OPTIMIZING TRAVEL INSURANCE PURCHASE DETECTION USING PREDICTIVE MODELS
As air travel remains a fundamental aspect of contemporary connectedness for individuals, enterprises, and governments, safeguarding travellersā€™ safety and financial security has become progressively essential. Travel insurance is a necessary protection against unforeseen financial losses during domestic or foreign journeys, encompassing trip cancellations, lost baggage, and emergency medical evacuations. The need for travel insurance has increased markedly in recent years. The United States Travel Insurance Association reports that Americans expended over $4 billion on travel insurance in 2018, signifying a 41% rise from 2016, with the number of insured individuals growing by over 10.7% from 2021. The rising quantity of insured passengers highlights an enhanced recognition of the advantages of these policies.
The COVID-19 pandemic underscored the essential importance of travel insurance. The extensive travel restrictions impeded global plans, rendering numerous uninsured tourists financially precarious. Forbes (2020) indicated that airline disruptions affected 62% of American people, resulting in financial losses for many due to insufficient insurance coverage. Although passengers obtained coupons or refunds, these instances underscored the financial perils encountered by uninsured travellers. In the post-pandemic era, sporadic disruptions such as weather-related cancellations and strikes have emphasized the essential role of travel insurance in trip planning. As a result, policies providing trip cancellation and interruption compensation comprised over 90% of travel protection products acquired in 2018.
Although the uptake of travel insurance is rising, research on the determinants of its purchasing behaviour is still scarce. Current research predominantly highlights qualitative assessments or concentrates on the medical dimensions of insurance. Although recent studies by Karl Kerr and Kelly explore risk perception and uncertainty in travel decisions, they neglect to consider the predictive modelling of insurance purchases. This gap necessitates a more profound comprehension of the demographic, social, and economic elements influencing consumer behaviour. Creating specialized insurance programs requires insights from predictive analytics, providing insurers with practical solutions to enhance product design and marketing.
RELATED STUDIES IN INSURANCE AND PREDICTIVE MODELS
Predictive modelling, sometimes called data mining, is described as extracting valuable information from existing data by statistical and machine learning methodologies. Despite the plethora of data mining techniques and applications available today, insurance, finance, and economics research remains scarce. A potential explanation for this is the scarcity of research data. Utilized six data mining strategies to assess the predicted accuracy of credit card client defaults, employing the sorting smoothing methodology to identify the artificial neural network as the most effective model. Recent prominent finance literature on credit fraud detection identifies the following classification algorithms as optimal: artificial neural networks, support vector machines, discriminant analysis, k-nearest neighbours, logistic regression, Bayesian learning, and random forests. The primary advantage of employing machine learning technology in the insurance industry is its capability to manage data sets effectively. All forms of data, including structured, unstructured, and semi-structured, can be altered by machine learning. The application of machine learning relies on the value chain, encompassing advanced precision, feature engineering, and, qualitatively, consumer behaviour.
TRAVEL INSURANCE LITERATURE
Risk perception and demographic parameters are the primary determinants of the propensity to purchase travel insurance. Indeed, an increased perception of danger correlates with a heightened need for insurance. Cheung and Law identified that acquiring insurance, providing additional funds, and pursuing current information regarding the destination were the three predominant tactics employed to mitigate risks in future travel planning. Experienced travellers were more inclined to employ these three tactics compared to inexperienced travellers, who preferred to consult family or friends, seek guidance from their travel agency, and participate in organized tour groups rather than travel independently.
The investigation into the correlation between risk perception and the decision to travel to China revealed that the propensity to acquire travel insurance was influenced by risk perception, readiness to pay a premium, duration of stay, and elevated monthly income. Al Mamun et al. investigated the willingness to purchase travel insurance for international travel among working adults during the COVID-19 pandemic through partial least squares modelling. They concluded that factors influencing attitudes towards travel insurance include insurance literacy regarding the product, perceived health risks associated with travel, and individual health consciousness. They advocate for enhancing travel insurance literacy and educating working persons about perceived health risks to motivate acquiring travel insurance plans for international travel.
PREDICTIVE MODELS
Classification models are frequently utilized for categorical response variables, rendering them appropriate for this study’s emphasis on a binary classification problem with two classes. Recent breakthroughs in machine learning for insurance modeling have offered novel methodologies to enhance classification precision and interpretability. Significant contributions encompass Althati, which investigates ensemble approaches for underwriting optimization, researches neural network architectures for estimating claim probabilities and employs explainable AI to improve decision-making in policy pricing. These studies emphasize the necessity of utilizing advanced categorization technologies to tackle particular issues in the insurance sector. This research examines the predictive accuracy of the categorization models presented, adjusting and comparing their efficacy in predicting travel insurance purchases.
LOGISTIC REGRESSION
Logistic regression quantifies the association between a categorical response variable and independent factors by estimating probabilities through a logistic function representing a cumulative logistic distribution. The conditional distribution of š¯‘¦ given x follows a Bernoulli distribution due to the binary nature of the dependent variable. It serves as an alternative to Fisher’s 1936 method, linear discriminant analysis, although it does not necessitate the multivariate normality assumption of the latter. It is widely recognized and user-friendly and remains one of the most utilized tools for data mining. It offers a valuable benchmark for evaluating the efficacy of novel methods and generating a straightforward probabilistic classification formula. The primary benefit of this method is its ability to develop a straightforward probabilistic classification formula.
K-NEAREST NEIGHBOURS (KNN)
The k-nearest Neighbours algorithm is a non-parametric, lazy learning technique for supervised learning, wherein an item is classified based on the majority vote of its neighbours, assigning the object to the most prevalent class among its k-nearest neighbours. In learning systems, generalization performance is influenced by a trade-off between the number of training instances and the learning algorithm’s capacity (e.g., the number of parameters). The primary benefit is that establishing a prediction model before classification is unnecessary. KNN does not train upon receiving the training data; instead, it merely retains the data without executing any computations during the training phase. Consequently, it possesses a lazy-learner characteristic. A model is not constructed until a query is executed on the dataset. KNN is optimal for data mining.
DISCRIMINANT ANALYSIS
Discriminant Analysis, also called Fisher’s rule, is a classification method that projects n-dimensional data onto a line by maximizing the mean between classes and minimizing the variation within classes, hence facilitating classification in this one-dimensional space. Two prevalent versions of Discriminant Analysis utilized in data mining are Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA). James, Witten, Hastie, and Tibshirani assert that Linear Discriminant Analysis (LDA) presumes the predictor originates from a multivariate Gaussian distribution characterized by a class-specific mean vector and a shared covariance matrix. This model is necessary because well-separated classes lead to unexpectedly unstable parameter estimates in the logistic regression model, the most commonly employed classical classification method. LDA does not encounter this issue and is even more favored, with many response classes exceeding two. Similar to the LDA, the QDA classifier is predicated on the assumption that observations from each class are sampled from a Gaussian distribution. Quadratic Discriminant Analysis (QDA) posits a distinct covariance matrix for each class, unlike Linear Discriminant Analysis (LDA). Recent methodologies encompass Shrinkage Discriminant Analysis (SDA) and Penalised Discriminant Analysis (PDA), yielding results congruent with the logistic classifier in this study.
CONCLUSION
This study examined the utilization of machine learning algorithms to forecast travel insurance acquisitions, providing substantial benefits to the insurance and aviation industries. The study identified critical indicators, including age, wealth, travel history, and degree status, demonstrating that advanced classifiers such as XGBoost, random forests, and gradient boosting surpass standard models like logistic regression in predicted accuracy and dependability. These findings offer practical insights for insurers to create customized insurance products, improve client targeting, and refine marketing methods. Additionally, airline businesses can utilize this knowledge to create customized travel insurance marketing, enhancing client conversion rates. Travelers benefit from cost-effective and tailored coverage, promoting a more inclusive insurance market.
This research highlights the significance of variable selection in predictive modelling beyond its practical uses. Emphasizing parsimonious models that prioritize the most pertinent variables mitigates overfitting and improves interpretability, ensuring solid predictions across various datasets.
This study offers valuable insights but recognizes limitations in data timeliness and breadth. The dataset, while enough, would improve with more recent, comprehensive data that includes additional relevant variables such as airline costs, insurance premiums, domestic versus foreign travel, and socio-economic aspects. Future research may rectify these deficiencies by examining data from pre- and post-COVID-19 eras, facilitating enhanced comparisons of travel insurance behaviour over time. Broadening the focus to encompass insurance for additional transportation modalities, such as automobile rentals or cruises, offers a viable avenue for future investigation. This study emphasizes the revolutionary capacity of machine learning in insurance modelling, providing a comprehensive framework for enhancing risk assessment and operational efficiency. Addressing current literature deficiencies facilitates the development of more inclusive, data-driven methodologies within the travel insurance sector.
