By Andreas Roell
When Google’s DeepMind beat the Go world champion in 2017, the AI world celebrated this accomplishment as a major milestone in the technical advancements of the industry. Rightfully so, because defeating the number one Go player at the time was considered beyond the reach of even the most sophisticated computer programs. The ancient board game, invented over 2,500 years ago, is famously complex with more possible configurations than atoms in the observable universe.
To explain briefly, Google’s AlphaGo used a combination of deep learning and neural networks in such an advanced manner, at that point in time, the algorithms continually reinforced themselves to improve by playing millions of games against variations of itself. This trained a “policy” network to help the models predict the next moves, which in turn created a “value” network to evaluate proposed positions. AlphaGo then looked at the possible moves and permutations to select the move the network deemed most likely to succeed.
When studying the moves of AlphaGo during its game against world champion Ke Jie, it is astonishing to see that the moves it executed were considered highly unusual or have never been seen before in certain situations of the game by even the most experienced players. If you have not seen the documentary about AlphaGo, I highly recommend it. It is available on Youtube here.
Often successful algorithmic approaches to the financial markets are compared to AlphaGo’s approach to the game of Go. The analogies are very logical from an algorithmic perspective when considering the extreme amount of permutations possible such as in the game of Go. However, comparing the level of permutations in Go to the financial markets is not considered a scientific challenge.
In addition, many scientists take the approach that training algorithms on the historical behavior of the markets will provide the model training necessary for algorithms to anticipate, predict and decide on the next trading moves. That’s why it is so common to hear conversations about how feeding more historical data into models for training will make the executions even better.
However, after analyzing the financial markets at a much deeper level, it becomes clear that the analogies between the game Go and the financial markets are overly simplistic and superficial. In reality, relying on a similar algorithmic approach will create a significant amount of risk to underperformance and volatility.
The fact is, the financial markets are significantly more complex than the game of Go. To explain it very simply, our Chief Scientist explained to me that the composition of “opponents”, or to put it nicely, “players” changes on a day to day and minute by minute basis. Different participants execute their beliefs on the best next move. They all have different strategies, needs, systems, capabilities, knowledge and even emotions.
In thinking about this complexity in relation to Go, you would have to re-analyze your opponent’s personality without ever seeing him before and prior to your next move. In addition, you have a high level of outside influences that have an impact on the markets. For example, breaking news and news stories that specifically have a major impact on the trading conditions of this unprecedented year of 2020.
To make matters worse, news may break in the middle of a trading day providing a significant amount of directional turns throughout the day. Driving this point home is that 18 out of the top 20 largest intraday directional market swings in market history occurred in 2020. Looking at 16 out of these 18 trading days, the intraday swing can be correlated with significant breaking news announcements, such as the vaccine trial breakthrough.
This brings me back to my comparison with the game of Go. The financial markets would be the equivalent of not just dealing with a changing opponent for every move, but an opponent changing before a move is made. In addition, the opponent also has exclusive permission to change his move after you have made yours. Tricky, right?
Leading an algorithmic-focused asset management firm, I have learned that the key to all of this is that models are built with the following key components to be successful:
- Ability to detect market rotations in order to identify underlying dynamics of the market
- Ability to train models based on very small data sets. The reason for this goes back to my earlier argument that every trade situation is unique and has never been seen before
- Participate in both long and short positions to provide an opportunity to participate in any market direction
- Ability to predict and function around market slippage when dealing with intraday trades
- Ability to function effectively across various type of securities to take advantage of de-correlated conditions typically between two asset types
I hope this provides you with a high level perspective on how one would either build an algorithmic trading framework or allow a potential investor into an algorithmic-focused offering to conduct more effective due diligence.