Custom loss functions for XGBoost using PyTorch

Data Science
3

Has someone tried to perform cross validation to this model?

The snippet below can replace the last line of mdo code…

param_fit_grid = { 'base_margin' : base_margin}

score = model_selection.cross_val_score(
                model_adj_sharpe,
                train[feature_columns],
                train[target],
                cv=3,
                n_jobs=-1,
                scoring=make_scorer(mean_squared_error),
                fit_params=param_fit_grid,
                error_score=123)

print(score)

However, it returns my error_score=123, after some investigation I guess the problem occurs here:

    # get correlations in each era
    for ee in era_idx:
        score = corr(ypred_th[ee], ytrue_th[ee])

More exactly ypred_th[ee], apparently after 1 successful cv-fold, it can´t find the the respective index on ypred_th tensor.

Moreover, if you replace de the objective function to squared_log example function as shown on xgb docs it works fine on cross_val_score.

One more thing

The order of the parameters in adj_sharpe_obj(ytrue, ypred) is flipped according to xgb_docs standards, not sure if it can create any noise.