Almost anyone involved in online trading on Forex, stock, commodity, or cryptocurrency markets has heard about neural networks and their use in trading with robots, also known as Expert Advisors (EAs). So what exactly are neural networks, what do they have in common, and how do they differ from artificial intelligence? What are their advantages and disadvantages? And finally, can neural networks become a reliable tool for every trader, ensuring consistent profits?
A neural network is an algorithmic construct inspired by the structure and operational principles of the brains of living beings. It is designed to process data through complex networks of interconnected nodes that mimic neurons. Each artificial neuron in such a network can receive, process, and transmit signals to other neurons. As a result, collectively, they are capable of solving tasks ranging from the simplest to highly abstract ones.
The concept of artificial neurons was proposed as far back as 1943 by American scientists Warren Sturgis McCulloch and Walter Pitts, who created a mathematical model of a neuron. Warren McCulloch, born in 1898 and having obtained a medical degree from Yale University, USA, in 1927, conducted research in psychiatry and neurophysiology, specifically studying the nervous system. It was then that the scientist seriously became interested in the possibilities of artificial modelling of the human brain. Walter Pitts, 25 years his junior and self-taught in mathematics and neurophysiology, exhibited outstanding abilities from a young age.
In 1943, Pitts met McCulloch at the University of Chicago, and this meeting began their fruitful collaboration. That same year, they published "A Logical Calculus of the Ideas Immanent in Nervous Activity," laying the groundwork for theoretical research on artificial neural networks. In their paper, the researchers proposed a neuron model based on mathematical logic and demonstrated how networks of simple artificial neurons could perform complex computational tasks if their interconnections were properly organized. This discovery shed light on the potential use of artificial networks for modeling cognitive processes and creating intelligent machines.
The development of this technology went through several important stages, including the creation of the perceptron by Frank Rosenblatt in 1957. A perceptron is the simplest type of artificial neural network used for data classification (i.e., dividing data into groups). It consists of inputs, each with a specific weight (a number indicating the importance of the input), and a single output neuron that sums the input signals multiplied by their weights. If the sum exceeds a certain threshold, the perceptron activates and outputs one result; if not, it outputs another.
Another significant step towards the emergence of more complex neural networks was the development of the backpropagation error algorithm, which appeared in the 1970s and became a cornerstone in training multilayer neural networks. This algorithm is a method for training artificial neural networks where the correction of neuron weights is based on errors the network made in its predictions. Initially, the network makes a prediction, then compares it to the correct answer and calculates the error. Information about this error is then propagated back through the network, allowing it to learn and improve its predictions as it processes more data.
The works of McCulloch, Pitts, and their successors played a fundamental role in the development of artificial intelligence concepts. Their research anticipated and stimulated the creation of deep learning models, which are used today in various fields from automatic translation and image recognition to self-driving cars and process automation, and of course, in financial trading.
The use of neural networks in financial trading began in the 1980s when computer technologies had advanced enough to process large volumes of data and perform complex calculations. However, real interest in them emerged in the 1990s with the development of machine learning and increased computing power, allowing more effective use of artificial neural networks for market data analysis.
In the last decade of the 20th century, the idea of using neural networks in trading robots, otherwise known as expert advisors (EAs), for analysing market conditions, predicting price movements, and automatically executing trading operations arose. These neural networks are trained on historical data on prices, trading volumes, market indicators, and other technical analysis tools. They can recognize complex patterns and dependencies that are not always obvious, even to an experienced trader-analyst. After training, EAs are capable of independently making decisions about buying or selling financial instruments in real-time.
The most significant development in the use of neural networks for automated trading has occurred in the last 15-20 years. During this period, their effectiveness in various aspects has been proven. However, it has also become clear that, like any other technology, the use of neural networks has its disadvantages, problems, and limitations. These include, for example, the need for initial training of EAs: a lengthy, difficult, and patience-requiring process. In some cases, the neural network may also require retraining. This is necessary when it adapts too precisely to historical data and loses its ability to generalize. The need for constant updating of data and algorithms to adapt to changing market conditions, as well as difficulties in interpreting the results of the neural network's work, remain relevant.
In this context, as many experts believe, one of the main directions for the development of neuro-EAs is the creation of adaptive systems capable of autonomously adjusting their parameters in response to market changes. Moreover, work continues on improving machine learning algorithms, including deep neural networks, which allows for more accurate predictions and more effective trading. Analyzing a larger number of variables and their combinations can also help improve the predictive power of the systems.
A neural network and artificial intelligence (AI) are terms that are often used together but actually denote different concepts. The main differences between neural networks and artificial intelligence are:
– Application Area: Neural networks are just one of the tools used in artificial intelligence because they specialize in learning and processing data based on provided examples. Artificial intelligence, however, encompasses a broader range of technologies and methods that are not limited to machine learning or neural networks alone. AI aims to be maximally universal, allowing it to solve a wide variety of tasks in diverse areas. Neural networks are often limited to areas where they can be effectively trained based on provided data.
– Functionality: Artificial intelligence strives to fully mimic human intellect and is capable of performing complex tasks such as reasoning, self-improvement, learning, perception, and even social interaction. Neural networks, on the other hand, focus on specific data processing tasks, their classification, and subsequent prediction.
– Adaptability: Neural networks perform well within the specific tasks for which they have been trained. Their effectiveness can significantly decrease with a lack of data or changes in the conditions of their application. Artificial intelligence includes systems that can evolve and adapt to new tasks and conditions with minimal prior preparation.
– Technological Differences: Neural networks are specific in that they operate on the principle of data transmission through layers of neurons, each of which transforms the input data according to set weights and activation functions. Artificial intelligence covers a much wider range of technologies and is capable of performing more complex and diverse tasks. To achieve intellectual functionality, it can employ a much more varied spectrum of methods, including logical programming and optimization algorithms.
In conclusion, let's present several quotes reflecting the opinions of leading experts on the importance of integrating neural networks and AI into financial trading:
– Catherine Wood, CEO of ARK Invest: "The predictive power of neural networks in stock market trading is revolutionary, potentially enhancing returns through more precise timing and risk assessment."
– Andrew Ng, Co-founder of Google Brain: "Neural networks have the potential to make financial markets more efficient, transparent, and accessible, but we must be cautious about their broad impacts on the economy."
– Rana Foroohar, Global Business Columnist and Associate Editor at Financial Times: "As neural networks grow more sophisticated, they could drastically change the landscape of trading by offering deeper insights into both high frequency and long-term investment strategies."
– Ray Dalio, Founder of Bridgewater Associates: "Artificial Intelligence and neural networks represent the next frontier in finance. Their ability to digest and analyze vast amounts of data can fundamentally reshape how we understand market dynamics and asset management."
