Evolving Techniques and Capabilities
By Paul E. Lacko

When and how will the economy react to the Fed's continuing increases in short-term interest rates? Is there a real estate bubble? When will the real estate bubble burst? When will the insurance cycle turn again? How can we guarantee that a seemingly adequately capitalized insurance company will still be in business two years from now?

So many questions. So few answers. But we do seem to be making progress.

When I started in this business, a model was a theoretical construct, not an analytical tool. The model guided one's selection of what variables to measure and what calculations to perform on the data. Mainframe computers collected a lot of data, and sophisticated in-house systems could even generate loss triangles and print them on microfiche.

When I started in this business, a model was a theoretical construct, not an analytical tool. The model guided one's selection of what variables to measure and what calculations to perform on the data.

Note to readers born after the mid 1960s: Microfiche, a form of hard-copy output, was an incredible leap forward in data storage technology. Dozens of pages of paper output could be printed in extremely small font on a single rectangle of translucent or transparent plastic. These were a wonderful invention at the time, because an entire file-room's worth of paper reports could be stored in a couple small boxes on a shelf at one's desk. All you needed was a fiche reader, which was a desk-top projection screen that worked just like a microscope: turn on the small light, place the fiche card on the glass slide, look at the screen in front of you, adjust the focus, and move the fiche around to view different pages.

The data then had to be converted to another medium before any serious data processing could be done. "Data processing" means someone, a clerical assistant, perhaps, wrote down the numbers on paper and then did all the calculations manually with a desktop calculator or entered the numbers into a computer program written specifically to accomplish the task at hand. (Some desktop calculators often did only arithmetic and subtraction.) Programming was tedious and time consuming, but it beat calculating manually. Most programming software allowed one- and two-dimensional arrays of up to several hundred elements. (APL, an important exception, was designed specifically to work with large n-dimensional data arrays. But most of us wrote programs in BASIC or FORTRAN or PL/1.)

With the program written, tested, debugged, and saved, calculations could be done again and again. Open the old program file, enter new data, run the program, and pick up the paper output.

Technology had not yet achieved the best of all possible worlds, though. Processing time could be very slow- several minutes to several hours. A simulation study with 10,000 iterations could take several days. All the output was hard copy, and printing was slow by today's standards. It seemed like printer speeds were measured in pages per hour.

Computer-generated graphs were poor-quality at best, so you had to do your graphs by hand with colored pens. Or maybe your actuarial department had purchased a "plotter," a special hardware tool that had moveable arms. Colored pens were stuck into the plotter's moveable arms. Your program moved the arms around and lifted the pens up and down, thereby drawing points and lines on the paper.

Things sure have changed, haven't they? Nowadays we download the data from one computer directly into another computer and import the data into a software package or a spreadsheet. We set assumptions and parameters in the software and get results almost instantly. We can fit data to curves, create graphs and charts, run thousands of Monte Carlos iterations using dozens of probability distributions as input variables, and never touch a piece of paper-until we report to senior management, of course.

Computer models have become increasingly important in the insurance industry over the last 25 years. We rely more and more on model outputs to generate distributions of outcomes, whether we're looking at pricing terrorism coverage, setting growth targets by territory for the workers comp book, or fine-tuning a private passenger auto rating plan.

In the predictive modeling session at the 2005 CAS Spring Meeting, panelist Cheng-Sheng Peter Wu pointed out that "predictive modeling" refers not so much to new techniques as to new capabilities. Data storage and data processing speed are no longer practical concerns for most of us. Computers can handle hundreds of variables and millions of data records in real time.

The hard part now is to pull together and refine the data. Daniel Finnegan, president of ISO Innovative Analytics and co-panelist at the predictive modeling session, described the huge project underway by his group to collect, correlate, and analyze hundreds of data variables from dozens of sources. They want to create a statistical model of the U.S. virtually by person and by street address. Their longer-term goal is to create personal lines pricing models for the entire U.S.

Despite what Mr. Wu said, I think that there are new techniques, certainly compared to 25 years ago. A couple that come to mind are genetic algorithms and neural networks. Neural networks have proven useful, in the sense of being adaptive models that, in effect, fine tune themselves to improve their modeling capabilities as more data pass through the model. Actuarial models may not yet incorporate genetic algorithms, but applications may arise in the future. Genetic algorithms are tools for constrained optimization, and most actuarial models can be characterized as such. (To be continued.)