Get best of the week

Neural networks in trading. Bury early

This article is my extended response to a recent publication, “Do Neural Networks Dream of Electric Money?” , in which the author verbose and in detail explains why neural networks just cannot work in trading and why price prediction is impossible.

Before justifying our disagreement with such a position, let's touch on the theory and techniques that are used in trading. The basis of most price charts is the so-called “candle”. This is the period of price averaging, within the period we ignore fluctuations in quotes, and leave only the minimum and maximum values, as well as quotes for the beginning of the candle (opening) and the end (closing). Candles can be from 1 minute to 1 year. Color the candle green if the price has moved up, red if it has moved down. As a result, we get a simplified, readable schedule, and most importantly, informative.

There is such a profession - a trader. This specialist who has studied his craft for a long time has extensive experience and, as a result, after analyzing a certain number of previous candles, he can predict the price movement on the next candle, i.e. predict the future. Of course, he does not just look at the chart, but uses additional tools called “indicators”. There is nothing magical in indicators, they formalize and mathematically describe the same experience of traders gained by generations. Working with indicators is called “technical analysis”. Unlike other types of analysis, technical analysis works only with the schedule, no news and other things.

Already clear what I’m leading to? I personally know of several traders who, using only technical analysis, have been successfully trading on the exchange for years and even steadily earning their bread. The obvious conclusion from all this is that the price of an asset in the future is somehow connected with the history of previous prices and this connection is sufficient for a living person to see and use this knowledge.

So why can't neural networks? A cat is distinguished from a dog, but here they cannot. It seems everything is obvious, it should work, but it does not work. From this place I will give my explanation why it does not work, or rather, it does not work for most.

Since we remembered the classic "Hello world" in neural networks - to distinguish a cat from a dog in photographs, let's remember what happens there. Neural networks for training show, for example, 10'000 pictures which in different situations depict a dog, then also with a cat. For each picture, the correct answer is given to who is on it. The neural network attentively looks at all this many times and establishes in its head certain rules by which it will be able to correctly answer the question “Is it a cat or a dog?” In the future. And this circuit works. Success rate 99.9%, bingo! So we apply this in trading.

Let's show the neural networks screenshots of graphs and give the correct answer where the price then went, she will learn this way and everything will be OK, she works with a cat. This is an example of entering the topic of a standard average researcher. And what does he get on the way out? Nothing ... Neural network is not learning. But our research is not simple and does not immediately give up: "We must submit the correct input!" and cycles of “right data” begin in the form of endless variations of ingenious vectors. And now the process has begun ... To understand when our researcher will get tired and write an article about the fact that it is impossible to train neural networks, you need to take the average value of the researcher’s diligence and multiply by the number of hours from one disappointment to another.  

But what is the correct answer, why not learn?
In fact, under the “researcher” I described myself, but only I was lucky, enough zeal to reach the first positive results. And here is my, purely objective, possibly incorrect, explanation of the problem.

Yes, quotes are chaos, but not 100%. In about 2% of cases, the next candle with a probability of about 70% is connected with the previous story. In fact, approximately the same principle is exploited by indicators, only in them it is called a “pattern” which, just like that, happens with approximately such a probability and the probability of working out for him is also not 100%. The values ​​of 2% and 70% are what I got today. I am sure that with proper training of the neural network this connection is much greater. And the approach to training with both cats and dogs does not work for a very simple reason. Showing the neural networks of the graphs and giving the correct answer, in fact, we do not show her a conditional cat or dog, but show clouds, butterflies, zodiac signs and only two percent of what is needed, i.e. 98% of our data is chaos.

It remains to understand how to catch these treasured 2% and only then make trading decisions on them. The option “we train showing only indicators” does not work, at least I didn’t succeed. As a result, I got the first results after 100500 selections of input parameters plus the correct analysis of what the network produces. A more detailed explanation is technically complicated and not for this article, here I just tried to argue with logical reasoning that the neural networks and trading are incompatible.

More articles:


All Articles