Neural networks in trading (FinTech)

Automation in the world takes place in all areas of life and business. Many processes have long ceased to depend directly on the manual labor of people and have completely moved into the field of machine control.

It was the turn of finance. How do neural networks imitate the functioning of the human brain and solve such complex problems? Or do they not cope with this?

The peculiarity is that the market of securities and oil is affected not only by the actual figures, on which the artificial neural network is based. Since it works with a huge amount of data regarding physical indicators and risks, in reality, other factors that are not related to numbers must be taken into account.

Neural networks are constantly learning and getting smarter, they are studying what was in the past in order to predict the future. Network training can be long enough so that they begin to independently find and make adjustments to patterns (trading patterns).
The acquired skills help neural networks to build on them, as from the main database, and compare them with each other. Thus they give out their forecast.

It is the constant training of neural networks that is their main advantage, since when they receive new data they use old forecasts, analyze risks and quickly run more information than a whole team of specialists would do in a long time.

The combination of technical and fundamental data is the main and essential plus for the use of neural networks in trading.

The main task of artificial neural networks is to predict. As already mentioned above, in order to make a qualitative forecast, the algorithms rely on a large amount of data and forecasts already existing in the history.

However, very often, numbers alone may not be enough. Making accurate forecasts of exchange rates or stock prices is difficult, and having only numbers does not always work. This is regularly emphasized by all experienced traders using neural networks.

Relatively correct judgments of the algorithm, of course, can be obtained if it has historical data on past correct network forecasts.

The accuracy of neural network prediction in trading is about 50-60%, which is not so much. That is, errors happen more often than we would like in an automated process.

It turns out that it is better not to use artificial neural networks at all for trading electronic securities and currency?

Experts recommend attributing neural networks to algorithms to search for patterns taking into account history, that is, use them as a method of technical analysis, and not rely completely on work.

Stuart Reed, an NMRQL hedge fund analyst, suggested that the tarnished reputation of neural networks was not the fault of developers or traders trying to work with them.
The problem is a misunderstanding of the operation of such systems.

Information is processed, analyzed, compared with old forecasts and then they begin to trade or make their forecasts.

Oddly enough, the lack of a neural network in this area is emotions. More precisely, the complete absence of emotions in the algorithms. Some believe the opposite, that this is an advantage.

But markets do not work without people, and people are impossible without emotions. So any emotional event can both raise the value of stocks or currencies, and bring them down at once.

Neural networks are created in the form of algorithms that do not understand how the illness of a tycoon or the owner of a corporation can affect the markets.

They are not ready for such a sharp surge of volatility, since they do not take into account human factors.

In order for the program to make a profit, it is necessary to constantly monitor, train and adjust it when errors occur. Among traders there are those who use it quite actively, but with regular monitoring.

It is important to remember that an artificial neural network is not a model of the human brain in its direct sense. It is like a multi-storey building, although it is built like a beehive, but it is not.

So a neural network is built as the relationship of neurons of the human brain with similar structural elements, but it can’t even replace it closely.

Reasoning on the work of the neural network in the financial sphere according to the principle of the human brain is rather nothing more than a delusion at the level of science fiction and rebellion of machines.

Modern artificial neural networks are best attributed to statistical methods, since they are more likely to correspond to regression and curve lines.

For example, a curve bearing a name is an approximation function. By the way, neural networks are often used to approximate complex mathematical functions.

It is also a mistake to consider a neural network a simplified form of statistics. The structure of the networks is such that they are layers of nodes interconnected. The nodes individually are perceptrons, similar to multiple linear regression.

Layers are of three types: input and output signals, hidden layers. The first layer uses patterns (patterns) from the received data or input information, and the second maintains a classification list or output signals in accordance with the scheme.

Hidden layers are used to adjust the weight of the input data. They regulate until they minimize the risk of error.

So, we summarize: using a neural network to trade in markets with lazy traders will not work, since it is not a simplified form of statistics or a technical tool.

A neural network for trading is not only primitive and does not necessarily consist of several levels of perceptron layers, as described above. And to be much more interesting and different in architecture.

● recurrent neural network - runs on serial data;
● Boltzmann neural network - can self-learn internal concepts and solve complex problems in combinatorics;
● deep neural network - perfectly solves problems in voice and image recognition;
● adaptive neural - is able to adapt to dynamic changes in financial markets.

Traders recommend trying more than one network in practice, but consider and work on several options in order to determine and choose the optimal one for the desired tasks.

For neural networks in trading, three basic learning strategies are used, each with its own data:

1. controlled;
2. uncontrolled;
3. reinforced.

More details on them need a separate article. But one of the main problems, the reasons for a non-working neural network, may be the poor preparation of this data before downloading to the system, and not the very amount of information, which can also be too much.

Redundant information should be deleted before it gets to the network. The more variables, the longer the network thinks.

Since the neural network is still just a machine, not a person, it only processes information, and the network does not know how to fully think.