| Marine finance is a critical component of the maritime economy and is playing an increasingly important role in financing,risk allocation,and resource transfer.The development of marine finance provides complete financial services and support to the Chinese shipping industry.Sustainable development of the maritime industry can also provide a steady stream of business needs and development momentum,thereby further promoting the development of marine finance.In the field of marine finance,marine insurance is particularly vital.Its functions include transferring ocean transportation risk,sharing and compensating losses,and other risk management activities that act as an intermediary between the financial and maritime sectors.It offers effective protection for a series of risk factors involved in the import and export of goods in China.A well-developed marine insurance industry can reduce uncertainty in cargo transportation,increase demand for shipping,and promote the development of import and export trade.Therefore,the development level of marine insurance has significant practical significance in implementing the national strategy of "sea power".It reinforces the importance of ongoing efforts to strengthen the development of China’s maritime economy,including comprehensive marine finance services and a flourishing marine insurance industry.The data from the International Union of Marine Insurance(IUMI)in 2017 reveals that China has become the world’s leading hull insurance market,with its premium income reaching 5.771 billion yuan by the year 2020.From a macro perspective,marine insurance plays a pivotal role in supporting the shipping industry and import-export businesses.During the critical period of China’s industrial transformation,analyzing basic data of marine insurance horizontally and vertically can reflect the monitoring data requirements of national strategic transformation.This reflects the growing need to support the implementation of China’s national revitalization strategy to the fullest extent possible.From a micro perspective,the hull insurance industry has shown a decrease in premium rates and a disorderly increase in risk exposure in the past four years,thereby resulting in loss-making.This situation highlights the urgent need to improve the pricing mechanism of hull insurance,determine premium rates scientifically,and improve the competitiveness of the industry.In conclusion,improving the pricing capabilities of hull insurance is crucial to ensure the healthy development of China’s marine insurance industry.It is essential to enhance the industry’s resilience,ensure its sustainable and stable growth,and realize its potential contribution to the wider maritime economy.The monopoly of companies,such as Lloyd’s,in marine insurance business,coupled with the uniqueness of hull insurance,and the limited availability of hull insurance data,has led to limited literature on the determination of hull insurance rates.To address this gap,this dissertation focuses on claim frequency prediction in hull insurance,drawing from theoretical and empirical aspects.The dissertation compares and analyzes the prediction and interpretability results of different models to determine the actuarial prediction model with the best fitting effect on the claim frequency of hull insurance.This modeling approach provides valuable insights into the determination of hull insurance premium rates for Chinese insurance companies.Theoretical research can be divided into theoretical research on actuarial prediction models and theoretical research on interpretability methods:Theoretical research on actuarial prediction models includes generalized linear models and models based on machine learning methods.Theoretical research on generalized linear models includes exponential distribution family,zero-inflation method,parameter estimation method and loss function setting of GLM,etc.Theoretical research on models based on machine learning methods mainly includes theories of regression tree,random forest and Boosting methods.Due to the lack of corresponding theoretical literature support for the prediction of claim frequency in hull insurance,this dissertation draws on the more mature theory of rate determination in the field of non-life insurance,combined with the inherent characteristics of hull insurance,to construct an actuarial prediction model suitable for hull insurance claim frequency.Theoretical research on interpretability methods aims to explain the prediction results of models based on machine learning methods and compare them with generalized linear models.The research content of the interpretability method includes variable importance,partial dependence,individual conditional expectation and interaction effect.The variable importance reflects the contribution degree of different explanatory variables to the prediction results,partial dependence and individual conditional expectation display the marginal prediction results of different variables,and the interaction effect measures the influence of the correlation between explanatory variables on the prediction results.Building on the theoretical frameworks,this dissertation employs R programming to conduct empirical analysis of various actuarial prediction models based on policy and claim data for hull insurance.The analysis yields fitting results for multiple models,while interpretability methods are utilized to explain the model outcomes.Univariate prediction analysis is also conducted to determine the claim frequency prediction trends for different explanatory variables in various models.By comparing the fitting results,interpretability analysis,and univariate forecast outcomes,the most appropriate actuarial prediction model is established for predicting the frequency of hull insurance claims.The main innovations of this dissertation are the following three aspects:The initial step of this study involves identifying and categorizing hull insurance-related risk factors,in line with non-life actuarial principles.The study then proceeds to predict hull insurance claim frequency employing generalized linear models and machine learning-based models.Compared to non-life insurance pricing research literature for other forms of insurance,the literature pertaining to risk factor identification and claim frequency prediction for hull insurance remains limited.As such,this study relies on generalized linear models,regression tree,random forest,gradient boosting trees,and XGBoost to forecast the claim frequency of hull insurance.Furthermore,a unified deviation calculation approach is utilized to evaluate all models,contributing to the broadening of non-life actuarial pricing theory and empirical research on hull insurance.In addition to prediction modeling,this study employs several interpretability techniques,including variable importance,partial dependence,individual conditional expectation,and Friedman’s H statistic.These methods enable in-depth analysis of the model results,providing a broader comparison perspective between traditional regression models and those based on the machine learning method.The fitting results of various models concerning hull insurance claim frequency are evaluated from multiple viewpoints,facilitating a better understanding of the prediction outcomes of the prediction model.The dissertation utilizes a comparative analysis approach to summarize the results of all actuarial prediction models.The models are evaluated and compared based on three key aspects,which include model forecasting performance,interpretability,and univariate forecasting.Ultimately,the optimal actuarial prediction model for hull insurance claim frequency is determined.The key indicators of model fitting effectiveness mainly comprise mean square error,in-sample deviation,and out-of-sample deviation.On the other hand,the comparison of interpretability results primarily centers on variable importance,which helps pinpoint the variables that contribute most significantly to the prediction.Additionally,the comparative analysis of univariate prediction results offers an intuitive way to analyze the claim frequency trend prediction of each explanatory variable.The main research conclusions of this dissertation include:First,adding interaction terms to the generalized linear model leads to a significant improvement in performance,with the generalized linear model under the zero-inflated negative binomial distribution being the most accurate model in predicting hull insurance claim frequency.And it was observed that the machine learning method took significantly longer to run compared to the generalized linear model,and the results were greatly influenced by preset parameters.To address this,the dissertation utilized a grid parameter optimization method that comprehensively considered three factors,namely prediction performance,running time,and robustness.Overall,the performance of machine learning methods was found to be generally superior to that of GLM,with gradient boosted trees being the best prediction model,followed by XGBoost,random forest,and regression tree.Second,the variable importance results indicated that the importance of variables in different models was relatively similar.The insurance amount,deductible ratio,tonnage,and ship type were found to have the most significant impact on the prediction results,whereas the coverage and age of the vessel had the lowest importance.Interestingly,after removing the coverage and age of the vessel,the mean square error and deviation of all models increased slightly,indicating that the seemingly "unimportant" explanatory variables still had a significant impact on the prediction results.The study thus emphasizes the importance of cautious variable screening in practical applications,as well as the need for comprehensive analysis of all available explanatory variables.Third,interpretability of the model was analyzed using partial dependence and individual conditional expectation.It was observed that the regression tree had poor interpretability,as the partial dependence could not accurately reflect the average forecast trend.However,the interpretability of other models,such as gradient boosted trees,XGBoost,and random forest,was relatively good.Fourth,the use of machine learning methods allows for complex interactions between explanatory variables,resulting in greater interaction effects.The study found that the interaction effect between tonnage,insurance amount,ship type,and other explanatory variables was the most prominent.To analyze the influence of interaction effects on the prediction of claim frequency,the dissertation utilized grouped partial dependence plots.These plots allowed for a detailed understanding of the interaction effects between explanatory variables and their impact on the prediction results.The fifth finding of the study indicates that based on univariate prediction results,no model was able to accurately predict the claim frequency trend of all variables.However,the gradient boosting tree model showed the best prediction performance by accurately predicting the trend of claim frequency of five variables.Additionally,the study revealed that the difference between the univariate predicted value and the actual observed value for each model was relatively small.This indicates that each model was able to generate reliable and consistent results.Based on the above conclusions,this dissertation puts forward suggestions on the product development and pricing of Chinese insurance company’s hull insurance business:Firstly,it is recommended that insurance companies standardize the policy and claims data for hull insurance based on their specific circumstances.They should then make optimal use of this data to identify and differentiate risk factors across multiple dimensions.Once classified,they can apply commonly used non-life insurance actuarial models for tailored pricing strategies,thereby improving their competitiveness in the hull insurance market.Secondly,it is advisable for insurance companies to explore machine learning techniques that offer superior predictive performance for actuarial pricing of hull insurance.It is recommended that they utilize different interpretability methods to provide reasonable explanations for the prediction results.In pricing marine insurance,it is important to consider not only the model’s predictive performance but also its running time and potential overfitting issues.A comprehensive analysis of these three factors can help identify the most suitable model. |
PDF Full Text Request