Is it Really Possible to Predict the Future?

The best known examples of where Artificial Intelligence predicts the future is Weather Forecasts. Most people don't even realize this but meteorologists have been using Artificial Intelligence to assist in weather forecasts for quite a while now. Are the predictions accurate? Not usually 100%, but moderately close most of the time*. If you could be as accurate as a good weather forecaster in the Stock Market, would it make a difference? You bet it would!

Is this technology sound?

Yes! The first thing is to unmask what a neural net is and explain how it works, since many people perceive this technology as nothing more than smoke and mirrors mostly because they do not fully understand it. For those that are math inclined this will be easier to understand. In short, a neural net is a multi-dimensional polynomial that trains (learns) on data that you want it to simulate (predict) such that it produces outputs that model the data it was trained on.

Neural nets work by providing them inputs (numerical values of some sort as can be seen by the smaller circles in the top layer) which are then multiplied by their connections (called weights - the part that is adjusted by training) and summed using various transfer functions by the receiving node (the circles on the second layer called a hidden layer) which are then multiplied by the next set of connections (again by another transfer function) to a final output (the circle on the bottom layer). This is what a neural net looks like visually:

In the example above there are 20 Inputs, 5 Hidden Nodes and 1 Output. In the first layer there are 100 connections that go to the second/hidden layer (20x5). In the second/hidden layer there are 5 connections that go to the output layer. So this system has a total of 105 connections or weights that are adjusted to produce a final value (this is how/why neural nets get their power). Let's give a super simple example of what is occurring by focusing just on the Hidden Layer to the Output Layer process. Lets say there are just two Hidden Layer Nodes and their values are [-0.60] and [0.25] (these values would be produced by the process from the Input Layer to the Hidden Layer). The connections (weights) are [-0.16] and [-0.45] respectively. The computation would then multiply the hidden nodes by their weights as follows:

-0.60 (Hidden Node 1) x -0.16 (Weights) = 0.0960
0.25 (Hidden Node 2) x -0.45 (Weights) = -0.1125

You would then sum the outputs via some transfer function. In this example it will be just a standard summation.

0.0960 (First Output) + -0.1125 (Second Output) = -0.0165 (Final Output)

Has this technology been proven?

Yes again. Aside from weather forecasts there are also many other uses for neural nets. Speech recognition, loan approvals, credit card transaction approvals, missile tracking, radar detection, dynamic noise filtering for cell phones are all specific examples of AI use. But it is more widespread then that with entire industries and fields benefiting from this technology. This includes the financial industry (of course), astronomy, mechanical engineers, banking, PC games, automotive, even the medical industry has been coming up with new medications developed by AI systems.

How does it learn?

From the example above we now understand how a neural net produces outputs and can now explain how it learns from them. When you train a neural net you provide it with both the inputs and the expected outputs (for normal training). In the detailed example above the final output was [-0.0165] but lets say the expected output was suppose to be [0.1250] which is quite a ways off from [-0.0165], [0.1415] to be exact. The neural net takes this difference (called the error) and sends it backwards through the net (i.e. from the output back to the individual inputs). Along the way it slightly adjusts each of the weights such that the next time the neural net calculates those same inputs, its output will be closer to what was expected. By repeating this learning process over and over again with all points from the training data, the neural net eventually learns how to model the data.

Why does modeling the data let us know what will happen in the future?

Now we all know that past performance is not an indicator of future results, yet no experienced investor I know will buy a stock until they've looked over the stock's charts and other bits of information at least to some degree especially since all this information is so readily available on the internet now. This information is used by the investor in hopes of achieving some level of insight needed to make a final decision. Since it is the investors themselves that are the driving factor behind the stock market, it is exactly this reason that information in past data can sometimes (not all the time) be used to potentially predict future direction and momentum of a stock (i.e. the neural net in essence is doing something similar to what a regular investor would do themselves). For example, if every quarter a company releases earnings announcements that have been favorable for over 4 years now making the stock jump in price both before the announcement in anticipation as well as after the announcement in confirmation and then soon fall back down to normal levels within a month afterwards, ANNI could pick this up and with a high chance of success, predict both the increase as well as the decrease in value. This of course is a very simple example especially since one could say "I could have figured that out myself just by looking at the charts" which is actually my point exactly (i.e. sometimes people do use past information to make future decisions even though past information is suppose to have no indication of future results). Since neural nets are far more powerful then just being able to find the strongest signal (such as the example above), they can also pick up weaker signals as well...all of which can provide profits throughout the year and not just at certain times of the year when things are obvious to everyone.

Just remember that neural nets are not 100% accurate and they can be wrong at times, especially more so with volatile stocks that have no chance of repeating any past moves with consistency or are temporarily in a volatile state because of some major event or change in the company that just occurred (the critical part of when a neural net can or cannot predict a stocks movement with any degree of accuracy and something many other similar programs do not properly address causing failure right from the beginning). ANNI has tools that will help you find suitable stocks to use and safety checks that will tell you if you are using a bad stock inside of ANNI after the fact in case something has changed since creating the file (something I believe is not currently in any other similar program). An investor should only use ANNI's outputs to assist them with investing and not to tell them how and when to invest (a critical mistake any inexperienced investor can make with any investment program). Using ANNI as an investment tool (rather then the hopes of it being a crystal ball) will help provide the highest level of success possible in the stock market.

<< Back | Tour Topics | Next >>

* Please be advised that there is no such thing as a crystal ball when it comes to predicting the future and any program that says it can do so 100% of the time is simply taking your money. ANNI utilizes many resources and technologies to assure the highest possible chance of accuracy but it will be wrong at times. Learning how to understand and work with ANNI's outputs will provide the best performance and profits possible which is the way ANNI was designed to be used in the first place.