Explaining Anomalies using Denoising Autoencoders for Financial Tabular Data

Denoising autoencoders can pinpoint specific data errors: Rather than simply flagging an entire record as anomalous, the framework identifies individual cells (data points) within financial records that contain errors and assigns confidence scores to each, making it actionable for domain experts reviewing hundreds of mixed-type columns.

The approach provides corrective suggestions: Beyond detecting errors, the model estimates what the correct values should be, enabling financial institutions to not only identify problematic data but also potentially fix it automatically or with minimal human intervention.

Performance validated across multiple datasets: The technique was tested on three public datasets (Credit Default, Adult, IEEE Fraud) plus proprietary Holdings data, with results showing the denoising autoencoder approach outperformed other methods in both error detection rates and accuracy of suggested corrections.

Addresses a real gap in financial AI explainability: The paper highlights that existing anomaly detection systems often lack the interpretability needed for financial domain experts, who need to understand why a transaction or record is flagged and what specifically is wrong—not just receive a binary anomaly alert.

Specialized loss functions improve reliability: The framework demonstrates that designing loss functions specifically for cell-level error detection further improves performance metrics, suggesting that tailored technical approaches can enhance both accuracy and trustworthiness of AI systems in regulated industries.