İflas Olasılığının Genelleştirilmiş Doğrusal Modeller ve Birleşik Aktüeryal Yapay Sinir Ağları Yöntemleri İle Değerlendirilmesi: Tamamlayıcı Sağlık Sigortası Uygulaması
Date
2024-02-16Author
Okunakol, Nermin Ödül
xmlui.dri2xhtml.METS-1.0.item-emb
Acik erisimxmlui.mirage2.itemSummaryView.MetaData
Show full item recordAbstract
In the insurance business, pricing is essential, especially for health insurance, where risk factors are taken into account significantly. Pricing is important because it can fairly estimate and distribute the cost of future health-related incidents. The cost of health insurance is determined by taking into consideration a number of risk variables, including age, lifestyle choices, pre-existing diseases, and local healthcare expenses. Insurance firms can set prices that represent the anticipated cost of meeting a person's or a group's healthcare needs by including these factors into the pricing model. Furthermore, accurate pricing helps insurers control risk by guaranteeing that premiums are high enough to pay for expected claims without compromising revenue. Inadequate pricing of health insurance policies may result in adverse selection, which increases the likelihood of enrollment for people with greater health risks and raises claims expenses and insurer losses. In light of this, risk-based pricing is crucial to preserving the health insurance market's sustainability and stability, which eventually benefits both insurers and policyholders. This study aims to use Generalized Linear Models (GLM) and Machine Learning (ML) methodologies in the field of health insurance for advanced risk assessment and premium calculation. Traditional actuarial approaches, typically based on GLMs, provide a robust foundation for modeling the relationship between various risk factors and insurance claims. Additionally, with the emergence of advanced machine learning techniques, there are increased opportunities to enhance prediction accuracy and capture complex models in the health insurance environment. In this context, the objective is to strengthen the model's prediction by integrating GLM outputs into artificial neural networks, a different approach in addition to the frequently used GLM in insurance product pricing by actuaries.
The thesis begins with a comprehensive examination of GLMs, emphasizing their applications in actuarial science and the field of health insurance. Subsequently, it explores the integration of artificial neural networks from machine learning algorithms to complement and expand the predictive capabilities of GLMs, explaining the Combined Actuarial Neural Networks (CANN). This study investigates how the combination of models and algorithms can effectively address nonlinear relationships, interactions between variables, and capture complex patterns in data. The proposed approach's analysis is conducted using real complementary health insurance data from a Turkish insurance company. Initially, the data structure is examined, and adjustments are made to fit the modeling, such as setting upper limits for claim numbers and categorizing some continuous variables.
Under the assumption of a classical risk model, a CANN model for both frequency and severity is developed, and the most suitable models are determined with significant explanatory variables for GLM. The obtained GLM outputs are aligned to the layered structure of artificial neural networks as input neurons. This process has increased the explainability and prediction power of the so-called black box models of artificial neural networks, implemented separately for frequency and severity models. The study also evaluates the performance of the GLM-Machine Learning combination against traditional GLMs in terms of prediction accuracy, model interpretability, and adaptability to dynamic health data. The impact of integrating machine learning techniques with established GLM methodologies is measured by examining the effect of modeling results using classic GLM on the profitability of the insurance company and comparing bankruptcy probabilities according to both approaches. Additionally, in the application phase of the study, the effects of incorporating machine learning into actuarial workflows are analyzed, considering both technical aspects and practical applicability in the insurance sector.
The study aims to investigate the applicability of combined actuarial neural networks as opposed to advanced modeling approaches to enhance risk assessment accuracy and encourage more informed decision-making processes in the health insurance sector.