Artificial intelligence is increasingly demonstrating its capabilities across various domains, from text generation to programming. One of its primary applications is data analysis, including the handling of vast datasets. Therefore, the prospect of using AI by financial analysts, investment advisors, traders, and investors seems promising. Will neural networks for investing become widely accessible, and is it worth trusting artificial intelligence in developing strategies and making trading decisions?
In this article, you will learn that:
- Neural networks show promise in investment applications today, despite limitations in predicting market behavior.
- While neural networks may not forecast market movements accurately, they can effectively identify trends and flat areas in trading scenarios.
- Teaching AI to trade is relatively simpler than predicting market trends due to the formalizable nature of trading decisions.
- Neural networks can assist in fundamental analysis across various levels, aiding in trend forecasting and asset selection for portfolios.
- Although AI can streamline data analysis for financial analysts and optimize investment strategies, it is not yet ready to replace human expertise in these domains.
Artificial intelligence in investing today
The successes of neural networks in various domains have been widely acknowledged. ChatGPT, for example, has become one of the most successful (or at least widely recognized) implementations of artificial intelligence, noted for:
- Writing books (available for purchase on Amazon)
- Passing exams, including those in complex fields like medicine
- Assisting in thesis projects
- Even distorting scientifically proven facts.
Given these capabilities, the desire to leverage AI in investing and trading appears natural. Especially since the process of training neural networks boils down to analyzing large datasets. Therefore, by incorporating price charts, macroeconomic indicators, financial and political news, company reports, and press releases into this process, there is theoretically a chance to achieve high-probability forecasting of future market developments.
Theoretically, there is a chance to achieve highly accurate forecasting of future events with AI. Several experiments already showcase this potential:
- At Seoul National University, ChatGPT was tasked with forming portfolios from a selection of stocks, cryptocurrencies, currency pairs, bonds, and commodities. The AI consistently outperformed random combinations, demonstrating effective diversification principles.
- According to the Financial Times, ChatGPT was tasked with creating a portfolio of 30 or more US stocks based on strategies used by leading investment funds. The AI’s portfolio of 38 securities showed a 4.9% growth over 8 weeks, outperforming top fund portfolios.
- Bloomberg announced plans to develop BloombergGPT, an AI network for generating analytical reports and datasets. The AI will be trained on decades of agency data and datasets from sources like The Pile, Wikipedia, and C4.
However, experimentation extends beyond ChatGPT, with various groups and investors creating neural networks with seemingly phenomenal results. While inspiring, investors should exercise caution and avoid rushing into decisions based solely on AI recommendations.
What neural networks are really capable of?
The emergence of ChatGPT and experiments with it are far from the first experience of applying neural networks in financial markets. As early as the early 2000s, many traders became familiar with the Trading Solutions package from Neuro Solution, which allowed building complex neural networks, training them on historical data, and obtaining a library for implementing trading robots.
With the advent of high-level “wrappers” for neural networks (such as Keras or FANN), development has become accessible to almost everyone, even those without knowledge of architecture, computational ratios, learning algorithms, etc. By the way, it was the terminology of these packages that played a cruel joke on a huge audience of traders and investors. Indeed, in these, the outputs of the neural network are called Predictions. Therefore, the vast majority of modern users are absolutely convinced that if historical data is “fed” into the neural network (provided at its inputs), then as a result of training, it will be able to predict the behavior of an asset in the current situation.
What neural networks cannot do?
A neural network is not a predictor, not an oracle; its essence is approximation. This means that based on the analysis of past data, it can approximate (i.e., calculate with some error) future outcomes. However, this only works if these outcomes can be approximated at all.
An interesting fact! To justify the use of neural networks in investments and trading, an example of using them for weather forecasting is often cited. Interestingly, it really works with weather forecasting! However, those who make such analogies rarely consider that the conditions for predicting these phenomena are fundamentally different.
In general, the theory suggests that any system can be modeled as long as it is self-sufficient. This means that its description (or data for analysis) fully represents all possible information about the system. As a result, it no longer needs other external data apart from the initial state. In this case, a neural network or artificial intelligence can provide accurate forecasts through the approximation of previous states.
In the case of analyzing and forecasting financial markets, the system is far from being self-sufficient. Quotations, no matter what historical interval they are collected from, are merely the result of numerous events that are not accounted for in this statistics. Even if the history of quotes is supplemented with a multitude of data from the economic, financial, political, and social spheres, the system still does not become self-sufficient.
The main component that must be introduced into the model in this system is human beings. It is their reaction to market behavior that determines the future behavior of the market (i.e., it is both the cause and the effect). This must be taken into account along with external influences (such as news) that are causes of changes in the system’s behavior. However, it will be necessary to model not just individual people, but the entire market participants, and not only the market (for example, the management of a company’s impact on the movement of its stock price).
Thus, AI will be able to forecast market behavior only if it can approximate the behavior of its participants and other people. The task does not seem unsolvable, but at the current stage, it will require a multiple increase in the flow of processed data and, consequently, the computational power of the system. In fact, this means that a neural network is currently unable to perform the direct task of forecasting.
Are there prospects for neural networks in investing today?
However, for enthusiasts of AI usage, it should be noted that things are not as bad. Although a neural network cannot yet predict market behavior, it is quite capable of solving other tasks.
For example, it is entirely possible to train a neural network to identify trends and flat areas on the market with a high degree of probability. In fact, solving just this task would already enable the use of AI in trading systems, with a sufficiently high level of profitability.
Experts working on neural networks in trading and investing say that teaching artificial intelligence to trade is much easier than predicting the market. Indeed, in most cases, making trading decisions (entering the market, taking profit, limiting losses) can be formalized into a limited set of rules. Even if a human cannot formulate them, this task is exactly what a neural network is for. It is enough to take historical data for training (possibly with additions of external factors, such as news) and precisely formulate the conditions for profit extraction and loss limitation. As a result, a quite decent trading robot will be obtained. Many such robots, although not widely publicized, already work quite successfully for the benefit of investors in large hedge funds and other market participants of the same scale.
Another possible application of neural networks in investments today is conducting fundamental analysis at all levels. In this regard, artificial intelligence can quite cope with the tasks of forecasting trends in the global economy and the economy of a country, as well as the situation in the industry. It will also help select promising companies for stock purchase, assess the prospects of including other assets in the portfolio, and ultimately ensure quality diversification with a specified level of risk and projected return over a certain interval (an example from one of the experiments mentioned above).
Artificial intelligence can also become an effective assistant to financial analysts, significantly simplifying data analysis for them and monitoring views expressed in other sources on the situation. It will also assist investment advisors by performing routine tasks related to optimizing strategies and their analysis. However, it is not yet ready to replace humans in these areas of work and is unlikely to be ready anytime soon.