[PIST6(競輪)を機械学習で儲かるか検証した話] Section 03-03 : 利益の検証とアンサンブル

def print_win_money_line(models, train_X, train_y, test_X, test_y, not_use_df, train_mask, test_mask):
    train_payout = not_use_df.loc[train_mask, 'ticket_payout'].reset_index(drop=True)
    test_payout = not_use_df.loc[test_mask, 'ticket_payout'].reset_index(drop=True)
    train_preds = [model.predict(train_X) for model in models]
    test_preds = [model.predict(test_X) for model in models]

    train_payout = train_payout.fillna(0)
    test_payout = test_payout.fillna(0)

    def calc_win_money_line(preds, hit, payout, th=0.5):
        bets = np.array([(pred > th).astype(int) for pred in preds])
        mode_bet, _ = stats.mode(bets, axis=0)
        mode_bet = mode_bet[0]
        hit = hit.to_numpy()
        payout = payout.to_numpy()

        return_money = mode_bet * hit * payout
        bet_money = mode_bet

        win_money_line = return_money - bet_money
        return np.cumsum(win_money_line)

    train_win_money_line = calc_win_money_line(train_preds, train_y, train_payout, 0.5)
    test_win_money_line = calc_win_money_line(test_preds, test_y, test_payout, 0.5)

    plt.figure(tight_layout=True, figsize=(12, 6))
    plt.subplot(121)
    plt.plot(train_win_money_line, label='train')
    plt.legend()
    plt.xlabel('tickets(day)')
    plt.ylabel('win_money')
    plt.subplot(122)
    plt.plot(test_win_money_line, label='test')
    plt.legend()
    plt.xlabel('tickets(day)')
    plt.ylabel('win_money')

    plt.show()

trial_num = 5
import pickle

def analysis_columns_effect(df, not_use_columns, not_use_columns_keywards, ticket_type):
    train_X, train_y, test_X, test_y, not_use_df, train_mask, test_mask = make_learning_data(df, not_use_columns, not_use_columns_keywards, ticket_type)

    models = []
    for n in range(trial_num):
        upsampint_X, upsampling_y = up_sampling(train_X, train_y, ticket_type)
        model, (lgb_train_X, lgb_train_y, lgb_valid_X, lgb_valid_y) = learning(upsampint_X, upsampling_y)
        print_win_money([model], train_X, train_y, test_X, test_y, not_use_df, train_mask, test_mask)
        print_win_money_line([model], train_X, train_y, test_X, test_y, not_use_df, train_mask, test_mask)
        models.append(model)
    print('---------ensemble result-------')
    print_win_money(models, train_X, train_y, test_X, test_y, not_use_df, train_mask, test_mask)
    print_win_money_line(models, train_X, train_y, test_X, test_y, not_use_df, train_mask, test_mask)

    with open(f'work/models/lgbm_{ticket_type}.pkl', 'wb') as f:
        pickle.dump(models, f)