Creation of Advanced Trading Systems

Overseeing the asset management is being taken over by trading systems based on data analysis, artificial intelligence, machine-learning methods, predominantly the neural networks. Robotic trading based on algorithms will outperform man. Do you have the feeling you have already read this somewhere? Most likely yes, recently, there have been a lot of marketing materials and academic publications issued, focusing on modern technologies enforcement into financial assets management.

The articles have a very similar structure – initially, they list “breaking-through” discoveries in the field of information technologies and data analysis, not forgetting to name magical words such as “artificial intelligence”, “data analysis”, “neural networks” etc. The readers are further introduced to the advantages these technologies bring in comparison to the financial assets management by a man of flesh and bones. These articles and publication materials are very similar in trying to get out the most of the upsurge of modern technologies. Thereafter, in most cases, there is information about the author’s individual solution or service and, there, the magical words start disappearing or they appear vaguely and on a general level. When such promoted services or solutions are investigated into greater detail, more or less the same approaches which can be handled by a better-developed spreadsheet editor are seen. The message could be conveyed roughly like this: “There are brilliant tools which are gaining ground in the asset management, we deal with them… but eventually, we still do the techniques which can be managed by a “better excel” because it is just enough for us.” It is a practical example of what is nowadays called buzzword. Most of the current solutions offered in the market say: “We need to follow the technological development and we are unique in this owing to our approach… as a result, we fill our portfolio with shares of excellent technological companies (e.g. A, B, C) and the following year we exchange them for X, Y, Z.” The point is that from the global view, these changes are only cosmetic and the high dependency on the global market development remains.

Yes, there are companies that really work on this basis and use a “state-of-art” approach which they successfully apply in practice. The pace is set by companies such as the US Renaissance or Two Sigma. They, however, never spread these words and what they show are real results. Why would we be using the top-notch technologies if they did not deliver an added value in the form of better results? This mind-set is so intuitive and straightforward that it is partly in contradiction with our rooted habits – for example, the dependency of strategy profits on the global stock market development. If a trading strategy is correlated with the activity of underlying assets or with stock index S&P500, it is not good and this strategy must be immediately adjusted or is completely removed from further development – this fact is researched in the first set of tests in the new trading system which is being developed.

The point of such an approach is an effort to maintain high stability, not higher profitability. Mathematically speaking, the optimisation problem is not maximising the profit but minimising the risk. The pursuit of maximal profit is the subject of many statistical delusions and mathematical paradoxes that lurk at each of research steps from the historical bias, overfitting, high degree of correlation with the global financial market up to the uncontrollable degree of risk. In the moment of reaching risk control, the trading system gains a truly powerful tool in the form of stability. Therefore, profit is “only” the function of available capital and willingness to undertake risks.

The basic optimisation condition is the starting point for a long journey of research and development which results in a viable and autonomous trading system.

• The first building block of a trading system is the identification of a trading opportunity. A considerable part of investors I met lives in the belief that “the secret” is just this, however, the truth is that it is only the first step to create a stable and profitable system. Trading opportunities can be of many forms, from “getting” the difference between offer and demand, technical analysis which is virtually nothing but looking for systematic deviation in the price development up to fundamental analysis which seeks trading opportunities based on the development of determinants (for example the weather at crops and gas, political situation and reserves at oil, indebtedness at companies etc.). This trading opportunity is represented by a set of rules, for example for opening or closing a trade. The power or, if you want, the size of such a trading opportunity is measured by the so-called alpha. The moment when your trading rules generate positive alpha, you made the first step leading to the creation of a profitable trading system.
• The next step involves brute force computing. There is often another mistake (in this case made by a little bit more advanced creators of trading systems). One group prefers using “brute force computing” for seeking trading opportunities without the first step (trading opportunity identification). This effort leads to the creation of a rule set that historically led to profits but the absence of trading and economic sense is the cause for very little information value on the system functioning in the future. This phenomenon is called overfitting and it is reported to have low stability of real performance. On the other hand, the other group denies the use of so-called “yield curve fitting” because of the above-mentioned reason and the rules are set based on the experience. By taking this step, the author can avoid the danger of overfitting, however, it is highly likely that they are located at a point outside the local maximum or, in other words, they do not use optimal rules for the particular trading opportunity by which they do not fit within the range of effective boundary of a set of possible solutions. Using this brute force computing for optimising parameters (using genetic programming or so-called parameter sweeping, less brute option are neural networks) in the paradigm of the particular trading opportunity moves the trading system to the local maximum of a set of possible solutions.
• In the third step, the trading system is placed into a larger context in a way so that outputs (realised trades) are compared with the development of potential determinants or predictors (among these determinants can be for example volatility, time premium of futures contracts, risk premium of option chains but also statistical quantities such as skewness, sharpness, average, Hurst exponent measuring the persistence of trend etc.). After identifying statistically significant determinants, there will be created a scoring model that evaluates the dependency of the trading system success on the development of significant determinants.
• The output of the model is information gaining binary values (imagine a traffic light on a pedestrian crossing) that tells the trading system when it is the most convenient time to trade and when it is better to hold back (when there is a higher likelihood to succeed).
• In case of having more trading systems of this kind, the trader can proceed to the last step in which the involvement of individual trading systems is optimised in a way so that the trader can be in the range of effective boundary of the portfolio or so that the maximal yield is gained at the required level of risk. 

In the end, a person enters the whole process as a checker and “maintenance man” of such a trading system. No strategy can run in the real environment without maintenance and calibration. This role is still held by man as well as the role of a checker. State-of-art technologies work as a good technical tool for achieving all goals – these technologies are faster, more accurate, more computing-efficient but they can never solve everything for man, they cannot define a problem and they cannot solve a wrongly defined problem.

Michal Dufek, Head of Financial Research Software Development