Identifying Driving Risk Factors to Support Usage-Based Insurance using Smartphone Personalized Driving Data

Transcription

1 Identifying Driving Risk Factors to Support Usage-Based Insurance using Smartphone Personalized Driving Data Yi-Chang Chiu 1, Yu-Luen Ma 2, Xianbiao Hu 3* 1 Associate Professor, Department of Civil Engineering and Engineering Mechanics, University of Arizona, Tucson, AZ, USA, chiu@ .arizona.edu 2 Professor, Katie School of Insurance and Financial Services, Illinois State University, Normal, II, USA, yma@ilstu.edu 3* Director of R&D, Metropia, Inc., Tucson, AZ, USA. xb.hu@metropia.com Words: 5908 Figures and Tables: 16

2 ABSTRACT Historically, automobile insurance premiums in the U.S. were largely based on various static factors, such as socio-demographic and geographical information, without taking individualized risk factors into consideration. Advances in the development of modern GPS devices have led some insurers to start looking into variations of Usage Based Insurance (UBI) or Pay-As-You- Drive-And-You-Save (PAYDAYS) research. Progressive Insurance Company, for example, monitors the driving behavior of their policyholders by installing a GPS device in their cars for certain period of time, offering adjusted premiums based on the GPS reading, although shortcomings with this approach have been observed. A comprehensive research framework and approach to analyzing vehicle use and/or driver behavior data to advance knowledge about driving exposure factors that are closely linked to crash risk is presented in this paper. The detection criteria for crash/near crash event will first be identified and carefully calibrated. Such criteria will be applied to the Metropia Mobile user database to automatically detect all crash/near crash events associated with Metropia Mobile users. The driving behavior evaluation model will then be built to quantitatively evaluate each user s driving behavior based on the available data including GPS trajectory, traffic speed, roadway geometry and crash event information. In addition, through exploring partnership with insurance companies, loss severity information for the detected crash event will be retrieved. The results of our study will further existing knowledge about driving exposure factors that are closely linked to crash risk. The manner in which these factors affect loss frequency and claim severity will also help insurers re-structure their existing pricing models to allow for variation in premiums based on these characteristics, and provide the actuarial foundation for advanced forms of PAYDAYS insurance pricing. Keywords: Usage Based Insurance (UBI), Pay-As-You-Drive-And-You-Save (PAYDAYS), Driving Risk Factors, Personalized Driving Data, Smartphone data collection, insurance pricing

3 1. INTRODUCTION Historically, automobile insurance premiums in the U.S. were largely based on various static factors, such as socio-demographic and geographical information, without taking individualized risk factors into consideration. Today, most people s driving patterns are not incorporated into actuarial pricing by insurers simply because such data has not been available, even though driving habits, traffic congestion and vehicle dynamics can all be potential hazards that lead to deadly accidents. The emergence and subsequent rapid advances with new information and communication technologies (ICT) such as GPS devices, cellphone, Bluetooth, etc., now offer the capability of collecting high-fidelity and high-resolution travel data in a cost-effective manner. These technologies, especially smartphones and mobile application platforms, also permit continuous data collection related to driver behavior, so long as the smartphone device remains in operation. Given the ever growing cellular phone market, 1 it has now become much easier to track and understand traveler activity, and travel patterns. Researches of applying the latest ICT to collect data and evaluate driving behavior started to show up in the last decade. A recent effort in the individual driving behavior is the 100-car naturalistic driving study [1]. Drivers are monitored and recognized as unsafe, moderate safe and safe according to frequencies of crashes/near-crashes. The results indicate that hard braking, inattention, and tailgating are the top three at-risk behaviors among drivers. Unsafe drivers are more likely to engage in the at-risk behaviors and decelerate/swerve greater than the safe drivers. The results also imply that improper braking and inappropriate speeds are positively related to crash/near-crashes. Different traffic and weather conditions are also studied separately for the driving behavior and crash risk. The unsafe drivers drive more aggressively regardless of traffic conditions. Another example of using new information and communication technologies in the driver behavior evaluation can be found in [2], who proposed a framework for profiling drivers by atrisk behavior using driving pattern, spatial and temporal characteristics and driver characteristics. Second-by-second GPS data observations are collected from 106 drivers in Sydney over several weeks. Behavioral measures are summarized as maximum, average, minimum and standard 1 A whopping 90% of adults in the United States own cellular phones, of which 65% own a smartphone.

4 deviation of speed, acceleration and deceleration, distance at 75% of speed limit or over speed, number of sharp celeration ( 4 m/s 2 ), etc. Also, in the last decade the Data Acquisition System (DAS) or In-Vehicle Data Recorders (IVDR) have been introduced to collect detailed driving behavior data, but the usage was limited due to the high hardware cost, and the research sample size are usually insufficient [3, 4]. More examples of applying ICT to evaluate driving behavior can be found in [5-8]. Those researches, however, are mostly limited in the model validations, and the amount of users in the experiments are usually low. Advances in the development of modern GPS devices have led some insurers to start looking into variations of Usage Based Insurance (UBI) or Pay-As-You-Drive-And-You-Save (PAYDAYS) research. Progressive Insurance Company, for example, monitors the driving behavior of their policyholders by installing a GPS device in their cars for certain period of time, offering adjusted premiums based on the GPS reading. There are two shortcomings with this approach. First, due to the high hardware cost, this short-term tracking does not permit the insurer to continuously monitor and track policyholders driver behavior, which provides little incentive for drivers to maintain safe driving habits once the trial period ends. Additionally, the data collected by Progressive s and other insurers systems depict only the speed and acceleration/deceleration of the vehicle. Other important risk factors such as location information (e.g. an intersection of higher risk than a mid-link) and traffic flow are not traceable by the GPS devices. In this 24 month research project titled Pay-As-You-Drive-And-You-Save (PAYDAYS) Insurance Actuarial Study 2, we propose to answer the following major research question: How can we collect and use various types of data, including vehicle trajectories, associated link traffic volume and speed data, time of day information, and desired claim data, to more accurately identify driving exposure factors that are closely linked to crash risk, and quantify the relationship between crash, hazard, driving behavior and congestion levels, as well as accident claims? In order to answer this question, a statistical analysis model pertaining to the technical narratives on our collected data as well as research methodology plan will be developed. The primary emphases include actuarial analysis based on vehicle trajectories, dynamic traffic data 2 This research is supported by Federal Highway Administration Broad Agency Announcement Pay-As-You-Drive-And-You-Save (PAYDAYS) Insurance Actuarial Study project. Contract award number DTFH61-13-C

5 and hazard proxy variables with the goal of establishing the linkage between driving risks and resulting claims. In this paper, the overall research framework to be used in this research project will be presented, which mainly includes data collection, general research work flow and actuarial analysis research approach. Current progress -- including crash detection threshold calibration and preliminary trajectory analysis -- will also be presented. The rest of this paper will be organized as follows: Section 2 presents the methodology being developed for this research project, including the overall research work flow, data collection which covers the list of data being collected in this research, and data collection mechanism, and the actuarial analysis research approach to identify the driving risk factors. Section 3 reviews the research progress to date, including the hard brake experiment performed earlier this year to better calibrate the crash detection threshold, and the preliminary analysis on the GPS trajectory data. Section 4 concludes this research and discusses future research directions. 2. RESEARCH METHODOLOGY 2.1. OVERALL RESEARCH WORK FLOW In this section, a comprehensive research framework and approach to analyzing vehicle use and/or driver behavior data to advance knowledge about driving exposure factors that are closely linked to crash risk is presented. We envision the research to span over a 24-month period, with the main modules in this research to include the following: Data collection: specifying what kind of raw data we are collecting, and how we collect these raw data Data processing: describing how different kinds of raw data will be processed in this research to extract useful information for further actuarial analysis purposes Actuarial analysis: explaining how the data will be used in this research for actuarial analysis purposes, what model will be built to analyze the data to identify the driving risk factors, etc. After the GPS trajectory is sent back to the cloud server, those data will be stored and processed for PAYDAYS research purposes. The detection criteria for crash/near crash event will first be identified and carefully calibrated. Such criteria will be applied to the Metropia Mobile user database to automatically detect all crash/near crash events associated with users. Such crash

6 detection module developed will be used to detect the crash event in a real time fashion, which means if abnormal situations are observed from the GPS trajectory sent back from user s smartphone and certain criteria is met, the system will consider the chance of this user running into crash to be very high. Certain subsequent modules will be triggered to verify the crash event, query the person injury and property loss information from users, and retrieve the claim cost from partner insurance company, if available. The driving behavior evaluation model will then be built to quantitatively evaluate each user s driving behavior based on the available data, including GPS trajectory, traffic condition, roadway geometry and crash event information. In addition, through exploring partnership with insurance companies, loss severity information for the detected crash event will be retrieved. The results of our study will further existing knowledge about driving exposure factors that are closely linked to crash risk. The manner in which these factors affect loss frequency and claim severity will also help insurers restructure their existing pricing models to allow for variation in premiums based on these characteristics, and provide the actuarial foundation for advanced forms of PAYDAYS insurance pricing. All of those data collected, together with the insurance claim data, will be used for further statistical and actuarial analysis to identify the risk factors associated with the driving behavior, and quantify their relationship with the auto accidents. The overall research work flow for this project can be found in Figure 1 below.

7 Figure 1. PAYDAYS overall research work flow The main procedure is as follows: Step 0: Determine the criteria for auto crash detection. This is an important module since accurately defining the logic and calibrating the threshold is critical to efficiently identify the potential crash event. A near crash detection experiment was performed earlier this year to achieve this goal Step 1: Collect GPS trajectory from user s smartphone, and such data will be sent back to the cloud server in a real-time fashion.

8 Step 2: Through the real time trajectory processing engine, the system will identify the potential crash event based on the comparison between user information and the pre-defined crash detection rules. Data cleaning will happen in this step prior to the crash event detection, a systematic method will be used to identify and remove invalid data points. Step 3: If potential crash event has been identified, an with pre-defined questionnaire will be automatically sent to the user, the purpose of this is to confirm the accident event. Step 4: After receiving the confirmation of crash event, if the user is also a customer of our partner insurance company, the claim cost will be retrieved when it becomes available for further research analysis purpose. Otherwise, a second with questionnaire will be sent to the user querying the estimated claim cost based on the person injury and property loss information. Step 5: Actuarial analysis with statistics modeling to reveal the driving risk factors and establish the linkage between driving risks and resulting claims. The system will continue steps 1 through 4 until it collects sufficient data samples for statistics and actuarial analysis purpose. Section 2.4 below will further discuss the actuarial analysis research approach (Step 5 above) for this PAYDAYS project DATA COLLECTION The dynamic new smartphone-based data collection mechanism opens up the possibilities of allowing long-term continual data collection that identifies true high risk behaviors. As one of the critical steps of this project, data collection provides the foundations for the different modules in the system research framework, including potential crash event identification, driving risk factors identification, actuarial analysis to quantify the relationships hazards and accident claims and so on. In this research, in order to answer the question of what are the factors that are attributable to auto accidents, and how can we use various types of data to more accurately quantify the relationship between driving, traffic and auto accidents, we propose to collect various risk factor data, as well as indicating whether an accident resulted from a particular journey. The primary data are categorized into: 1) Vehicle trajectory data 2) Roadway geometry 3) Time-varying traffic dynamic

9 4) Claim data In this study we use Metropia Mobile, a recently developed smartphone-based app designed to improve mobility management, to keep track of people s driving behavior, surrounding traffic, and other trajectory data to develop models that investigate the relationships between various driving hazards and insurance claims. Metropia Mobile is a recently available mobile traffic app that uses prediction and coordinating technology combined with user rewards to incentivize drivers to cooperate, balance traffic load on the network, and reduce traffic congestion. When a user starts a trip, the internal GPS module is activated and starts to record the second-bysecond latitude/longitude data location and instantaneous moving speed. These data allow detailed position, velocity, acceleration and deceleration details to be stored and analyzed. One unique feature of the Metropia data is that its backend server also estimates traffic speed and volume for each link that the vehicle traverses. Such data -- combining both user trajectory and link speed/volume information -- has rarely been seen in prior research, permitting a unique opportunity to link critical traffic congestion factors that lead to certain driving behaviors and crash potential. For example, if a driver exhibits stop-and-go or abrupt accelerate/decelerate behavior, it is usually difficult to tell if this is simply due to the driver s behavior, or because of heavy traffic conditions. When both driving and traffic data are linked together, one can discern hazards caused by driving behavior from those caused by congestion levels VEHICLE TRAJECTORY DATA Detailed trajectory data and driving behavior data for each trip are collected during the trip validation process. When a user starts a trip with the app, the internal GPS module built in the smartphone is activated and starts to record the second-by-second latitude/longitude data location and instantaneous moving speed. These data, including detailed position such as latitude, longitude and altitude, heading, velocity, acceleration and deceleration will be collected at fine time interval and sent back to the cloud server, where they will be stored and used for further analysis both in the online or offline fashion. The main attributes of the trajectory data can be found below: Table 1. Vehicle trajectory data Attribute UserID Description The unique ID of the user

10 TripID Latitude Longitude Altitude Heading Speed Acceleration Timestamp LinkID The unique ID of the particular trip The latitude of this particular GPS point The longitude of this particular GPS point The altitude of this particular GPS point The heading of vehicle at this moment The instantaneous travel speed of the vehicle The instantaneous acceleration value of the vehicle The timestamp of this particular GPS point The ID of the road segment that user is currently driving on ROADWAY GEOMETRY The geographic network stored on cloud server with the main purpose of providing routing and navigation functionalities to the users. A typical traffic network usually includes a set of links and nodes, i.e. G = (N, A) where N is the set of nodes {1, 2,, n} and A the set of link. The key attributes relevant with this research is the latitude and longitude of the nodes, the link types, i.e. whether a road segment is located on freeway, arterial, local streets, the speed limit of the road segment, the number of lanes of the road segment and so on. Table 2 describes the list of data attributes for the geographic network dataset. Table 2. Roadway geometry data Attribute LinkID Speed Limit Facility Type Number of lanes Link length Start nodeid Start latitude Description The ID of the road segment that user is currently driving on The speed limit of the current road segment The type of links, such as freeway, arterial, local street, ramp Number of lanes of the road segment that user is currently driving on The length of the road segment The ID of the start node for this particular road segment The latitude of the start node

11 Start Longitude End nodeid End Latitude End Longitude The longitude of the start node The ID of the end node for this particular road segment The latitude of the end node The longitude of the end node TIME-VARYING TRAFFIC DYNAMIC On the cloud server, the time-varying traffic dynamic, i.e. the traffic dataset including timedependent traffic speed, traffic flow, etc., with the time interval of five minute at each traversed link is hosted and keeps updating every 5 minutes. The main purpose of these data is to feed into the routing engine and guide users to the desired destination while avoiding the traffic congestion. The traffic dataset of interest mainly includes time-dependent link speed and volume information. List of attributes of the traffic data can be found in Table 3 below. Table 3. Traffic dataset Attribute LinkID Date Timestamp Description The unique ID of a road segment The date of the traffic data The timestamp of this particular traffic data Link average speed The instantaneous average speed of this road segment Link volume The instantaneous traffic volume of the current road segment CLAIM DATA In this research, we consider two possible scenarios with or without insurance company s claim data. In the scenario in which at least one insurance company is willing to partner with the research team in this project and is willing to share claim data as they come in, when an incident occurs and a claim is filed, if the policyholder is a Metropia traffic app user, this claim will be cross-referenced with the corresponding reservation and trajectory record. Claim data for the participant who is not insured with a partnering insurance company and is confirmed to have been involved in an accident will be collected through questionnaire, which is automatically sent out to the driver sometime after the accident happens.

12 2.3. DATA PROCESSING DATA CLEANING Data cleaning process will be applied prior to the data being used in the research, which is very important to ensure the quality of system input and accurate analysis result. A systematic method has been developed to identify and remove invalid data points. Some typical invalid data scenarios include but are not limited to: GPS being inaccurate under certain scenarios, such as when it s around tall buildings or indoors. In a dense city packed with skyscrapers like Los Angeles and New York, finding an unobstructed patch of sky to divine your location can be a serious challenge. GPS being inaccurate due to network issues. There s a component of assisted GPS to the system, which requires a cell signal and numerous nearby towers with which the phone can triangulate its location. If the cellphone network connection is weak, for example under 3G or 2G network, the location result may not be as accurate as with strong signal. GPS point jumps when the speed of vehicle moving is low. It is sometimes observed that when the vehicle driving speed is low, there s a chance that the heading, location of the GPS will not be very accurate Issues with the data transmission. During the data transmission from cell phone to the cloud server, due to some known or unknown issue, the data may be lost or sent over multiple times. Network data quality issue, such as missing speed limit for the road segment DATA JOINS AND SAMPLE DATA For the actuarial analysis purpose, if we perform data joins between the multi-source data mentioned in Section 2.2, in the end for actuarial analysis purpose, for each GPS point, we will have the following attributes: Table 4. List of data attributes after data joins Attribute UserID TripID Latitude Longitude Altitude Heading Speed Acceleration Timestamp LinkID Description The unique ID of the user The unique ID of the particular trip The latitude of this particular GPS point The longitude of this particular GPS point The altitude of this particular GPS point The heading of vehicle at this moment The instantaneous travel speed of the vehicle The instantaneous acceleration value of the vehicle The timestamp of this particular GPS point The ID of the road segment that user is currently driving on

13 Distance to Downstream Intersection Next Movement Facility Type Number of lanes Link average speed Link volume Speed limit Distance to the next intersection that drivers need to turn The next turning movement at the intersection, such as turn right, turn left, U turn The type of links, such as freeway, arterial, local street, ramp Number of lanes of the road segment that user is currently driving on The instantaneous average speed of the current road segment The instantaneous traffic volume of the current road segment The speed limit of the current road segment Most attributes in the above table are pretty straightforward and could be obtained by simple data joins. One thing to mention is some special data processing is needed for the Next Movement and Distance to downstream intersection, which requires the algorithm to scan through the user s whole trip trajectory, extract the decision points where drivers need to make the turns, and compute these two values for each GPS point ACTUARIAL ANALYSIS RESEARCH APPROACH The general research work flow in Figure 1 described the systematic framework of how the GPS trajectory data will be collected and applied to identify crash event, how the crash event data and claim cost data will both be obtained and correlated to facilitate the actuarial analysis. In this section, the overview of PAYDAYS auto insurance actuarial analysis research approach will be presented. The emphasis of this actuarial analysis will be placed on several aspects including risk measurement design, statistical modeling and data analysis ACTUARIAL ANALYSIS RESEARCH APPROACH FOCUSES To achieve the objectives of this insurance actuarial study, the actuarial analysis research approach proposed in this project will involve three main steps. We will first define and explain how we identify the driving risk factors and develop a baseline analysis that describes the characteristics of our sample. The risk factors that we believe will affect driving behavior and may potentially lead to accidents will firstly be defined. The characteristics of those risk factors, such as distribution of driver accelerations, decelerations will then be analyzed with the sample data from the database. Secondly, a correlation analysis of risk measurements that are likely to be linked with accidents will be performed, the purpose of this analysis is to reveal the relationship between different risk

14 factors, validate various assumptions on the risk measurements that may or may not be contributing to the accidents, and provide theoretical foundations to the full statistical analysis later. For the full statistical analysis which is the third step, generalized linear models (GLMs) will be utilized for predicting both loss frequencies and loss severities and provide implications for insurance rate making. GLM is a flexible generalization of ordinary linear regression that allows for response variables that have error distribution models other than a normal distribution. The GLM generalizes linear regression by allowing the linear model to be related to the response variable via a link function and by allowing the magnitude of the variance of each measurement to be a function of its predicted value. By removing the restrictive assumptions usually seen in the linear models, GLM eliminates the chances of the models being built violate real world practice, and it s extensively used in the statistics KEY VARIABLES DESCRIPTION We envision the key variables of measuring driving hazards to include the variables that can be observed from the user GPS trajectory or can be derived from joining trajectory data with the geometric network as well as dynamic traffic data. For example, since driving speed and acceleration/deceleration can be obtained from the second-by-second detailed GPS trajectory, by associating the GPS point data with the network road segment speed limit information, it can be determined whether the driver drove at a speed higher than the speed limit. In additional, it can also be inferred how many times the driver has made harsh brake during the trip by looking at the deceleration values. For each trip, the following key variables can be defined and calculated: Table 5. Key trip-based variables Attribute Speeding Relative Speed Braking Description if driver is driving at a speed higher than speed limit of the road segment by a certain threshold if the driver s driving speed deviates from other drivers on the same road at the same time for more than a certain threshold if deceleration is lower than given threshold

15 Time in traffic Acceleration Peak Time Late time Left Turn Right Turn Mileage Travel time Accident Crash Loss the time when traveling in traffic and driving slowly if acceleration is higher than given threshold the time when driving during peak hours time when driving at midnight or early morning the number of times driver makes a left turn during the trip the number of times driver makes a right turn during the trip the total distance of the trip the total length of time traveled during this trip a Boolean variable, the value is true if this trip is associated with an accident, otherwise false. the claim amount, if this trip is associated with a crash event 3. CURRENT PROGRESS In this section, current progress of this PAYDAYS research project will be presented, which mainly includes the hard brake experiment to calibrate the crash detection module, and a preliminary analysis on the GPS trajectory data HARD BRAKE EXPERIMENT The capability to automatically detect an accident is one of the key requirements to this PAYDAYS research. As shown in the overall research work flow in Figure 1, the activities to verify accidents either via questionnaires or through an insurer partner will be triggered only after a crash event has been detected. The GPS modules in the smartphone is constantly updated in fine time intervals and send the data to the trajectory database on the cloud server, and these acceleration/deceleration values can be computed in a real-time fashion. Combined with other supplemental information, such as speed afterwards and location changes, a harsh deceleration can be detected and flagged as a potential crash/near crash event if it exceeds a certain threshold. In this design, the detection is primarily based on scanning the speed-related data in the recorded vehicle trajectories and identifying the hard breaking/large deceleration occurring in most of the crash/near crash

16 events. 3 In the past, we ve relied on literature to define this type of threshold value; however, such a value may not be applicable for our purpose because of the differing measurement instruments used in literature. The experiment conducted earlier this year is specifically for the PAYDAYS research study. We collected a set of field data from a controlled experiment that involved actual hard breaking behavior by different drivers using various types of vehicles. These data will help determine appropriate deceleration thresholds for the detection of potential crash/near crash events by the system EXPERIMENT SETUP Time: Weekend (Saturday) morning Location: Shopping mall parking lot (EI Con Mall, Tucson, Arizona) Conditions: Clear, dry and warm Total number of vehicles: 5 (4 sedans + 1 SUV) Total number of phones: 7 (4 android phones + 3 iphones) Equipment: 1. A light-weight baseball training net 2. Paint bucket 3. Tripod and camera/camcorder Procedure: 1. Place the baseball net at a fixed location on the test track. 2. Measure feet upstream and place the bucket on the right side of the track. 3. Ask the driver to reach a speed higher than 30 mph. 4. The driver cannot break until the vehicle passes the bucket. 5. The driver needs to break hard enough to avoid hitting the net. A video showing an example of the procedure for this test can be viewed at 3 It is possible for a crash to occur without a hard break if the driver is not cognitively capable of processing dangerous driving conditions or perform a timely reaction (e.g. driving under influence), or if the weather condition is icy and a crash occurs as a result of vehicle skidding. These types of cases won t be detected by the proposed approach. Such events, however, account for a relatively small percentage of crash incidents when compared to those with normal driving reaction.

17 Figure 2. Experiment location overview

18 Figure 3. Hard Break Field Testing photos from a test conducted in Tucson, Ariz COLLECTED DATA A total of 78 samples were collected in the experiment. Below is a brief visualization of the testing procedure and speed fluctuations. The maps of Figure 4 and Figure 5 show vehicle trajectory as collected by phone, where it can be seen that the vehicle drove back and forth during the experiment period. The graphs of the two figures show the vehicle was first speeding up, and then the speed quickly goes back to 0 as the testers brake very hard (X axis time, Y axis - speed) Figure 4. Data collected by Phone 1 (upper - GPS trajectory, lower speed fluctuations)

19 Figure 5. Data collected by Phone 2 (upper - GPS trajectory, lower speed fluctuations) PRELIMINARY RESULTS The deceleration values of the hard brakes during the experiment were plotted in the historical diagram below. As can be seen, the deceleration data exhibits a bell shaped distribution with just a few observations of larger deceleration values. Figure 6. Hard Brake Deceleration Histogram A more detailed statistical results analysis can be found in Table 6 below. The mean value of deceleration/ft. per square second is about -21.1, the median is -19.7, and the standard deviation is At confidence level of 95%, Z value is 1.681, and confidence interval is [- 22.5, -19.7]. Table 6. Detailed statistical results of hard break deceleration Deceleration/ft. per square sec. Mean

20 Standard Error Median Mode Standard Deviation Sample Variance Kurtosis Skewness Range Minimum Maximum Sum Count 76 Confidence Level (95.0%) PRELIMINARY TRAJECTORY ANALYSIS In this section, sample user trajectory data is extracted from the user database and examined to better understand user driving behavior, including user s instantaneous speed fluctuation, the frequency of hard brake and/or jackrabbit starts, the acceleration and deceleration values, etc. We present some brief examples below. Figure 7 through Figure 9 visualize the GPS trajectory of three users, with each including the time-dependent speed fluctuation and celeration (acceleration is the number is positive, and deceleration if negative) changes. A preliminary analysis reveals that User 3 braked not only the most frequently, but also the hardest, so it s reasonable to assume he/she was driving more aggressively than the other two users. 41 Some of the deceleration values were observed to be as high as -40 ft/s 2, which according to our preliminary analysis was due to the Anti-lock Braking System (ABS).

21 Figure 7. User 1 morning trip analysis (speed fluctuation and celeration changes) Figure 8. User 2 late night trip analysis (speed fluctuation and celeration changes)

22 Figure 9. User 3 evening trip analysis (speed fluctuation and celeration changes) The GPS trajectory is also correlated with the road network geometry information to better understand users driving behavior and how deceleration behavior/action relates to roadway information. In Figure 10, the red signifies the driver stopping or driving at relatively low speeds, and green indicates the user driving at higher speeds. The graphic allows us to easily determine that the traffic was light and the red dots indicate slow movement mostly at intersections.

23 Figure 10. Driving speed fluctuations along the travel trajectory The case studies in this section demonstrated the feasibility of applying the data collected by smartphone to quantitatively evaluate the driving behavior. All the trajectory data collected will be further analyzed in this research by correlating with other data source including roadway geometry, time-varying traffic dynamics and claim data. Statistical models will be built to performed to reveal the relationship between different risk factors, validate various assumptions on the risk measurements, and in the end, to predict both loss frequencies and loss severities and provide implications for insurance rate making. 4. CONCLUSIONS AND FUTURE RESEARCH Advances in the development of modern GPS devices and smartphone technologies have made the application of Usage Based Insurance (UBI) or Pay-As-You-Drive-And-You-Save (PAYDAYS) possible, which has been demonstrating significant advantages over the traditional insurance pricing model. A comprehensive research framework and actuarial analysis research

24 approach to analyzing vehicle use and/or driver behavior data to advance knowledge about driving exposure factors that are closely linked to crash risk is presented in this paper. The next steps in this 24 month research project include defining and computing risk measurements, performing the correlation analysis between different risk factors, and building actuarial analysis model to quantify the relationships hazards and accident claims. We believe that as time goes, more personalized driving data will become available and used in the statistical model for analysis, the research team can further move this research forward, so that in the end, the results of our study will further existing knowledge about driving exposure factors that are closely linked to crash risk. The manner in which these factors affect loss frequency and claim severity will also help insurers re-structure their existing pricing models to allow for variation in premiums based on these characteristics, and provide the actuarial foundation for advanced forms of PAYDAYS insurance pricing. 5. REFERENCES 1 Klauer, S.G., Dingus, T.A., Neale, V.L., Sudweeks, J.D. and Ramsey, D.J Comparing Real-World Behaviors of Drivers With High versus Low Rates of Crashes and Near- Crashes. Virginia Tech Transportation Institute. 2 Ellison, A.B., Greaves, S.P. and Daniels, R Profiling drivers' risky behaviour towards all road users. A safe system: expanding the reach: Australasian College of Road Safety national conferencesydney. 3 Toledo, T., Musicant, O. and Lotan, T In-vehicle data recorders for monitoring and feedback on drivers. Transportation Research Part C: Emerging Technologies, 16(3), Toledo, T. and Lotan, T In-Vehicle Data Recorder for Evaluation of Driving Behavior and Safety. Transportation Research Record: Journal of the Transportation Research Board, 1953,

25 5 Af Wåhlberg, A.E Driver celeration behaviour and accidents an analysis. Theoretical Issues in Ergonomics Science, 9(5), Young, M.S., Birrell, S.A. and Stanton, N.A Safe driving in a green world: A review of driver performance benchmarks and technologies to support smart driving. Applied Ergonomics, 42(4), Ericsson, E Driving pattern in urban areas-descriptive analysis and initial prediction model. Lund Institute of Technology, Department of Technology and Society, Traffic Planning, Lund. 8 Ericsson, E Independent driving pattern factors and their influence on fuel-use and exhaust emission factors. Transportation Research Part D: Transport and Environment, 6(5),