Hedgeweek exclusive: Asset management is on the threshold of deploying investment strategies using next-level AI. Patrick Ghali (pictured), Managing Partner of investment consultant Sussex Partners, gives his thoughts on how best to prepare investors.
- Asset management is on the threshold of deploying next-level AI
- Investors tend to struggle with the idea of an AI taking decisions
- But there are ways – like risk management – to provide comfort
By Patrick Ghali
Managing Partner, Sussex Partners
Much has recently been written about AI and it seems impossible today to have a conversation about the future of asset management without ChatGPT and similar tools being mentioned as potential game changers.
Systematic investing, however, is nothing new. As the simple models of the past have become less productive, they have given way to ever more complex ones, with armies of PhDs looking for ever more esoteric relationships to model and exploit. First, this has led to advances in machine learning, and now it seems we are on the threshold of AI as the next phase in this evolution.
To consider what approaches already exist in asset management in respect of AI and machine learning, and which of those are the most promising, it helps to think of the integration of AI in other fields such as autonomous driving where there are different levels of sophistication, each representing a substantial leap towards the ultimate goal of a system fully controlled by AI.
The levels of AI
At its most basic level, Level 1, there are classical statistical models where AI optimizes a very limited set of parameters in a predefined and simplistic rule-based setup. Outputs can be correlations, portfolio weights, optimal parameters to technical indicators, risk limits and so on. They are used to assist portfolio managers in their decision-making processes, and the use of the results ultimately is at the discretion of the human trader.
Level 2 AI is exemplified by data mining and pattern recognition. Classical machine learning algorithms such as random forest, kernel regression, support vector machine and shallow neural networks are able to identify repeating patterns within large pools of data and give complex insights. However, they are still a tool to the human trader and provide results in the form of sentiment scores, chart alerts, and trade recommendations.
Level 3 is the most sophisticated approach to AI. Here complex (typically deep-learning) models are created where the strategy is not limited by the model, but only by the data. This is achieved by the use of deep reinforcement learning, large-scale neural networks and suchlike. These systems have all degrees of freedom to decide, without human limitation on features generation, asset selection, investment decisions, risk management, optimal allocations, and trade creation and even execution optimization. Once Level 3 is reached, the system can autonomously trade and constantly adapt to changing environments.
While Levels 1 and 2 can add value to asset managers, it is Level 3 that is seen as the holy grail. It will allow asset managers to trade markets in completely new ways, and to come up with strategies that will be fundamentally different to anything else currently known. In theory at least, this should lead to non-correlated and highly attractive risk adjusted return profiles.
One question which arises is whether Level 3 could lead to capacity constrained trading strategies should multiple investors all launch their own “AI portfolio managers” which hone in on similar market inefficiencies. However, according to Dr. Michael Kopp, the Founding Director of the Institute of Advance Research in Artificial Intelligence in Vienna, such AI trading systems would likely be able to learn from each other and collaborate hence potentially mitigating this risk.
The ability of an AI to digest vast amounts of data very quickly makes it in many ways vastly superior to any human trader. Additionally, it should be possible for an asset manager with access to such an “AI portfolio manager” to literally create new “AI portfolio managers” at the click of a button. “”AI portfolio managers” that strictly adhere to risk limits, never get tired, and never lose focus.
Easing investor discomfort
So where are all these “AI portfolio managers” today, and what about the black box aspect of any such system?
While there are a numerous, very well established and successful firms employing traditional systematic investment techniques, and a number of companies employing Machine Learning (e.g., Millburn) or simple AI driven investment solutions (e.g., Qraft, Boosted.Ai), most seem to still be employing Level 1 or Level 2 rather than Level 3 techniques. It doesn’t seem as if anyone as yet has been able to successfully launch a Level 3 product focused on traditional assets, but Rubinstein & Schmiedel may be the closest so far to achieving this holy grail.
With regard to the black box concern, the same can be said of many more traditional systematic approaches. It is true, that allocators tend to struggle with the idea of an AI taking decisions which cannot be explained in simple ways. There is a level of comfort in knowing that losses stem from a “bad decision” – markets not reacting as expected to fundamental data, historical correlations breaking down or, at worst, a breach of risk management rules by a human trader.
There are, by contrast, a number of ways in which AI can be assessed which should help provide a certain level of comfort to investors. For one, it is possible to apply similar standards to AI as one would apply to a human trader. The “AI portfolio manager” can be presented with different market situations and the AI’s behaviour studied for each such scenario (arguably easier to do with AI than with a human).
Additionally, Algorithmic testing can be employed, and thousands or even millions of scenarios tested and outcome distributions evaluated (including maximal losses, etc.). Additional risk management measures such as diversification among different AI strategies, especially uncorrelated ones, can also help to mitigate any AI specific risk (much like investors do already by allocating capital to multiple PMs or funds).
Lastly, fixed boundaries can be provided so that an AI can only make decisions within those predefined boundaries, thereby avoiding style drift. This should ensure that even in a worst-case scenario, losses are limited. While this last aspect will likely be the most effective at limiting loss potential, it may come at the price of potentially lower returns and less efficient use of capital. The most sensible approaches to AI trading will likely make use of all three of the strategies mentioned to gain additional insights and limit risk.
One further benefit of any systematic strategy, including AI is the ability to replay what has happened, and to further learn from this. This of course is much harder to do with human traders.
Looking ahead
As with any new development, investors will need time to adjust and digest both the potential and the risks AI represents to their portfolios. It seems inevitable though that AI will end up playing an important role in future asset management. Whether that is simply by providing human traders with additional inputs, or whether it will replace portfolio managers completely remains to be seen.
With the value of studiously acquired knowledge quickly losing its importance now that real time answers to any conceivable question can be produced at the push of a button, the real killer app of the future may to be the ability judge and interpret information, and to draw the right conclusions from this. Perhaps, for this, human input will continue to be crucial.
Patrick Ghali is Managing Partner at Sussex Partners, an alternatives-focused investment consultant, and serves on the advisory board of Rubinstein & Schmiedel. The topic of AI and hedge funds will be explored in this month’s Hedgeweek Insights Report. To subscribe to the Insights series and receive this report upon its release, follow this link.
