METHOD FOR DETERMINING A GROUP OF DRIVERS TO ESTABLISH A PEER-TO-PEER INSURANCE SCHEME WITH HOMOGENEOUS RISK

Abstract

METHOD FOR DETERMINING A GROUP OF DRIVERS TO CONSTITUTE A PEER-TO-PEER INSURANCE SCHEME WITH HOMOGENEOUS RISK The present invention relates to a method for determining a group of drivers participating in a peer-to-peer insurance program, implemented by computer, comprising the steps: a) determining a knowledge graph (2) of a plurality of drivers (A, B, C, D, E); b) determining a community of drivers from among the plurality of drivers, presenting a homogeneous accident risk, based on accident risk scores; c) determining a first driver from among the community of drivers and inviting said first driver to participate in the peer-to-peer insurance program; d) determining a subsequent driver from among the community of drivers by following a shortest Hamiltonian path in the knowledge graph;e) iteration of step d) until the sum of the insured amounts of the drivers reaches a maximum insured amount. (Figure to be published with the abbreviation: Figure 1);

Description

Title of the invention: METHOD FOR DETERMINING A GROUP OF DRIVERS TO ESTABLISH A PEER-TO-PEER INSURANCE SCHEME WITH HOMOGENEOUS RISK technical field

[0001] The present invention relates to the field of motor vehicles, more particularly to the field of insurance for drivers of motor vehicles. Prior art

[0002] Conventionally, known peer-to-peer (P2P) insurance schemes offer a better approach than traditional insurance in that these schemes exploit the principles of sharing and pooling risks between insured peers.

[0003] However, within these P2P insurance schemes, optimizing the selection of insured parties is paramount. A major challenge arises when P2P insurance grows, leading to an unequal distribution of risks and instability.

[0004] Indeed, as the P2P insurance group expands, the risk of imbalance in risk distribution increases. This imbalance stems from the potential disparity in risks among group members, which could potentially threaten the financial stability of the entire system. This instability represents a major problem to be solved, particularly in the context of P2P insurance schemes made up of motor vehicle drivers.

[0005] The published patent document US 2015 / 0161538 Al discloses a method for identifying individual profiles to create a peer-to-peer insurance scheme based on a criterion based on their shared interests on a social network.

[0006] However, the solution proposed in the document does not effectively resolve the technical problem stated above, because it groups risks based on risk factors, increasing the possibility of excess losses due to the simultaneous occurrence of risks. For example, the identification of insured profiles may be based on individuals sharing risky leisure activities, such as skiing, which makes the P2P insurance scheme unstable, since most skiers concentrated in the same P2P insured community are at increased risk of loss during the winter. Description of the invention

[0007] The present invention aims to overcome at least one of the drawbacks of the aforementioned prior art. More specifically, the invention aims to provide a simple and efficient solution for effectively selecting insured profiles sharing a homogeneous risk in a peer-to-peer insurance scheme.

[0008] To this end, the invention relates to a method for determining a group of drivers participating in a peer-to-peer insurance program, implemented by computer and comprising the following steps: a) determination of a knowledge graph of a plurality of drivers where each driver is linked to each of the other drivers by a weight weij representing a similarity of accident risk score, based on driving measures of said drivers; b) determination of a community of drivers among the plurality of drivers, presenting a homogeneous risk of accident, based on accident risk scores; c) determination of a first driver from the community of drivers and invitation of said first driver to participate in the peer-to-peer insurance program; d) determination of a next driver from the community of drivers by following a Hamiltonian shortest path in the knowledge graph, and invitation of said next driver to participate in the peer-to-peer insurance program; e) iteration of step d) until the sum of the insured amounts of each of the drivers accepting the invitation reaches a maximum insured amount.

[0009] According to one embodiment, the weight weij linking each of the drivers to another of the drivers in the knowledge graph is determined on the basis of a difference in the accident risk scores of said two drivers.

[0010] According to one embodiment, the weight weij linking each of the drivers to another of the drivers in the knowledge graph is determined, in addition, on the basis of a proximity measure my of risk linked to exogenous factors between the two drivers.

[0011] According to one embodiment, the weight weij linking each of the conductors to another of the conductors of the knowledge graph is weij = | UBI ( ni ) - UBI ( n, )| + (1- ) with mÿ = M( nj ) e [ 0; 1 ] where n, and correspond to two conductors, where UBI(ni) and UBl(nj) correspond respectively to the risk score relative to drivers rit and n-, ; and where M(nt, rij) is a proximity of exogenous risk factors between drivers rit and rij.

[0012] According to one embodiment, the weight weij connecting each of the conductors to another The similarity of accident risk scores between drivers in the knowledge graph is all the smaller as the similarity of accident risk scores is greater, and vice versa.

[0013] According to one embodiment, in step b), a plurality of communities is established and ordered according to the sums of the accident risk scores of the drivers of each of said plurality of communities, the community with the smallest sum of accident risk scores being selected.

[0014] According to one embodiment, the plurality of communities is established among the plurality of drivers by means of a Louvain algorithm.

[0015] According to one embodiment, in step c), the first driver among the community is determined on the basis of a minimum weight sum weij.

[0016] According to one embodiment, for each driver accepting the invitation, an insurance premium amount is determined.

[0017] The measures of the invention are advantageous in that the shortest Hamiltonian path drawn on the knowledge graph makes it possible to identify and classify the drivers of motor vehicles efficiently, thus guaranteeing a risk distributed homogeneously in the group of insured P2P up to the predefined insured capital.

[0018] The present invention makes it possible, on the one hand, to select the drivers presenting the lowest risk in terms of car insurance, and on the other hand, it allows drivers presenting a low risk to benefit from P2P car insurance with homogeneous and controlled risk, allowing them to considerably reduce their deductible and / or the amount of their insurance premium. Brief description of the drawings

[0019] [Fig-1] schematically represents a knowledge graph comprising a group of drivers participating in a peer-to-peer insurance program, said group being determined by means of a shortest Hamiltonian path in said knowledge graph, according to a method according to the invention. Detailed description

[0020] Fig. 1 schematically represents a fully connected knowledge graph 2 comprising a group of drivers selected by means of a method according to the invention to participate in a peer-to-peer (P2P) insurance program.

[0021] The method of the invention corresponds to a computer-implemented method for determining drivers suitable for joining the P2P insurance program, said method comprising a first step a) determining a fully connected knowledge graph from a given plurality of drivers, and a second step b) determining a community Cx of drivers from among the plurality of drivers, presenting a homogeneous accident risk as a function of accident risk scores based on driving measurements of said drivers.

[0022] Indeed, for each driver, the calculated risk score is preferably based on a record of telemetry data relating to the driving of a motor vehicle by said driver. Preferably, the risk score is also based on exogenous risk factors related to the driving of the motor vehicle. The risk score being: [Math 1] uBi( ni ) e [ai] where UBI(ni) corresponds to the risk score relative to a driver n, the score being between 0 and 1, with 0 corresponding to a high risk, and 1 to a non-existent risk.

[0023] The calculation of the risk score is not part of the scope of the present invention, but will nevertheless be explained later in this description.

[0024] In step b), a plurality of communities Cx is established and ordered according to the sums of the accident risk scores of the drivers in each of said plurality of communities Cx, the community with the smallest sum of accident risk scores being selected. This is the one represented in knowledge graph 2, with nodes A, B, C, D and E representing respectively one of the drivers in community Cx who presents the lowest risk.

[0025] Preferably, the plurality of communities Cx is established among the plurality of drivers given by means of a Louvain algorithm, in which each node is exclusively present in only one of the established communities Cx.

[0026] The nodes are connected by edges representing a risk link between two drivers, said link comprising a weight weij representative of a similarity of accident risk score.

[0027] The weight weij linking each of the conductors to another of the conductors of the knowledge graph 2 is determined according to the equation: [Math 2] wei . = | UBI ( n, ) - UBI ( nj) | + ( 1 - my ) where rit and nj correspond to two drivers, and where UBI(ni) and UBI(nj) correspond respectively to the risk score relative to said drivers n, and n, ; 1 JJ with my being a measure of proximity of risk related to exogenous risk factors between the two drivers n, and nh: 1 J [Math 3] niij = MÇ n,) s [ai] °ù Wni> nj) describes a proximity in the space of exogenous risk factors, in which my is a projection from the k-dimensional space (where k is the number of standardized exogenous risk factors) to the One-dimensional measure [0; 1]. If two nodes are very close in terms of risk, then this measure is close to 0; otherwise, it is 1. Calculating the proximity measure is outside the scope of the present invention.

[0028] In this configuration, the weight weij linking each of the drivers to another of the drivers in the knowledge graph 2 is smaller the greater the similarity of accident risk score and vice versa.

[0029] Thus, nodes having similar and close UBI(ni) risk scores in the space of exogenous risk factors have a lower cost or weight weij on the edge that connects them.

[0030] On the example of the knowledge graph 2 illustrated in [Fig. 1], we can see the different weights weij linking the nodes in the form of a number (in parentheses) between 0 and 1.

[0031] Advantageously, the method of the invention includes a step c) of determining a first driver among the community of drivers in the knowledge graph 2, in order to launch the shortest Hamiltonian path in said graph 2 and thus successively invite the drivers connected by said Hamiltonian path to form the P2P insured group.

[0032] Preferably, the first conductor is determined on the basis of a minimum sum of weights Weÿ. In this respect, a degree (or valence) weighted by previously determined weights wetj is calculated for each of the nodes of the knowledge graph 2, and the node with the smallest of the calculated weighted degrees corresponds to the first conductor.

[0033] In knowledge graph 2, the calculated weighted degrees are represented in the following table: [Tables 1] Weighted Degree Node A 1.4 B 1.5 C 2 D 1.8 E 1.7

[0034] Since the weighted degree of node A is the lowest, then the corresponding driver A is the first driver to be invited to join the P2P program.

[0035] The method includes a step d) of determining the next driver from among the community of drivers by following the shortest Hamiltonian path in the knowledge graph 2. This time, the choice of the second node is based on the weight weij, the latter being the weakest among those of the different edges.

[0036] Preferably, the selection of the next node is accompanied by a calculation of the insured amount for the corresponding driver, so as to verify that the sum of the insured amounts of the drivers connected by the shortest Hamiltonian path does not exceed a maximum insured amount corresponding to a maximum insurable capital predefined prior to the selection of each of the subsequent nodes. Determining the insurable capital is not part of the scope of the present invention. A formulation of the insured capital threshold for P2P insurance is given by: (Denuit, Stop-loss protection for a large P2P insurance pool, 2020).

[0037] Advantageously, step d) is repeated until a community of P2P insureds reaches the predefined maximum insurable capital, the drivers selected by means of the Hamiltonian path are invited as they go along to join the group of insureds.

[0038] The Hamiltonian path is represented by a dotted line. It can be seen that the latter connects, in order, the nodes A, C, B, D and E.

[0039] For each driver accepting the invitation, an insurance premium amount is determined. The proposed premiums follow a credibility method, with the premium proportional to the expected individual risk, adjusted for community risk (outside the scope of the present invention). The method for calculating the premium amount can be based on credibility theory (Kaas Rob, 2008) (Ronka-Chmielowiec Wanda, 2005) and, particularly for automobile insurance, on telemetry signals (Corradin Alexandre, 2021).

[0040] The calculation of the UBI(ni) risk score for a driver n is disclosed below, although it is not part of the present invention.

[0041] To predict a probability pt of reporting a claim from a certain node: [Math 4] PUBI a gradient reinforcement classifier (machine learning algorithm) is chosen based on the telemetry driving data of the corresponding motor vehicle, as well as other risk-related data projected over a defined period, which could be, for example, a second, a trip, a day, a month, etc.

[0042] The input data set is represented as follows: [Math 5] where y is a binary target-variable vector indicating a loss within the specified timeframe, and where X is a matrix of explanatory variables composed of: [Math 6] X separating exogenous risk factors from other data. where XM£SFs are variables indicating risk factors leading to mass accidents if they occur (e.g., proximity to rivers, areas where hailstorms are frequent, members sharing the same driving habits, such as risky behavior only in winter, etc.), and where Xo is a matrix of all other risk factor candidates, excluding exogenous factors.

[0043] For classification, a differentiable loss function is defined: [Math 7]

[0044] Gradient-enhanced classification requires starting with an initial value of: [Math 8] F0(x) ^argmin^^L^y^ y)

[0045] Next, the iteration from m=l up to M is considered as follows: 1- Residues: [Math 9] _ [^(^))1 for an th observation, and for one of my iterations; 2- The rim residuals are the object of the next iteration for fitting a regression tree with a leaf (or terminal node of the tree) RJm for j = 1, ..., Jm; 3- For each sheet j = 1, Jm, reset: [Math 10] Yjm = argmin^^.^yf, F^x^ +y) 4- A new Fm(x) is calculated based on the estimates from the previous iteration, the fixed learning rate v and the estimates from the last iteration as follows: [Math 11] Fm-M + ^^L[yy]

[0046] Based on the final definition of Fm(x), the probability pt of declaring a loss by a given node n; is established for {nB n2, ..., nm}.

[0047] The work of JH Friedman on gradient strengthening is the reference for further explanations and variations (Friedman, 2001), (Friedman JH, 1999).

[0048] Advantageously, the present invention makes it possible, on the one hand, to select the drivers presenting the lowest risk in terms of motor insurance, and on the other hand, it allows low-risk drivers to benefit from homogeneous risk P2P motor insurance enabling them to considerably reduce their deductible and / or the amount of their insurance premium.

Claims

Claims

1. A method for determining a group of drivers (A, B, C, D, E) participating in a peer-to-peer insurance program, implemented by computer and comprising the steps of: a) determining a knowledge graph (2) of a plurality of drivers (A, B, C, D, E) where each driver is linked to each of the other drivers by a weight weij representative of a similarity of accident risk score, based on driving measurements of said drivers; b) determining a community of drivers (A, B, C, D, E) from among the plurality of drivers, having a homogeneous accident risk, based on the accident risk scores; c) determining a first driver (A) from among the community of drivers (A, B, C, D, E) and inviting said first driver (A) to participate in the peer-to-peer insurance program;d) determining a next driver (C) from the community of drivers (A, B, C, D, E) by following a shortest Hamiltonian path in the knowledge graph, and inviting said next driver (C) to participate in the peer-to-peer insurance program; e) iterating step d) until the sum of the insured amounts of each of the drivers accepting the invitation reaches a maximum insured amount.;

2. Method according to claim 1, in which the weight weij connecting each of the conductors (A, B, C, D, E) to another of the conductors (A, B, C, D, E) of the knowledge graph (2) is determined on the basis of a difference in the accident risk scores of said two conductors.

3. Method according to claim 2, in which the weight weij connecting each of the conductors (A, B, C, D, E) to another of the conductors of the knowledge graph (2) is determined, in addition, on the basis of a proximity measure my of risk linked to exogenous factors between the two conductors.

4. Method according to one of claims 1 to 3, in which the weight weij connecting each of the conductors (A, B, C, D, E) to another of the conductors (A, B, C, D, E) of the knowledge graph (2) is u ;. ; = J UBI ( ni ) - UBI ( nj ) | + ( 1 - ) with mi, — M ( Hj, n,) e [ 0; 1 ] where n, and n, correspond to two conductors, where UBI(ni) and UBl(nj) correspond respectively to the risk score relative to drivers n, and n, ; and 1 JJ where M(nh rij) is a proximity of exogenous risk factors between drivers n, and nh

5. Method according to one of claims 1 to 4, in which the weight weij connecting each of the conductors (A, B, C, D, E) to another of the conductors (A, B, C, D, E) of the knowledge graph (2) is all the smaller as the similarity of accident risk score is great and vice versa.

6. A method according to one of claims 1 to 5, wherein, in step b), a plurality of communities is established and ordered according to the sums of the accident risk scores of the drivers (A, B, C, D, E) of each of said plurality of communities, the community with the smallest sum of the accident risk scores being selected.

7. The method of claim 6, wherein the plurality of communities is established among the plurality of drivers using a Louvain algorithm.

8. Method according to one of claims 1 to 7, wherein, in step c), the first driver (A) among the community is determined on the basis of a minimum sum of weights weij.

9. Method according to one of claims 1 to 8, in which for each driver (A, B, C, D, E) accepting the invitation, an insurance premium amount is determined.

Images