Derivative Lasso: Credibility-Based Signal Fitting for GLMs

Akur8 release a new research paper for the actuarial community entitled “Derivative Lasso: Credibility-Based Signal Fitting for GLMs.” Developed by a team of actuarial scientists at Akur8, including  Mattia Casotto, Head of Product for the U.S. and Principal Scientist, and Thomas Holmes, Chief Actuary for the U.S., the paper proposes a solution to the challenge actuaries often encounter: whether to build transparent GLMs that require significant cost and expertise, or to automate the model building process using non-transparent techniques that could limit the ability to deploy models in production.

Actuaries apply mathematical and statistical methods to provide sound estimates of risk. For the past two decades, the de-facto standard statistical methodology has been Generalized Linear Models (GLMs). GLMs have gained widespread popularity due to their ability to explicitly define statistical assumptions in data, model correlations among variables, and directly generate output in the form of a rating table. While in recent years there has been an increase in the usage of other predictive modeling techniques from the machine learning field, such as GBMs and Random Forests, these techniques lack transparency for the models created. This has made it difficult for actuaries to interact with and adjust their models to incorporate selections and actuarial judgment.

“The objective of this new paper is to propose a solution to this artificial dilemma – that GLMs can natively detect non-linearities and manage low-credibility segments,” noted Mattia Casotto, Head of Product for the U.S. and Principal Scientist at Akur8. He continued, “We will illustrate how lasso allows GLMs to include credibility considerations, and how the derivative lasso, a variation of the lasso penalty, natively models nonlinearities. The modeling technique preserves the standard GLM parameterization and results in only minor adjustments to the GLM optimization formula.”

“Our team of actuarial scientists at Akur8 worked closely together to produce this comprehensive research paper. We are thrilled to publish this new information in an effort to continue to expand the literature available to the actuarial community on this important topic,” noted Samuel Falmagne, Co-founder and CEO at Akur8. 

Specifically developed for actuaries and predictive modelers, Akur8’s solution enhances pricing processes by automating risk and rate modeling using proprietary transparent machine learning technology. The core benefits for insurers include a reduction in data preparation and modeling time, automated model building, transparent GLM outputs, data-driven underwriting and no coding required, all of which effectively accelerates time to market while ensuring full transparency and control of the models created.

The paper can be downloaded here: https://www.akur8.com/white-papers/derivative-lasso-credibility-based-signal-fitting-for-glms