Phd Thesis

**Subject**:
**Context of risk for young driver**:
In some insurance markets, the inclusion of real driving data (RDD) is now common for young insurance : in Italy, UK (25% of the market of young driver), US,This data has mainly two objectives:

- better selection
- better driving through a change of behavior (monitoring effect).

Even if there are several colloquial examples of change of behavior (including a reduction in time of the improvement of claims experience potentially linked to a tendency to _forget _that we are monitored), this is often hard in practice to split the part linked to selection and the part linked to change in behavior.

**Link between **_Real Driving Data_** and Driving behaviour : literature & R&D survey**:
Telematic insurance is now developed in several countries but literature is rather limited.

However, there is some literature through:

- the reading of the various offers proposed in various markets (cf Progressive Insurance communication to investors).
- theoretical and practical discussion about behavioral impact in car insurance vs autoselection

***_ Real Driving Data _**in the context of Driving Licence Education**:
Real Driving Data in the context of driving licence Education can be used in several supervised problems:

- To give practical feedback to the learner and improve its way of driving
- to define a score of Exam readiness
- to propose adapted insurance according to specific segments (for instance, UBI for those who have strong change of behavior / monitoring effect, )
- to predict Claims potential.

In addition, it can be retrieved in various environments:

- with the driving instructor (limited experience of driving of the driver)
- under parental supervision (“conduite accompagnée”)
- alone, in the context of UBI insurance.

**Dimension reduction**:
We need first to clean data (for instance, when we are in a tunnel with no information about driving behavior) and then to reduce dimension, through several dimension reduction techniques (PCA and Auto-Encoder), in order to simplify features _X_i_ into some “intense” codes _Code_i_ that will then be used as inputs for the various supervised models mentioned above.
- The issue will be to interpret the obtained axes or codes of the dimension reduction techniques, ie PCA or AutoEncoder.
- PCA, Principal Components Analysis are linear combination of features preserving the maximum level of information and orthogonal one by one. ACP “axes” are generally well understood.
- AutoEncoders are a class of Neural network with two phases, one of encoding and one of decoding codes. AutoEncoder Codes are more complex to interpret but potentially able to capture non linearities (potentially strong in insurance).
- He will also look at recent development in AutoEncoders, (_masked autoencoder ), _in order to solve the difficulty of heterogeneous tracks (some clients having only a track on theoretical driving license, other practical). He will propose alternatives to current existing models.

**Work on unsupervised / Supervised problems**:
**“One self vs several selves”**

**Can we recognized the pattern of a driver from his first driving lesson to his UBI experience**:
We have seen that there were several different environment of driving for the young driver : with an instructor, a parent or alone, under the “unseen” supervision of an insurer. Can we recognize common “patterns” for a same driver (“one self”) or is the _driving style _very different in these three context ? In addition, can we spot the input of an instructor that would have a positive (hopefully) impact on his students ?

A kaggle done on AXA telematics 6 years ago has already tackled some of these questions (see Driver Telematics Analysis ) with probably too much weights given to the roads used by the driver for a context of “young drivers”.

**Give practical feedback to the learner and improve its way of driving**:
Instructors are required to describe each hour of driving lesson. Through a comparison of NPL to analyse their comments and the scores/codes obtained on the driving of the student, can we predict comments / feedback on the driving?

**Define a score of Exam readiness**:
Can we predict a level of driving licence exam according to the level of scores/codes obtained?

**Propose adapted insurance according to specific segments**:
According to the behavior of the student measured through the obtained scores/codes, propose him/her adapted insurance product, especially UBI for those who have strong change of behavior / monitoring effect.

**Predict Claims potential**:
Reinforce current claims potential score (called score Edu) by relevant information retrieved from practical experience with instructor or conduite accompagnée.

**Impact of feedback on Driving behavior**:
We will send regularly feedbacks to drivers and measure through _Real Driving Data_ if we can see some impacts of these feedbacks

**Split between moral hazard