Autoencoder Based Anomaly Detection of Indonesia's Leading Banks

Penelitian

Abstract

This study This study builds and tests an autoencoder model that hunts for multivariate anomalies inside the financial data of Indonesia's banking sector. Three issuers anchor the analysis, PT Bank Mandiri (Persero) Tbk (BMRI), PT Bank Rakyat Indonesia (Persero) Tbk (BBRI) and PT Bank Central Asia Tbk (BBCA). The goal is not fraud detection in the narrow sense but something broader, isolating financial behavior that breaks the pattern investors expect and that might signal deeper trouble. The method layers three techniques. Correlation analysis maps how variables move together. Z Score standardization exposes univariate outliers. An autoencoder trained on an 80 to 20 split between training and test data hunts for anomalies that only reveal themselves across several variables at once. On the test set, BMRI and BCA produced zero anomalies and earned a low-risk label. BBRI stood apart. It logged two anomalies out of thirteen test observations, a rate of 15.38 percent that places it in the moderate risk bracket. Both anomalous quarters, the fourth quarter of 2008 and the first quarter of 2011, trace back to sharp swings in the Z Scores for Revenue, Net Income, Operating Cash Flow, Investing Cash Flow and Financing Cash Flow. The results argue for the autoencoder as a genuinely capable tool for multivariate anomaly detection in financial data. They offer real insight into how issuer risk profiles differ. Still, the test set is small. That limits how far these conclusions should be generalized.