Data management and interactions with predictive models and simulations

In a risk management-based approach, future models will be produced to compute scores, manage underwriting and handle claims. They will have less “black box” traits and will be more sophisticated with machine learning aspects. Enterprise Risk Management considerations will be constraining the use of these new models. For example, the ERM will ease the characterisation of which analytics models should not be used for a specific purpose because they may create too much risk. Also, what are the cases of good practice when using these new predictive techniques?  In terms of data, what data strategy should be used and how can existing systems be transformed to adapt to big data while continuing to manage risk properly? This corresponds to modelling management issues for analytics.

Some questions of speeding up calculation times could also be addressed from a predictive modelling standpoint in connection with the Chair’s previous simulation studies on simulations for Solvency II.

The topic of attitudes to models could also be studied as part of analytics. How might managers react to a model with little capacity for interpretation when the analytics actuary is unable to recount how it delivered the results? What is the technical cost of interpretability (how much is lost in quality if one requires a minimum degree of interpretability?) How can the interpretability of results from an analytics-type model be precisely defined? How might policyholders react to the possibility of a rate that varies widely over time due to segmentation pushed to extremes? How can segmentation, ethics and controlling the stability of rates over time be reconciled? What is the cost in performance terms of constraining the average stability of rates over time for a policyholder?

Questions on selection bias and the risk of fraud or data manipulation by policyholders could also be addressed. How could this risk be managed, any eventual manipulations be detected and how could selection effects be taken into account over time?

Finally, consideration could also be given to studying changes that enable approaches linked to the use of data from connected objects. Some teams at BNP Paribas Cardif have to monitor digital developments in the insurance world while others are introducing connected solutions, for example, home insurance policies (Habit@t, Italy) using a “box” which monitors the house using sensors. There are also vehicle insurance policies via mobile-phone “Pay how you drive” apps that analyse the driver’s behaviour behind the wheel and reward careful drivers (UK). Can this data alter the traditional insurance model by introducing incentives and prevention in a much more pronounced way than at present?