Neural Networks are a form of artificial intelligence in which a computer simulates the way a human brain processes information. A while ago, expert systems were all the rage in various sectors of business as the stake through the heart of inefficiency. In my opinion, they're more problematic than practical, and when it comes to using an expert system as a trading aid I think reality has met its match.
But neural networks and expert systems are very different creatures. An expert system uses a previously defined set of software rules to guide someone through a decision making process; the rules can be updated and refined, but they do not modify themselves. A neural network is a style of program that specializes in pattern recognition, using interconnected software pieces (organized not unlike neurons in your brain, hence the name) to process information much more rapidly than a conventional sequential program would do. Once you have the program created, you can feed it data to give it examples of things you want to look for. If you're using a neural net to translate handwriting into a document, you'd give the net (or "train it with") samples of different styles of letters and handwriting. Once it has been trained, it will analyze new handwriting samples and attempt to find close matches to the examples it already knows about. In the absence of an exact match (which realistically wouldn't happen often with handwriting), the neural net will perform a most-probable analysis on what it does have in an attempt to guess the correct translation.
Neural nets are being used for a wide variety of applications these days, translating documents to an electronic form, matching fingerprints, playing chess, trading futures. Trading futures? .
It's the same idea. You create a neural net application that knows about market motion, price progressions, highs and lows and closes, etc. Then you feed it price and other relevant data from whatever sources you think are important. Given its preprogrammed understanding of what markets do and what the goal is, which is to take long or short positions in a market to generate a profit, it searches through the data to find patterns and inter-data relationships that appear to precede a profitable trade. Once the net has been trained, you then run it with real data. When a pattern or relationship emerges that it recognizes, it issues the appropriate trading signal based on what's expected to happen.
Is the net doing things that a human couldn't do? Not really; it's not that the computer is finding invisible relationships. It's that the computer is working in a heavily parallel way (i.e. working on several interrelated things at the same time) to compare and contrast a large volume of information far more quickly than a human would be able to. There's also the element of tedium in that same human endeavor; very few people would be excited about the task of looking at thousands of price bars in a handful of different markets for small, tradeable patterns, or for seeing if Corn is a leading indicator for the stock market across the last century. Frankly, it's a boring idea, and if you spent week after week in that pursuit, you'd get burned out and forget where you started.
Is it a holy grail, then? No, it's more of a holy pain in the butt. This isn't my area of expertise, but my understanding is that working with neural nets is a full time occupation. There are some systematic problems with using nets for trading. First of all, you have to find relevant sets of data to use. Remember, "garbage in, garbage out?" The net cannot magically distill good trading signals from nothing. It has to be fed datasets that actually do contain real patterns and relationships. How do you know? Experiment! If it were easy to grab a chart and pick out the patterns that worked, you wouldn't need your computer friend. I've heard of folks looking at price bar patterns, leading and lagging markets (e.g. seeing if Corn and Wheat have a signalling relationship across time), market behavior against natural cycles, behavior in different contract months during different times of the year, all kinds of things. It's not a trivial exercise to find something that works.
Next is the concept that markets are not static beasts. That is, they don't have the same patterns, the same personality all the time. They change, they shift, they have periods of unpredictability. Sometimes the changes last a few days, sometimes they behave a certain way for months or years. Interrelationships between markets and external information also change and shift over time. Bonds lead stocks. Bonds follow stocks. Bonds trade inverse to stocks. Take your pick; you can find any and all of those relationships. A neural net attempts to identify patterns, but it is searching for relationships in a chaotic environment. So once you find a fertile field of data, you have to continually re-train the net with new information, or you risk having it begin to make bad decisions because its assumptions are based on last quarter's market behavior profile. In a way, it's a near-ultimate form of curve fitting your system.
There are some other considerations that move the current neural net technology beyond the typical trader that I won't spend a lot of time on, such as cost, availability and complexity. Nobody's come out with The Norton Neural Net, with help text that's just as E-Z as the stuff that comes with Microsoft Office. That was a joke, and if anybody has a net that will help me search for information in this PC online help crap, do let me know... But it's not useless technology by any means, and maybe someday they'll have a NeuralStation program that you can use to do neural net analysis, for just 25 easy payments of £199 a month. Though I guess then you'd need another little hardware thingy to stick on your computer's parallel port, and how many of those do you really want to have?
The content of this site is copyright 2016 Financial Spread Betting Ltd. Please contact us if you wish to reproduce any of it.
