Supervised machine learning methods, in which no error labels are present, are increasingly popular methods for identifying potential data errors. Such algorithms rely on the tenet of a ‘ground truth’ in the data, which in other words assumes correctness in the majority of the cases. Points deviating from such relationships, outliers, are flagged as potential data errors.
This paper implements an outlier-based error-spotting algorithm using gradient boosting, and presents a blueprint for the modelling pipeline. More specifically, it underpins three main modelling hypotheses with empirical evidence, which are related to (1) missing value imputation, (2) the loss-function choice and (3) the location of the error. By doing so, it uses a cross sectional view on the loan-to-value and its related columns of the Credit Registry (Hitelregiszter) of the Central Bank of Hungary (MNB), and introduces a set of synthetic error types to test its hypotheses.
The paper shows that gradient boosting is not materially impacted by the choice of the imputation method, hence, replacement with a constant, the computationally most efficient, is recommended. Second, the Huber-loss function, which is piecewise quadratic up until the Huber-slope parameter and linear above it, is better suited to cope with outlier values; it is therefore better in capturing data errors. Finally, errors in the target variable are captured best, while errors in the predictors are hardly found at all. These empirical results may generalize to other cases, depending on data specificities, and the modelling pipeline described underscores significant modelling decisions.
Keywords: data quality, machine learning, gradient boosting, central banking, loss functions, missing values
JEL codes: C5, C81, E58