Using Neural Networks to Track Stocks
Watch the bouncing ball, and profit from it! ... Sure!
----------------------------------------------------------------------------------
------------------------------------------------
In the technical approach to the stock market, traders depend on price action for buy/sell information. The claim is that whatever is known about equities has to be seen in their prices -- or more correctly, in the pattern of prices over time (the so-called time series) as it might appear in charts.
Different price patterns give different information (show different trading activities). So a pattern of stock accumulation, say, is different from a pattern of stock distribution. To time your trades, you must look for telltale signs of impending price movement. The aim is to detect and identify the patterns and act appropriately when the time is right. Easy to say!
Unfortunately, chart patterns don't usually live up to their billing and aren't easy to identify. Patterns often start out one way and gradually morph into something else. Also, patterns are seriously non-linear. The plan in this tutorial is to "let a neural network do the job for us," if we can only design and build one that can learn how! It's a nice idea, but, as you might expect, it's easier said than done!
---------------------------------------------
To design and train a network to track stock prices and predict where the price will be at some point in the future. We have to build and train a golden goose....
Hey! Why not! Let's go for it! A few golden eggs wouldn't hurt anybody. But how and where do we begin? Yes, that is a problem. So where do we start? ... Any ideas?
The thought I've had for a while is to compare price-tracking with tennis tracking to see if one can help the other. Both tasks are time-series in nature, one having to do with prices, the other dealing with something like successive trajectory positions of the ball. Also, associated with each trajectory is a projection into the future space-time of its respective environment -- i.e., the market and the tennis court.
Unfortunately, it isn't clear what characteristics are important for predicting future states. So it makes sense, initially, at least, to include as many characteristics as make sense as inputs for the respective neural networks. It would be up to the network, through many learning trials, to select the properties that give an advantage in the transformation of the network. The next step would then be to list the possible candidates for the job.
---------------------------------------------
What properties of the price trajectory are relevant when it comes to predicting future prices? If we knew this exactly, we probably wouldn't need a fancy network to help us. What we do know is that the problem is by no means easy, if it has a solution at all. If there's a formula that gives an answer, that formula is not only non-linear, it's probably severely non-linear. It can even be chaotic at times. But that's where the neural network shines, so there may be hope for a good outcome after all.
A great deal of work with a lot of experimenting has already been done to find useful market predictors -- models. These are the so-called market indicators. It's a shame to waste them, so we might draw from them to try to identify potentially significant properties. There are also more obvious items like the prices themselves, or the difference between successive prices, or the proportionate difference between successive prices, or whether the trend is up, or down.
Imagine, if you will, that a starting time for a price examination has been selected. From this point forward we might arbitrarily construct a price trajectory to a future date, which is the date for price prediction. This trajectory would be one of many possible trajectories that could be constructed. Now think of all the possibilities making up a space of trajectories, each with a certain probability of occurring, depending on the market forces at work. Viewing the problem this way, it would be up to the neural network to pick out the most likely one or two or three.
---------------------------------------------
It's possible that Elliott wave theory may work as the basis of identification for neural networks. Not sure how this goes, but it may be possible to define wave characteristics that can be used as inputs to the networks.
--------------------------------------------
In my attempt to formulate a model for the market, a number of difficulties have become evident. Some have to do with defining price relationships. Historical data failed to make clear how obviously related financial instruments are to be connected quantitatively. So I couldn't write the connecting formulas. It's possible, however, that a properly designed neural network can find connections and produce more precise relationships than can be constructed in a formal, a priori manner.
---------------------------------------