Minnow Bear Computers Michael Nunamaker's selections: Where they've come from, where they are going

Michael Nunamaker's selections: Where they've come from, where they are going

1) The beginnings
2) The search begins: Impact value systems
3) The next step: fundamental ratings
4) Bob Selvin, National Turf, and regression
5) Self-modifying analysis software
6) Status in 1999
7) What will the next decade bring?
8) A graph showing Mike Nunamaker's improvement over time

The beginnings

Way back into 1974, Frederick Davis published "Thoroughbred Racing: Percentages and Probabilities" and "Thoroughbred Racing: Probability Computation". These booklets contained the first impact value based system that I had ever seen. While it had a relatively low winning percentage (around 23% in my tests), it also was considerably profitable. I believe that this simple system remained profitable until April 1, 1992. Until then, my own handicapping didn't go beyond simply applying Davis's system. After all, why tinker with success?
But on April 1st, 1992, the Daily Racing Form started publishing Beyer speed ratings. The result was that any handicapper could see at a glance which horses were the fastest. As a result, many systems and methods that had worked for decades ceased to be profitable. While they still picked winners as frequently as they had in the past, the mutuel pools had become too efficient for profits. Indeed, one commercial method that I'm sure used to work, now shows a -20% ROI on my post Beyer database! Obviously, the only way to do well at the track was to use much better handicapping methods.

The search begins: Impact value systems

My first work in trying to improve upon Fred Davis's system was to simply try to modernize impact value systems. After all, Fred Davis's method's didn't assume the availability of sophisticated computers and modern speed ratings. After a considerable amount of work, I was able to raise my winning percentage on all races from 23% to 26%. Now those of you who are familiar with impact value based systems may think this winning percentage is low, and it is compared to most literature. The reason is that I am reporting the winning percentage for all races. Most people skip maiden races or races with many first time starters and other difficult races. For example, my tests have shown that Paul Peterson's Morning Liner system wins 24.9% of the time when applied to all races. Let me make clear that I don't advocate playing every race on the card. However, I do think that playing every race on paper is an extremely useful measurement of the quality of a handicapper.
The crown jewel of my work in this period of impact value research was my impact value search software. Using a variant of the Fast Fourier transform butterfly, I wrote software that can search through millions of different impact value based methods very rapidly. For example, on a 90 Mhz Pentium and a 10,000 race sample, I can calculate the performance of 2 million impact value systems in one night! The only other software I'm aware takes this long to test a single system.
As you can well imagine, being able to examine a type of system two million times faster than anyone else does have it's advantages. The impact value systems that I developed have significantly higher winning percentages and ROIs than anything I've seen published anywhere at anytime. The final result of all this was that by the end of 1993, my winning percentage on all races stood at 26%. A big improvement, but there was still a negative ROI.

The next step: fundamental ratings

My research at this point moved to a much more fundamental level. For example, virtually every handicapper agrees that speed ratings are extremely useful, but what is the best way to use them? I wrote software that examines thousands of different ways to utilize speed ratings. I tested pace line selection, regression weighting, time decay weighting, simple averages, complex averages based on distance patterns, best of ratings, worst of ratings, and many many others. I also wrote software to generate many speed, pace, and other ratings of my own. Because these would be available to no other handicappers, I thought they would give me an advantage over other players. I was right in the winning percentage and ROI improved, but still no consistent profits. I made no progress for about a year on overall winning percentage simply because that's not what I was working on. I was aiming at finding profitable spot plays.
I actually found quite a few that worked (and many are published in my books Favorites!, Longshots, and my track specific studies. The problem with these spot plays is that the profitable ones generally have a relatively small number of plays. There simply isn't enough action in the ones that make money to have the kind of income I can generate from my computer work. This kind of research continued until the end of 1994. In November of that year, I received a phone call that wound up significantly changing the direction of my research.

Bob Selvin, National Turf, and regression

In November of 1994, Bob Selvin of National Turf Phone Seminars called me and invited me to join their staff of handicappers. The National Turf handicappers are the finest handicappers that Bob can find and they all record daily seminars for the Southern California circuit. Bob had noticed that my books had gotten more "9" and "9 1/2" ratings from the Phillips Racing Newsletter than any other author in history. He figured if my computer work could produce books of that quality, I should be able to do it for handicapping. I told him that there were several kinds of races that I had never done any work on and I needed some time to research them and write software to analyze those race types. For example, I had never looked at races full of first time starters or turf races. He said "OK, let me know when you're ready."
It took several months of preparation. I took all of the fundamental advances that I had made and wrote special step-wise multiple regression analysis software to use my new ratings to predict every kind of race offered. If you look at the graph of my handicapping progress, you can see this was the period of the fastest increase in my overall handicapping winning percentage.
I went on-line with National Turf on April 28, 1995. Since then, I've made it my quest to improve the overall accuracy of my prediction software in every type of race. All throughout the second half of 1995, I made progress in improving the winning percentage of my prediction software. And when the winning percentage increases, the overall accuracy of the prediction for every horse in the race increases also! While the headway I was making was gratifying, it was also annoying that I had these incredibly powerful computers to work with, and they weren't actively helping me to do any research. In August of 1995, it occurred to me that I could write software that could essentially perform independent research on its own.

Self-modifying analysis software

There is nothing revolutionary about the idea of a program teaching itself. Indeed, that is the entire point behind things such as neural nets and genetic algorithms. My idea was to go up to a slightly higher level than this. I wanted routines that could learn in several different ways. My initial work on this idea uses standard multiple regression as a framework. As of June 20, 1996, I use the ninety two most predictive factors that I've found in racing. These are converted into numeric form and normalized where needed. It's at this point where things get interesting.
My software can then take the entire database and automatically perform lengthy step-wise regression analysis on every conceivable sub-set of my database and find out what set of factors produces the best results. The resulting equations are automatically fed back into my software with no additional programming needed. This allows me to easily update my regression equations as I add more data to my database. Likewise, new equations that utilize new racing ratings that I program are extremely easy to generate. In effect, I've converted the computer into my research assistant. Early this year, on this software's maiden run, it was able to increase the winning percentage of my top selections by almost half a percent. I've since converted these new routines into the "production routines" that I use for my daily selections.

Status as of 1999

In 1996, I began acquiring data for all tracks in the country. This caused the winning percentage of my top pick to drop quite dramatically to 27%. While the overall winning percentage at the larger tracks was probably still the same, I always look at my entire database, not just part of it. Also in 1996, I was hired as a consultant by Experian. That is the company that produces everyone's credit reports. Since that time, I've put most of my time into writing software for them. However, in late 1998 and 1999, I decided I should try some analysis that I had been thinking about for a long time, but computers were never powerful enough to do it. 250 _trillion_ calculations later, and the winning percentage on my top pick has increased a whopping 2.55%!!! I'll be playing with refining this software over the next year. It should be interesting to see what I come up with.

What will the next decade bring?

What does the long term future hold? My long term goal is nothing less than trying to build prediction software that is more accurate than any human handicapper on the planet. While many will scoff at this goal, I see no reason why this can't be done within five to ten years. As recently as two or three years ago, there were still chess experts predicting that a computer would never beat the world chess champion. A chess playing computer beat the world chess champion in 1996.
While handicapping is more complex than chess, it is very much a game of statistics and information. These are things that computers excel at processing. And as my database goes, even my existing software will be able to look deeper than it has in the past. But rest assured I will not remain content with my existing software. I shall push the state of the art as hard and as vigorously as anyone ever has. I hope you'll be along for the ride.

The following graph shows the progress of my handicapping software over time. For reference, I've included Paul Peterson's Morning Liner system and some marks on "good public handicappers." It's very difficult to come across information on how handicappers perform when handicapping every single race. Most like to brag about their specialty, or they don't even maintain statistics. But in his book, "Woulda, Coulda, Shoulda", Dave Feldman states that good public handicappers win between 25% and 29% of their races. From what I've read from other sources, I think this might be slightly overstating it, but it is a useful benchmark nonetheless. As you can see, my prediction software currently performs better than the high end of this range.