have been used as effective tools in nearly every business sector of the global
economy. They are used in hospitals in help predict the staging
and prognosis for cancer patients and they are used by credit card companies to
anticipate fraudulent activity before it occurs.
Simulators have been quite useful
in predicting the future of economic activity and the subsequent price
movements of stock, commodities, and currencies. The ability to predict the
future accurately and be financially rewarded accordingly has grown
exponentially with the mammoth advances in technology over the past two
The ability to organize and
structure enormous databases of information has become the foundation of nearly
every business enterprise. The goal is to simply learn from the past mistakes
and successes and then apply that accumulated knowledge to the current
environment. Ultimately, our goal here is to learn when teams have the greatest
probabilities to win or lose game.
The statistical method called
backpropagation is the foundation for
many numerical problems. For example, currency prices for a specific country
can be forecast based on current levels of economic activity and through the
use of historical patterns. History does not always repeat itself in the
financial markets, but there are times when historical outcomes are very useful
to predicting a future movement of a stock market.
There are similarities between
sports betting and the stock market. In both cases, the linesmaker (market
maker) does not have the priority of wanting a specific outcome. Their primary
focus is to determine what the market will bear and establish a price level
where there are equal buyers and sellers. The profits of both types of business
are essentially made the same way. For the financial markets, commissions, and
market pricing make up the large part of profits. In sports gambling, the betting odds maker charges a 10% ‘fee’ for the opportunity to bet a game.
Backpropagation was developed by
Paul Werbos in the early 1970’s and grew in popularity and acceptance with the
work done by Rumelhart and McClelland in 1986. This work was centered on
proving why the human brain is more powerful than a computer in being able to
learn from experience. They used several case examples, one of which was how
children learn the past tense of a verb. I will not bore you any longer with
the details of this work, but it is worth noting as it is landmark research in
the evolution of artificial intelligence modeling.
Backpropagation and Sports
With high powered computers now
available on a laptop, large statistical databases on sports stats can be saved
and manipulated to form new information. The BCS standings in College Football
are a perfect example of the direct output from a large pool of data. Of
course, too, it is not a perfect science, and the computer rankings associated
with the BCS poll have been largely scrutinized and their validity debated
endlessly. However, I have found that using backpropagation can determine the
outcome of sporting events generating enough winners to make it a viable
investment option. Sports betting has tremendous risk and in my opinion, must
be done in a highly disciplined and controlled manner. We all know how fast money can be lost simply betting on a ‘gut feel’ or ‘with the heart’. So, I
have always emphasized the need to recognize that none of us truly knows what
the exact final score will be for a game and no matter how strong the play
appears to be, it can lose against the spread. Wagering the same amount per *
unit and not allowing any emotion to enter into the decision making process is
essentially the only way to make money in sports wagering.
The simplest example of a model
would be to use two inputs to produce one output. For example, we could use
offensive and defensive scoring stats for two teams to predict which team will
win the game and by how much. However, this single-layered approach has obvious
shortcomings; these start with the location of the game, time of year, what
part of the season, common opponents and the like.
What the high powered computers
allow us to do that our brains cannot is to process large mountains of data.
So, in the case of baseball, 15 years of box scores and all the games stats can
be built into a large database. A computer programmed with a backpropagation
algorithm can quickly determine and
match historical situations that match the current game’s matchup. So, the
model can alert us to a possible money-making opportunity.
In any given neural network
simulator, there are hundreds of input variables. Adjusting these weights is
always necessary to get the most
reliable and meaningful results. For example, 30 years ago, starting pitchers
in baseball completed a large number of their games. In the modern game,
pitching has evolved and now we have stats named quality starts, and the relief
pitching has become a very important part of the game. So, the weightings for
bullpens is far greater now than it was in decades past.
This is just a brief start to artificial
intelligence modeling and neural network
simulators. In the weeks ahead, I will delve into more specific definitions and
provide enough information so that you can try to build one on your own to
augment your current handicapping methods.