Provide gridsearch support for time series

[SamVanhoutte]

See how to leverage this method (I once built) to gridsearch time series prediction algorithms (which are not compatible with GridSearchCV).

Starting from this code:

def GridSearchModel(model, fit_params: dict, 
                    X_fit: pd.Series, y_fit: pd.Series, y_val: pd.Series, horizon:int,
                    ctor_params: dict = None, validation_method = None, verbose=False):
    __best_error_score = 10000000
    __best_params = {}
    __best_predictions = None
    __best_fitted_model = None
    
    y_val = np.array(y_val)
    
    param_combinations = list(product(*(fit_params[par_name] for par_name in fit_params)))
    # Looping through all constructor possibilities
    for param_combination in param_combinations:
        kwargs = dict()
        for par_idx, param_name in enumerate(fit_params):
            kwargs[param_name] = param_combination[par_idx]
        if verbose:
            print('Applying parameter set', kwargs)
        model_fit = model.fit(**kwargs)
        __current_error_score = 0
        __y_pred = np.array(model_fit.predict(len(X_fit), len(X_fit) + horizon - 1))
        if validation_method == None or validation_method.lower() == 'mape':
            # Mean absolute percentage error (MAPE)
            __current_error_score += np.mean(np.abs((y_val - __y_pred) / y_val)) * 100
        elif validation_method == None or validation_method.lower() == 'mae':
            # Mean absolute error (MAE)
            __current_error_score += np.mean(abs(y_val-__y_pred))
        elif validation_method == None or validation_method.lower() == 'rmse':
            # Root mean squared error (RMSE)
            __current_error_score += math.sqrt(np.mean(y_val-__y_pred)**2)
        
        if __current_error_score < __best_error_score:
            print(f'Better score of {__current_error_score} with parameters', kwargs)
            __best_error_score = __current_error_score
            __best_params = kwargs
            __best_predictions = __y_pred
            __best_fitted_model = model_fit
    
    return __best_fitted_model, __best_params, __best_predictions, __best_error_score