FAQCTUARY

          FAQctuary-V2

          The FAQctuary series

          The FAQctuary are a series of educational papers dedicated to the insurance sector. Each issue addresses a specific actuarial topic, expressed as a question asked by market players. FAQctuary's are published by members of the Detralytics team and written in a clear and accessible language. The team combines academic expertise and business knowledge. Detralytics was founded to support companies in the advancement of actuarial science and the solving of the profession’s future challenges. It is within the scope of this mission that we make our work available through our DETRAnotes and FAQctuary series.

          Subscribe

          List of publications

          Does autocalibration improve goodness of lift?
          Date : July 2022
          Author : Nicolas Ciatto, Harrison Verelst, Michel Denuit and Julien Trufin
          Publisher : Detralytics
          Language : English

          Autocalibration is a desirable property, intimately related to the method of marginal totals that predates modern risk classification methods. Autocalibration can easily be implemented using the practical method proposed by Denuit, Charpentier and Trufin (2021), consisting in an extra local regression step. Professional practice favors Lorenz curves and Gini coefficients to assess the performances of a candidate premium. Under autocalibration, Denuit and Trufin (2021) established that the advanced diagnostic tools proposed by Denuit, Sznajder and Trufin (2019) reduce to Lorenz curve and Gini coefficients, respectively. The present note aims to assess the impact of autocalibration on the goodness of lift. Lift graphs compare average predicted and average actual loss costs when policies have been ranked according to the candidate premium and grouped within buckets. It is shown on a case study that autocalibration does not only restore global and local balances but also improve lift.
          Try our autocalibration dashboard!
          How to assess broker's performances in an elementary way?
          Date : February 2022
          Author : Michel Denuit and Julien Trufin
          Publisher : Detralytics
          Language : English

          In order to rate intermediaries, Detralytics has developed a credibility model separating pure randomness from structural differences in loss ratios. The output results in a reliability index quantifying the confidence one can place on underwriting experience for a particular intermediary (based on the credibility coefficient) and a performance index assessing the quality of the intermediary’s production (based on the credibility predictor of individual random effects). In this note, an elementary approach is proposed, treating the broker’s efficiency level as a fixed effect (rather than a random effect in credibility). This provides the actuary with a preliminary ranking pointing to the ineffective intermediaries.
          ICC, deviance, AUC : What are the differences ?
          Date : April 2021
          Author : Louise d'Oultremont, Michel Denuit and Julien Trufin
          Publisher : Detralytics
          Language : English

          Assessing performances of different models is essential in modelling processes. Many measures exist to determine these performances and draw conclusions. This FAQctuary compares the behaviour of three measures on different datasets. The measures under consideration are the integrated concentration curve, the deviance and the area under the curve (AUC).
          Features with flat partial dependence plots up to a certain level: not important?
          Date : November 2020
          Author : Elke Gagelmans, Michel Denuit and Julien Trufin
          Publisher : Detralytics
          Language : English

          This paper focuses on partial dependence plots which are often used when modeling with machine learning techniques in order to better understand the effects of the features on the conditional expectation of the response variable. However, these plots must be interpreted with caution. Indeed, they can easily lead to wrong interpretations in case the analyst is not enough familiar with these plots. As noticed in a previous FAQctuary, a typical situation is the case where a feature is important because of its interactions with others while its partial dependence plot is flat. In this FAQctuary, we go one step further and we consider a very simple example with a three-way interaction effect and we show that only looking at partial dependence plots for each feature and for two features may indeed lead the analyst to wrong conclusions.
          Generalization error for Tweedie models: can it be decomposed as for Normal and Poisson and can it be reduced with bagging?
          Date : 2020
          Author : Michel Denuit and Julien Trufin
          Publisher : Detralytics
          Language : English

          (This letter has been published)
          Wüthrich and Buser (DOI:10.2139/ssrn.2870308, 2020) studied the generalization error for Poisson regression models. This short note aims to extend their results to the Tweedie family of distributions, to which the Poisson law belongs. In case of bagging, a new condition emerges that becomes increasingly binding with the power parameter involved in the Tweedie variance function.
          Features with flat partial dependence plots: not important?
          Date : April 2020
          Author : Elke Gagelmans, Michel Denuit and Julien Trufin
          Publisher : Detralytics
          Language : English

          This paper focuses on partial dependence plots which are often used when modeling with machine learning techniques in order to better understand the effects of the features on the conditional expectation of the response variable. However, these plots must be interpreted with caution. Indeed, they can easily lead to wrong interpretations in case the analyst is not enough familiar with these plots. A typical situation is the case where a feature is important because of its interactions with others while its partial dependence plot is flat. In such a case, an analyst who would only base his analysis on this plot could be tempted to conclude that the feature is not important to explain the conditional expectation of the response while he would be wrong. In this FAQctuary, we aim to illustrate such a situation with the help of a simulated example that is very simple.
          How to combine market and internal medical indices in orsa ?
          Date : January 2020
          Author : Candy Mahirwe, Audrey Meganck, Michel Denuit and Julien Trufin
          Publisher : Detralytics
          Language : English

          “As part of the Own Risk and Solvency Assessment (ORSA), health insurance companies using the standard formula are required to assess the appropriateness of this formula for among others the medical inflation risk covered within the Health SCR module. However, for some companies, the medical index does not seem to reflect the real evolution of their own medical expenses. Therefore, we choose as topic of our first FAQctuary, the medical index in which we will attempt to answer this following question : - How to combine market and internal medical indexes in ORSA ?”