Calibrating TabTransformer for financial misstatement detection

In this paper, we deal with the task of identifying the probability of misstatements in the annual financial reports of public companies. In particular, we improve the state-of-the-art for financial misstatement detection by training a TabTransformer model with a gated multi-layer perceptron, which encodes and exploits relationships between financial features. We further calibrate a sample-dependent focal loss function to deal with the severe class imbalance in the data and to focus on positive examples that are hard to distinguish. We evaluate the proposed methodology in a realistic setting that preserves the essential characteristics of the task: (a) the imbalanced distribution of classes in the data, (b) the chronological order of data, and (c) the systematic noise in the labels, due to the delay in manually identifying misstatements. The proposed method achieves state-of-the-art results in this setting, compared to recent approaches in the literature. As an additional contribution, we release the dataset to facilitate further research in the field.