Türkiye'de Otomobil Sigortası Sahtekârlıklarının Makine Öğrenmesi Yöntemleri ile Tespit Edilmesi
xmlui.mirage2.itemSummaryView.MetaDataShow full item record
The insurance sector has inherent high risk of fraudulent activity due to its trust-based nature. Insurance fraud is often considered to be less risky and easier than other criminal activities like robbery and smuggling. This has also caused the fraudulent behavior to be widespread in this domain. Since insurance frauds are relatively more common and difficult to detect, it has significant impact on insurance companies and other stakeholders in this sector. Even simple frauds may lead to a financial burden that is not expected by the insurer, and this impairs the financial balance of the insurance sector. As a result, deviations occur in the actuarial calculations, claim reserves become insufficient. At the end of the day, increased insurance costs are paid as insurance premiums by the insured who are not guilty. Insurance fraud is one of the most important problems of the sector in Turkey as well as in other countries. Awareness about fighting against insurance fraud has increased for both insurance companies and regulatory authorities, as they have started to establish units related to fraud and increased the controls on this subject. This study aims to detect insurance frauds in Turkey by using a large dataset. Since the automobile insurance covers a significant share of the sector, the registration systems for automobile insurance data are more advanced in Turkey, and data can be recorded more accurately. However, previous studies that use comprehensive autmobile insurance data to detect fraud in Turkey are not available. We aim to understand the important variables related to automobile insurance fraud, and evaluate the performance of different machine learning algorithms in this domain. We have applied various statistical and machine learning methods that are used for similar problems in the previous literature (Logistic Regression, Decision Trees, Artificial Neural Networks, Support Vector Machines). In addition, we have also used Bayesian Networks with three different learning algorithms, which has not been applied for automobile insurance fraud detection in previous studies. All methods had high performance for fraud detection in our evaluations but none of them were superior to the other.