Insurance premium prediction app:Journey from data to prediction

Lotus Labs
9 min readAug 13, 2024

--

Insurance companies need to accurately predict the cost of health insurance for individuals to set premiums appropriately. However, traditional methods of cost prediction often rely on broad actuarial tables and historical averages, which may not account for the nuanced differences among individuals. Data science addresses these issues by analyzing variable interactions to uncover patterns and trends. Using clustering algorithms, insurers can segment customers based on similar characteristics, enabling more targeted sales strategies and accurate pricing models. By leveraging predictive machine learning techniques, insurers can predict more accurately the insurance costs tailored to individual profiles, leading to more competitive pricing and better risk management.

The Data and insights

The dataset includes 986 samples and 11 attributes related to health and demographic factors. These attributes cover various health conditions, medical history, and basic physical measurements. The data provides a comprehensive overview that can be analyzed to understand patterns and correlations affecting health insurance premiums.

Distribution of variables

Analysis of the central tendencies of the dataset and visualized the data using histogram-q plots and kernel density estimates (KDE) were done. These visualizations helped in understanding the distribution of each attribute. Additionally, conducting normality hypothesis tests, such as the Shapiro-Wilk test, to determine if the data follows a normal distribution. The results indicated that the data is non-normal, necessitating non-parametric statistical methods for further analysis. This insight is crucial for applying appropriate data science techniques to gain accurate and meaningful insights. Decile checks and outlier analysis revealed a few outliers in weight, premium, and BMI, while all deciles seemed normal. The mean premium check of categorical variables showed that people with health issues are charged higher premiums as expected. There is a very low distribution of customers with transplants, chronic diseases, a family history of cancer, and a high number of surgeries, indicating these high-risk cases are less preferred and are charged higher premiums. The highest premiums are observed for customers with transplants, with no special cases noted otherwise.

Correlation analysis

The correlation analysis of numeric variables indicates that the premium is most strongly correlated with age, followed by BMI. Weight and height are naturally correlated with BMI. Scatter plots do not suggest any clear linear relationships. Hypothesis tests, such as Spearman correlation, confirm the Pearson correlation findings and visual insights, reinforcing the observed patterns.

Violin plots of categorical variables with respect to the premium reveal non-normal distributions. Non-parametric tests, such as the KS test, indicate that all variables influence the premium except for the number of surgeries, likely due to its rarity. Chi-square tests for categorical variable interactions show varying levels of influence between different categories. These analyses highlight the complex factors affecting premium pricing.

Clustering and Cluster analysis

K-means is a clustering algorithm that partitions data into K clusters by minimizing the variance within each cluster. K-means++ is an improved version that initializes centroids to be as far apart as possible, leading to better and more consistent clustering results.

Clusters are typically selected based on the elbow curve method, which plots the explained variance against the number of clusters. The optimal number of clusters is identified at the “elbow” point, where the rate of variance reduction slows down. For our insurance data, the elbow curve suggested four clusters. These clusters might represent distinct groups such as low-risk customers, moderate-risk customers, high-risk customers, and very high-risk customers, each with varying premium levels and health characteristics.

Plotting a pair plot segmented by clusters provides a visual representation of group attributes, helping to identify distinct patterns and relationships. Aggregating data at the group level offers insights into the average premium and other numeric attributes for each cluster, as well as the most common health conditions like diabetes and blood pressure issues. Further analysis, such as drilling down to the decile level and examining value counts, reveals more detailed characteristics of each group. This comprehensive approach helps in understanding the distribution and prevalence of attributes within each cluster, aiding in targeted strategies for insurance pricing and sales.

Modeling

Before diving into the modeling, we preprocess the data to ensure optimal model performance. This involves standardizing the numerical features (age, BMI, number of surgeries) to have a mean of 0 and a standard deviation of 1.

To find the best model for predicting insurance premiums, we evaluate several machine learning algorithms:

  1. Linear Regression
  2. Ridge Regression
  3. Decision Tree Regressor
  4. Random Forest Regressor
  5. Bagging Regressor
  6. AdaBoost Regressor
  7. Gradient Boosting Regressor

Linear Regression:

Linear Regression is a fundamental and widely used model that assumes a linear relationship between the input features and the target variable. We start with this model to establish a baseline for comparison.

Ridge Regression:

Ridge Regression is a linear model that includes a regularization term to prevent overfitting. It is particularly useful when dealing with multicollinearity.

Decision Tree Regressor:

The Decision Tree Regressor is a non-linear model that splits the data into subsets based on the value of input features, creating a tree-like structure. This model can capture complex interactions between features but is prone to overfitting.

Random Forest Regressor:

The Random Forest Regressor is an ensemble learning method that combines multiple decision trees to improve performance and reduce overfitting. By averaging the predictions of individual trees, it achieves better generalization and robustness.

Bagging Regressor:

The Bagging Regressor is an ensemble method that improves the stability and accuracy of machine learning algorithms by combining the predictions of multiple base estimators.

AdaBoost Regressor:

AdaBoost Regressor is an ensemble technique that combines the predictions of multiple weak learners to create a strong predictive model. It works by assigning weights to instances based on the error of the previous model.

Gradient Boosting Regressor:

Gradient Boosting Regressor is an ensemble learning technique that builds models sequentially, with each model attempting to correct the errors of the previous one. This method is highly effective for regression tasks.

Comparison of Models:

After training and evaluating the models, we compare their performance using the Mean Average Error (MAE),Mean Squared Error (MSE) and R-squared (R²) metrics. These metrics provide insight into the accuracy and explanatory power of the models.

Why Random Forest Regressor is the Best

The Random Forest Regressor outperforms the other models, as evidenced by its lower MSE and higher R² values. Here are the key reasons why the Random Forest Regressor provides the best results:

  1. Robustness: By averaging the predictions of multiple decision trees, the Random Forest Regressor reduces the risk of overfitting, which is a common issue with individual decision trees.
  2. Handling Non-linearity: Unlike Linear Regression and Ridge Regression, the Random Forest Regressor can

Deployment through streamlit

Preparing the Environment:

Before we start building the Streamlit app, ensure you have the required libraries installed. You can install Streamlit and other necessary libraries using pip, we use srteamlit,pandas,scikit-learn

Loading the Model and Scaler:

First, save the trained model(in our case we observed RF model to be the best based on MSE and R2) and the scaler used for preprocessing the data. This ensures that the same transformations are applied to the input data during prediction.

Building the Streamlit App:

Create a new Python file, app.py, and start by setting up the Streamlit application. Streamlit provides a simple and intuitive API for building interactive web applications.

Creating Input Fields:

Add input fields for the features required by the model. Use Streamlit’s widgets to create interactive input fields for user data.

Preprocessing the Input:

Calculate the BMI and standardize the numerical features before making predictions.

Making Predictions:

Use the preprocessed input data to make predictions using the loaded model. Display the prediction to the user.

Application link:https://lotuslabs-premiumprediction.streamlit.app/

Observations and suggestions

Data Structure:

  • The dataset contains 986 rows and 11 columns, with ‘premium’ as the target variable.

Variable Types:

  • Numeric variables: age, height, weight.
  • Ordinal variable: surgery number.
  • Binary categories: diabetes,bp etc.

BMI Index:

  • BMI, a known risk factor in insurance, can be derived from weight and height.

Missing Values:

  • The dataset has no null values.

Outliers:

  • A small number of outliers are present, indicating no mandatory treatment is needed.
  • The outliers seem reasonable upon inspection.

Data Distribution:

  • The numeric data does not follow a normal distribution, confirmed through visual inspection, Q-Q plots, and hypothesis testing.

Premium Analysis:

  • Visual and group-by analysis shows higher age and BMI are associated with higher mean premiums.
  • Presence of health conditions also leads to higher premiums.

Demographics:

  • 60% of the insured individuals are aged 22–50 and have a decent BMI level according to decile analysis, with premiums ranging from 15k-28k.
  • 30% fall into the obese category, aged 56–66, with premiums ranging from 29k-40k.
  • 10% are in the 18–22 age range and are underweight.

Risk Factors:

  • Older age (56+) and higher BMI indicate the highest premiums, implying the highest risk.
  • Underweight individuals are fewer across all age categories, indicating either high risk or generally low numbers.

Health Conditions:

  • Health conditions increase premiums due to higher risk.
  • Variables exhibit expected average premiums; individuals with health issues are charged more.

High-Risk Categories:

  • Few customers fall into sub-categories such as transplants, chronic diseases, cancer in family, and high number of surgeries (2–3), explaining high risk cases are not preferred, and those accepted are charged more.

Premium Correlation:

  • The highest premiums are observed for customers with transplants.
  • Premium price is highly correlated with age, followed by BMI.

Linear Relationships:

  • Hard to see any linear relationship of premium price with variables except age and BMI.

Clustering Analysis:

  • Highest risk groups:
  • Overweight, middle-aged, with transplants.
  • 50+ years old, overweight, with diabetes, BP, and one major surgery.
  • 36–45 years old, obese, with no health conditions.
  • 26–35 years old, normal weight, with no health conditions.

Model Insights:

  • Random Forest model suggests age and BMI are the top two indicators of premium.

Based on the observations, here are some suggestions for better business management, particularly in the areas of sales, pricing, and risk planning:

Sales Strategy

Targeted Marketing:

  • High-Risk Groups: Focus marketing efforts on high-risk groups such as older adults (56+), those with higher BMI, and individuals with specific health conditions like diabetes or history of surgeries. Tailor messages to emphasize the importance of comprehensive coverage.
  • Young Adults and Low-Risk Groups: Market to younger, healthier individuals by highlighting lower premiums and long-term benefits of starting insurance early.

Customized Plans:

  • Offer customized insurance plans based on the specific needs of different demographic segments. For instance, plans with more preventive care benefits for older adults or wellness programs for young adults.

Health Incentive Programs:

  • Introduce wellness and health incentive programs that reward policyholders for maintaining healthy lifestyles. This can help in reducing overall risk and encouraging more people to join.

Pricing Strategy

Dynamic Pricing:

  • Implement dynamic pricing models that adjust premiums based on real-time health data and lifestyle changes. Use wearable technology and regular health check-ups to monitor policyholders’ health and adjust premiums accordingly.

Risk-Based Pricing:

  • Continue using age and BMI as primary factors for pricing premiums. Additionally, include other risk indicators like health conditions and surgery history to create a more accurate risk-based pricing model.

Discounts and Rewards:

  • Offer discounts for policyholders who actively participate in health and wellness programs. Provide rewards for those who demonstrate consistent healthy behaviors over time.

Risk Planning

Predictive Analytics:

  • Utilize predictive analytics to identify potential high-risk individuals and offer them targeted interventions. This can help in mitigating risks before they result in costly claims.

Outlier Management:

  • Regularly review and analyze outliers to understand underlying causes. Implement strategies to manage these high-risk cases, such as offering specialized coverage or personalized health management plans.

Preventive Health Measures:

  • Invest in preventive health measures for policyholders, such as regular health screenings, fitness programs, and chronic disease management. This can help reduce the overall risk and cost of claims.

Customer Retention and Satisfaction

Customer Education:

  • Educate customers on the importance of maintaining a healthy lifestyle and its impact on insurance premiums. Provide resources and support to help them achieve their health goals.

Enhanced Customer Service:

  • Offer superior customer service with personalized support and timely assistance. Ensure that customers feel valued and supported throughout their insurance journey.

Feedback Mechanism:

  • Implement a robust feedback mechanism to gather insights from policyholders about their needs and preferences. Use this feedback to continuously improve products and services.

Innovation and Technology

Technology Integration:

  • Integrate advanced technologies like AI and machine learning to better assess risk and personalize insurance offerings. Use telematics and IoT devices for real-time monitoring and data collection.

Data-Driven Decisions:

  • Make data-driven decisions for product development, marketing strategies, and risk management. Leverage data analytics to gain deeper insights into customer behavior and risk patterns.

By implementing these strategies, the business can enhance its sales, pricing, and risk planning processes, leading to better management and profitability.

Dashboard for interactive insights

A dashboard with interactive chart objects and filters provides rapid insights, enabling the firm to make swift, data-driven decisions. The ability to click through different visualizations and apply filters ensures a comprehensive view of the data, while real-time data refreshing guarantees that the management team has access to the most current information. This dynamic approach facilitates timely actions and strategic adjustments, enhancing the overall efficiency and responsiveness of the organization.

Github link:https://github.com/sram77/insurance_premium_prediction

--

--

Lotus Labs
Lotus Labs

Written by Lotus Labs

Transform your business into an AI-driven enterprise. We specialize in Machine learning for Retail, Insurance, and Healthcare industries. www.lotuslabs.ai