Inexplicable, Say I

Stephen recently questioned whether the hype around AI models for Life Insurance might be a case of The Emperor's New Clothes. In this blog we discuss an important point of difference: whereas in the fable, a youth reveals the expensive "invisible" new clothes have no substance at all, in our scenario, we find precisely the opposite. AI models utilising machine learning are, far from being see-through, simply not transparent enough.

Stephen's post aptly described the relationship between an AI Modelling actuary and the data being modelled as "a dangerous mental distance". Some might see this aspect as less critical for data scientists dealing with population-level datasets where individual correctness is far beyond the purview of the analyst. In such vast seas of records, small-scale data problems might be expected to average out. However, in portfolio experience analysis, where datasets are not generally of that scale, data issues affecting small numbers of high-value records can make the difference between profit and loss. Increased mental distance can be so dangerous in that context that professional and regulatory obligations apply to the modelling actuaries perform.

As one example of this, in a very early blog, Stephen discussed the Technical Actuarial Standards (TAS) applying to the UK Actuarial Profession at that time (other professional bodies doubtless carry similar obligations). Now that the older TAS-D, TAS-M and TAS-R have evolved into a suite of generic and specific TAS guidance, it is worth considering TAS within the context of AI modelling.

Starting at the model foundation, TAS-100 Principle 3 covers data, and places obligations on the actuary to understand the data underpinning any actuarial outputs. This includes consideration of data accuracy and completeness, the application of suitable checks, and identification of any material bias. This means an actuary cannot exploit the short cut offered by unsupervised machine learning techniques, since it isn't good enough for the neural network to "understand" the data - the actuary must do so. While auto-magical pattern identification within a neural network might be helpful for unregulated businesses, the obligation to understand the data beneath a financial model is very far from optional.

A lynch-pin of TAS guidance is the reliability objective:

To allow the intended user to place a high degree of reliance on actuarial information, practitioners must ensure the actuarial information, including the communication of any inherent uncertainty, is relevant, based on transparent and appropriate assumptions, complete and comprehensible.

General Actuarial Standards (Purpose, Clause 1.3)

A key focus in this objective is transparency, such that the actuarial deliverable can be understood, deemed appropriate and used with confidence. An issue immediately arises that work based on AI modelling must admit a severe lack of transparency. These models have a notorious "black box" problem, and largely reach their conclusions without an actuary having any hope of fully understanding why.

The problem deepens when we consider the applicable guidance to the use of actuarial models:

Practitioners must ensure they understand the models used in their technical actuarial work, including intended uses and limitations

General Actuarial Standards (Models, Clause 5.1)

The requirement for understanding a machine-learning model might be politely termed a big ask, since an actuary is expected to vouch for (and ultimately explain) the character and suitability of an opaque network of thousands of numerical weights. For this reason, the examples referencing AI in the 2024 TAS Modelling guidance include preferring models based on inherent explainability and, where needed, using the Local Interpretable Model-agnostic Explanation (LIME) technique. LIME involves using the predictions from the AI model, while taking explanations from a simpler surrogate regression model. The LIME approach, sadly, has a variety of such significant problems that the data scientist writing the previously linked document concludes by recommending the use of fully-interpretable models. Other research highlights the as-yet unsolved problem of concretely evaluating the performance of LIME-based explanations. Explaining AI model output, it seems, is an area replete with unsolved problems.

The bottom line is that fully explaining AI model output is not currently possible. The problem remains so "interesting" (read intractable) that it has inspired a highly active field of research. The upside of all this activity is that breakthroughs could well emerge over the next few years. The downside is, until that happens, it remains difficult to justify choosing AI models that cannot meet actuarial and regulatory standards when existing statistical models have a long pedigree of ticking the necessary boxes.