Abstract
This chapter presents a thorough descriptive analysis of automobile fatal accident and insurance claims data. Major components of the artificial neural network (ANN) are discussed, and parameters are investigated and carefully selected to ensure an efficient model construction. A prediction model is constructed through ANN as well as generalized linear model (GLM) for model comparison purposes. The authors conclude that ANN performs better than GLM in predicting data for automobile fatalities data but does not outperform for the insurance claims data because automobile fatalities data has a more complex data structure than the insurance claims data.
TopData Description
There were two groups of data used in this study. The first one is were extracted from the FARS (Fatal Accident Reporting System) collected and organized by the United States Department of Transportation. 2016 to 2019, 4 years of complete data were extracted. One is fatal accident data by person, which includes all relevant information by person who were involved with an accident that had at least one death during the year. We will call this data source 1 in the following content.
The second data used is based on one-year automobile insurance claims data from either 2004 or 2005, downloaded from the following website: http://www.afas.mq.edu.au/research/books/glms_for_insurance_data. The data was used as sample for the book Generalized Linear Models for Insurance Data (Jong & Heller, 2008). In this chapter, the data is analyzed to provide a benchmark comparison between three main statistical models. We will call this data source 2 in the following content.
Key Terms in this Chapter
Automobile Insurance Claim: One of the datasets analyzed in this chapter is that captures an automobile insurance company’s claims information from 2004 to 2005.
ReLU: ReLU is for rectified linear unit. It is a type of activation function that converts input value to a range of (0, 1) exclusively. It is a commonly used activation function for probability model output layer.
Hidden Layers: Hidden layers are essential components of artificial neural network which significantly enlarges the model’s mathematical complexity and improve its capability. It serves as ANN model parameter compared with other traditional statistical modeling.
Batch Size: The batch size is the number of samples selected from training dataset. Possible batch size can be any integer ranging from 1 to the total data number in training set.
Artificial Neural Network (ANN): A machine learning technique that can be used to learn historical patterns and make future predictions. It works as a simulation of our brain nervous system, each node in ANN represents a neuron, and one or more nodes in each layer.
Logistic Sigmoid: Logistic sigmoid is often called sigmoid. It is a type of activation function that converts input value to a non-negative value. It is a commonly used activation function for hidden layers.
Predictive Modeling: A process that utilize historical and current data to capture patterns, then create models to make future predictions.
Automobile Fatal Accident: One of the datasets analyzed in this chapter is from FARS (Fatality Analysis Reporting System), which is recorded for any accident happened in the United States that has at least one person dead accidentally due to automobile accident.
Generalized Linear Model: A modeling process that build relationships between variables. It relaxes the restrictive normal assumptions for the traditional linear regression model, which allows GLM capable of analyzing non-normal data.
Activation Functions: Activation function assists neural network to learn from training data. It transforms values from the previous layer into the next layer, with adjusted weights on parameters and adjusted bias values.
Linear Activation: Linear activation function has an output range from negative infinity to positivity infinity with a constant coefficient to adjust the weight.
Neurons: Neurons are key components of ANN, which is responsible processing information by carrying the calculated value forward from input layer to output layer. Too few neurons in each layer could result in large errors or under-fitting, and too many neurons could result in overfitting.
Step Function: Step function is a type of activation function that converts input value to a binary 0 or 1 output. It has a high calculation efficiency.
Model Validation: A process of model construction that used to validate the model efficiency. Commonly used process for validation include 50%(Training)/50%(Testing), 75%(Training)/25%(Testing), 80%(Training)/20%(Testing), Even-Year(Training)/Odd-Year(Testing), K-Fold Cross Validation, etc.