Data Imputation in Large Datasets: A Comparative Study of PCA and Machine Learning Approaches

Abstract This paper contributes to the financial econometrics literature by providing a systematic comparison of machine learning techniques and principal component analysis (PCA) methods for data imputation in large financial datasets. Missing data is pervasive in empirical finance and economics, and imputation accuracy determines whether the entire dataset can be used in analyses such as modeling credit risk. We introduce a fully linear autoencoder with a loss function tailored to observed values, which performs on par with PCA across various missingness mechanisms while offering greater flexibility. Although random forest-based methods are less accurate in large datasets with a strict factor structure, they demonstrate superior performance in lower-dimensional settings where the primary goal is outcome prediction. This is particularly relevant in applications such as credit default prediction, where identifying risk factors from incomplete borrower data is the main objective. In two such examples, the MissForest random forest technique outperforms others by achieving lower imputation error and higher predictive accuracy.