November 2018

Introduction to Machine Learning

By Bill Rearden

Actuaries have long held strong strategic positions at the core of insurance. Artificial intelligence (AI) at the moment promises to simplify much of the quantitative analysis done by actuaries for policy underwriting. Not surprisingly, many actuaries like myself are curious to know more about AI and how it will influence the profession.

The Entrepreneurial & Innovation Section of the Society of Actuaries has sponsored sessions on AI at meetings, online webcasts, and virtual townhalls. From these events we have learned that AI is still in its infancy, but it is likely to improve exponentially with computational efficacy. At its core, AI is a prediction technology to improve decision making. In the near future, the accuracy of AI predictions is expected to make professional judgement 1 more important.

Machine learning (ML) is one of the drivers of AI. As we continue to generate oceans of data, ML allows us to rapidly uncover interesting relationships. This means that actuaries will have new data for potentially more powerful underwriting analysis. More importantly, with ML, we can develop better predictions and expand the areas where they can be applied. In this article, I present simple and helpful answers to explain in detail the particulars of ML.

ML differs from the conventional analysis in that it takes a more “down-up” approach. It searches through the data looking for relationships to make predictions. That is, ML has no prior knowledge or bias of the system generating data under consideration. The goal of ML is not to explain the underlying system, but instead, to generate the finest “out-of-sample” predictions. ML is a tool that let the data tell their own story. 

Supervised ML is based on algorithms that automatically search for best predictive interactions. Here is how supervised ML works 2:

  1. Randomly partition the data into training and validation sets,
  2. find interactions that generates best fit to training data—regularization,
  3. test the prediction power on validation dataempirical tuning,
  4. repeat steps 1 to 3, and
  5. Stop when optimal predictive model is found or at iterations limit

The objective of Step 5 is to find the optimal prediction model by maximizing the regularization, subject to the empirical tuning constraint. The constraint helps prevent over fitting models that have zero predictive power when applied to new data. Learning models for regularization can be as simple as linear models or exotic and complex as deep neural nets. ML can take huge amounts of data and divide them onto thousands of neural net layers. This predictive capability using neural net layers summarizes the anticipated excitement currently surrounding artificial intelligence.

The biggest challenge with ML is that it can produce different models with similar predictive results. Consider predicting life insurance premiums from detailed characteristics—age, gender, smoker, etc. Two ML runs can predict the same premium. But one model may place more weight on age, while another on smoking. This happens because the ML process starts without any prior knowledge about the data; it blindly begins calibrating algorithms by randomly selecting input variables. The guiding principal of ML is empirically tuning the regularization process until the best out-of-sample prediction model is found. The instability of variables explains why machine output needs human judgment for assistance.

With faster computers and the introduction of quantum computing, AI is likely to automate much of the quantitative actuarial analysis. However, generally speaking, the biggest hurdle with AI is the way in which ML works. The motto, “garbage in—garbage out still applies to the data supplied into ML analysis. This is an all too familiar concept for actuaries and is a concern in regard to AI automation. The silver lining, at least for the foreseeable future, is that AI will not eliminate or displace actuarial work. More importantly, the value of actuarial judgement is likely to be heightened in this new arena of AI automated risk underwriting.

Bill Rearden, ASA, MA, is co-founder and CEO at Ironbound Consulting Group on Wall Street in Manhattan. He can be contacted at

1 Agrawal et al, “Prediction Machines: The simple Economics of Artificial Intelligence”, Harvard Business Review Press, 2018.

2Mullainathan and Spiess, “Machine Learning: An Applied Econometric Approach”, Journal of Economic Perspectives, Volume 31, No. 2, Spring 2017.