruptures_df.columns = ['date','old_account','new_account','value']

    return ruptures_df