-
Investments based on mathematics
How Quantum Capital invests using machine learning — in an interview with the company's co-founder Akhmet Byashimov
It's been five months since our last meeting. How did you get through this difficult period?
is a really difficult year. The summer was very good for the markets, but since September, volatility has increased due to overheating in technology stocks, another wave of coronavirus, and the constantly delayed fiscal stimulus in the United States. Because of this, it is difficult to enter stocks now, but the scarier it is to buy, the more correct the moment of entry is, historically it has happened that way.
Since June, we have made a lot of money for our clients, which means more trust and confidence in our strategy. Quantum Capital has doubled in size in five months, we have expanded the team and implemented machine learning in the investment process.There is a lot of hype around artificial intelligence right now. Many people say that they are introducing AI into their processes. What do you mean by machine learning?
— There is a lot of confusion with the terminology of the actual. At the academy, artificial intelligence is any program that can make conclusions or predictions based on data, no matter in what way – deep neural networks or if—then rules. And machine learning is a subsection of artificial intelligence, in which the algorithm must independently find statistical dependencies in the data. But in everyday life, people usually mean reinforcement learning, computer vision, and text comprehension by artificial intelligence.
For example, the Deep Blue chess computer is part of "artificial intelligence", but not "machine learning", since its developers wrote down the rules in it, and it was not he who found the winning strategies. Most algorithmic trading strategies also work. Developers automate the conditions for buying or selling stocks in response to an event.
As a result, companies can actually use the simplest linear regressions or a preset set of "if — then" rules and at the same time honestly declare the use of artificial intelligence. But consumers/investors remain deluded.
By machine learning, we mean precisely the property of our algorithms to learn independently from data, but we do not use the term "artificial intelligence" in order not to create confusion.
How is machine learning applied in the investment field?
— In our industry, machine learning has a nuance that distinguishes it from the general science of data analytics. So, the rapid pace of development of machine learning is primarily due to the fact that it is mainly open-source projects and research. In other words, almost all the latest developments are publicly available and everyone can contribute to this science. But machine learning in the investment field is an exception in this regard – leaders cannot share their profitable developments, as this essentially dissolves the opportunity to make money. In investment machine learning, everyone is on their own and each company is, in fact, a research center that tests different theories in order to find the very gold mine.
It is also worth noting that financial markets are efficient enough to make the process of creating a machine learning algorithm a difficult task. There are many scientific papers in this field that claim to create a profitable model, but there are also many refuting articles. If you study the methodology of the articles with sufficient care, in many cases you can find flaws that make the work inapplicable or fundamentally wrong. There is also a lot of work covering less accessible markets such as Brazil, Turkey, India, and so on. This means that the scientific literature is only a starting point in our work and real algorithms need to be created independently.How does Quantum Capital solve this problem?
For example, the Deep Blue chess computer is part of "artificial intelligence", but not "machine learning", since its developers wrote down the rules in it, and it was not he who found the winning strategies. Most algorithmic trading strategies also work. Developers automate the conditions for buying or selling stocks in response to an event.
As a result, companies can actually use the simplest linear regressions or a preset set of "if — then" rules and at the same time honestly declare the use of artificial intelligence. But consumers/investors remain deluded.
By machine learning, we mean precisely the property of our algorithms to learn independently from data, but we do not use the term "artificial intelligence" in order not to create confusion.
And what are your results?
— We started by trying to use recurrent neural networks to predict the financial performance of companies, which in turn can predict stock prices. Unfortunately, the model did not produce adequate results due to the high probability of overfitting in long-term forecasting. We also tested many technical analysis indicators for return, giving machine learning models the opportunity to find the best parameters for each indicator. The result also did not meet expectations, as the market adjusts quickly and indicators lose their predictive ability.
In our case, finding the right formulation of the problem turned out to be the most important part of the solution, and we came to search for companies with a high probability of moderate growth in the short term.We have collected information on more than 4,000 companies since 2000 and have trained an algorithm that could accurately detect such events. Based on the test data, the median returns of our models were 2.7-4.1% over the 2-3 weeks during which the position was open. This does not mean that the model allows you to double your capital in a year, but it can improve the performance and diversification of a standard portfolio.
How much can historical results be trusted?
is the most critical issue in our work. The highest priority for us is the correct methodology for testing on test data (backtest). It is better to have more modest but honest results and be confident in them than to think that everything works because of a wrong methodology. And this is the most common mistake in algorithmic trading and even in scientific papers.
For example, a very common flaw is testing on training data. That is, the model finds the best patterns in the training data and, of course, the result will be great there. But this does not mean that it will be the same in the future. A less obvious "systematic survivor error" can also be noted. For example, the current members of the S&P 500 index have not always been in the index and it is not certain that they will remain in it. Doing testing only on these companies means knowing in advance who will be among the largest 500 companies in the United States. Such small things can also affect the reliability of the results.
To be honest, we devoted much more time to the quality of the test than to optimizing the algorithm. This allows us to trust our tests and put the model into practice. We have already closed 12 real deals, which have brought in an average of 3.5%, despite the negative dynamics of the markets in recent weeks.
Do you have any competitors in this field?
In Kazakhstan, we are the only asset management company that already manages assets using machine learning. There are many reasons for this: excessive skepticism, a lack of understanding of data analytics, and a shortage of suitable personnel. We ourselves see competitors outside of Kazakhstan, for example, "quantum" hedge funds. But we also have an advantage over them – our investors' assets are not in the same offshore account. Each of our clients has his own brokerage account opened with the largest banks in the USA.
What are your plans for the future?
— We already have two investment strategies with different levels of risk. Both are showing good results this year (+34.54% and +20.6% for a more or less aggressive strategy), because we have protected ourselves from the coronavirus crisis and entered the market on time. A small proportion of assets are already being invested according to the forecasts of the model, and we plan to launch another strategy with a focus on machine learning.