Artificial Intelligence and Stock Market
Michael Henry,  Top-Down Market Research, LLC

The type of artificial intelligence software that we use at Top-Down Market Research, LLC is very similar to what is commonly known as a neural network. Our neural network codes have been written from scratch and include additional features not found in commercially available neural network software. Understandably, our neural network codes are proprietary and specific details of our codes cannot be revealed. However, some general background information on neural network technology is provided here.

Biological Basis for Neural Network
Neural networks were first developed as an outgrowth of the study of the human brain and nervous system. The brain and nervous system are  composed of cells called neurons. Neurons do not die and replace themselves like other cells in the body, which probably explains why our memories are retained. 

Estimates are that we have as many as 100 billion neurons in the human brain, each one connected to as many as 1000 neighboring neurons. Some of the electrical signals transmitted between neurons pass through signal   modifiers called synapses. Learning occurs as the synapses increase or decrease the signals passed between neurons. In this way, neurons and synapses work together in groups called networks.

Neural networks are capable of making sense out of complex patterns that would otherwise be unrecognizable. A good example of how this works is human vision. The retina in each eye has approximately 120 million light collecting cells. The cells convert the light energy to electrical impulses which are carried to the brain via the optic nerve.  The brain is given the task of decoding millions of electrical impulses so that they can be assembled into a picture that makes sense. Can you imagine someone giving you a jigsaw puzzle with millions of pieces and asking you to put it together in a fraction of a second?

Neural Networks Change and Learn
Neural networks learn cause and effect relationships. Remember the first time you tried something new? For instance, imagine you are learning to play tennis for the first time. You have never even picked up a tennis racket before. You don’t know how to angle the racket or how hard to swing at the ball to keep it in the court. Furthermore, you don't know what the trajectory of the ball will look like as it crosses over the net into your side of the court. How will the ball bounce? Will the spin and speed of the ball effect it’s bounce? Will the wind have a significant effect on the trajectory of the ball? Where do you need to run on the court to meet up with the ball at exactly the right time? Needless to say, there are a lot of things that must be learned to play tennis effectively.

So, ask yourself  this question: Before you go out on the court to hit the ball, are you going to sit down and run calculations based on the laws of physics to account for all of these factors? No, of course not! You just do it! You use what little information you already have about the sport and swing at the ball! Perhaps the first time you swing at the ball, your racket is too high and you completely miss it. You realize your error and make a correction. The next time you swing a little lower and make contact, but because of the angle of the racket, the ball goes over the fence. Over time, through trial and error, you continue to make corrections until you find that the ball starts to go where you were aiming.

Neural Networks on Computers
That's sort of the way a neural network computer program learns too. In the case of predicting the direction of the stock market, a neural network looks at a large amount of historical economic information and attempts to make a prediction as to what will happen next. It doesn't run any complex supply and demand calculations. It just does it! It then compares it's prediction with what really happened and makes adjustments to compensate for it's error. Essentially, the neural network “lives” through history time after time until it becomes proficient at predicting the future. In essence, the program has learned what factors have significant effects on the future prices of stocks. Some of the factors that affect future stock prices are hidden and are not easily recognized. But they exist nonetheless. The neural network program learns what the cause and effect relationships are and is also able to quantify how much of an effect each factor will likely have on a stock's price.

The Wisdom Of A Lifetime
Have you ever wished that you could have all of the investing experience of a 70 year old veteran of the stock market? Can you imagine how valuable it would be to have already made every investing mistake in the book and to have learned how to avoid repeating them? Since you would have seen it all before, you wouldn't be so easily fooled by every breaking news story and each twist and turn of the market.

Think of how many times you would have observed the business cycle - how the stock market explodes as the economy accelerates out of the bottom of a recession; interest rates are lowered, jobs are created, and consumers begin to spend again - then how the stock market loses momentum as the economic expansion slows; consumers accumulate debt and spend less, business inventories grow, and interest rates start to rise in an effort to quell inflation - and then how the stock market, seemingly without warning, plunges as the economy slips back into another recession; jobs disappear, spending slows to a crawl, and bankruptcies accelerate. Can you imagine how much wiser you would be having observed this cycle a dozen times? You would have a much better understanding of the inter-relational aspects of world events. You would be able to focus on the truly important long term events and would be able to ignore the unimportant short term fluctuations of the market.

Not Quite As Good As It Sounds
Unfortunately, there are a number of reasons why all this experience probably wouldn’t help much in real life. First, by the time someone has the chance to accumulate a lifetime of investing  experience, there’s not enough time left to make a substantial difference in one’s wealth. Second, although the human brain is a marvelous creation, it is nonetheless fallible. The more time that goes by, the more things we forget. Some of our most valuable lessons in life fade away into the past and  are not easily recalled to assist us in our present circumstances.

Furthermore, there are multitudes of hidden relationships in our modern economy that are extremely complex and hard for even the human mind to understand. The stock market, interest rates, money flows, demographics, and an endless number of economic variables are all intertwined in an extremely complex system of cause and effect.

For example, experts on the financial markets agree that as interest rates or inflation rates go down, stocks tend to do well. Why? Because during these times investments such as money market accounts, CD’s, real estate and gold become low yielding assets. So investors react by selling those and buying other investments with greater potential, such as stocks or mutual funds. That sounds good in theory, but is this always the case? No, not always. As the chart below shows, over the last 60 years there have been many exceptions to the rule.

(The Dow Jones 20 Bond Index is inversely related to long term interest rates)

Looking closely, you will notice that there have been many periods during which the well known "rules of thumb" do not apply. In fact, the stock market is interwoven with the global economy in such a way that any number of events can cause ripple effects that find their way into stock market prices. Most of these economic inter-relationships are hidden and are not easily observed for processing by the human mind. But computers that are trained to look for them can find them.

Gaining A Lifetime Of Experience - The Quick Way
Today, with the advent of computer technologies such as artificial intelligence, all of these problems can be overcome. It is now possible to acquire decades of wisdom in a relatively short time using a neural network computer program. Once a neural network is provided with historical data, it can be programmed to make stock market predictions as it steps through time. As it steps through time it measures the errors of its predictions and makes corrections to the weights (or synapses) in the neural network in order to reduce the errors. Essentially, the neural network "lives" through history over and over again until it is able to effectively predict the future. 

So How Well Does Our Neural Network Predict The Market?
To date, no human or computer program has yet been able to predict the future of stock prices with certainty. There are simply too many uncontrollable variables in the world. Events such as natural disasters, terrorist attacks, and financial meltdowns in foreign countries are just a few of the types of global occurrences that are for all practical purposes unpredictable. Yet these events can have dramatic effects on the US stock market. As an investor, you simply must accept these risks.

That being understood, the goal of our forecasting systems has always been to use as much pertinent data as possible from the global economic environment so that the neural network's predictions will be right most of the time. Therefore, keep in mind that each individual 12 month prediction should not be considered as a definite target to be hit. Instead, the 12 month predictions are used to rank the investment choices that are available so that we will most likely be in the best industries and the best stocks at the right time.

Since we do not know what future stock prices will be, there is only one way to measure a system's ability to forecast the future - past performance. To date, the historical performance of the Top-Down Market Research, LLC (TDMR) neural network models have been fantastic. Through rigorous back testing, the models have proven themselves to be robust without the benefits of curve fitting.

Our Computer Models
The neural network computer models that we use took many years to develop. A large part of the work involved the tedious creation of database files that are used to train the neural network computer programs. Much of the data was not available in a digital format and had to be input by keyboard. Some of the information goes back to the  1800's, but the most useful data covers the post World War II era. Three different databases are used - one for macroeconomic indicators, one for overall stock market indicators, and one for the stocks of individual companies.

The macroeconomic database includes a subset of various measures of interest rates, inflation, currency exchange rates, money supply, Federal Reserve data, production statistics, demographics, sales, orders, inventories, U.S. trade, commodity prices, employment, spending, debt, income, and various data on foreign countries.

The overall stock market database includes a subset of index prices, volume, short interest, short interest ratio, margin account credits and debts, earnings, dividend yields, mutual fund flows, and industry index prices.

The individual stock database includes 132 different parameters derived from the balance sheet and profit and loss statements of approximately 9000 companies. Measurements of insider trading are also included.