Using AI to Power Trading Strategies that Never Stop Evolving

Can you briefly explain the difference between quantitative trading and AI-based trading for our audience, and how they have evolved over the years? 

There are two key fundamental differences. The first difference is the benefit of future knowledge that the quantitative trader (or trading system) has (and cannot get rid of, creating perception bias), versus the AI, which has used the same information to build its decision-making capabilities, but subsequently has had that knowledge stripped out, leaving only the decision-making process intact. Put another way, quantitative solutions, when backtested, cannot “forget” the knowledge of what’s about to happen, whereas the AI systems have literally been given amnesia, and act as if they are “seeing” the test information for the very first time. The second key difference is that quantitative trading is fixed; modifications of the trading system or strategy require human intervention. AI trading systems by definition have a machine learning layer and retrain at regular frequencies. This means that (in a well-developed AI trading system), no human intervention is required for the AI to self-evolve and adapt the trading strategy to changing market conditions. 

There are various misconceptions about quantitative trading and AI-based trading. How do you address the common misconception that AI-based trading systems are “black boxes” that cannot be understood or controlled? 

Well, I’d start by saying that it’s not necessarily a “misconception”; there are black box AI trading systems (Renaissance Technologies being perhaps the most famous user of those systems). ​I’d argue that black box systems are the most profitable without question, because of the dynamism the AI enjoys via the very light restrictions on decisions it might make, or risks it is permitted to accept, and yes, it is true that the decisions black box systems make cannot be easily explained, or explained at all. This is disquieting to many investors because humans use linear logic paths to make investment decisions, and AI does not. But humans forget that we too make similar, non-linear decisions – for example when we see a piece of art we like or hear a song that is pleasing to us, or find someone we just met physically attractive. At the moment, if challenged, we would have a difficult to impossible time articulating the “why” of those instant impressions, because they are not made via linear processes we can walk back through.   

Most systems are not black boxes, but rather employ varying layers of human-directed guardrails, and follow specific rules tied to the investment strategy, manager’s ideologies, and overall risk envelope that’s deemed appropriate for the fund and its investors.

By and large, however, most systems are not black boxes, but rather employ varying layers of human-directed guardrails, and follow specific rules tied to the investment strategy, manager’s ideologies, and overall risk envelope that’s deemed appropriate for the fund and its investors. While buy/hold/sell decisions within these systems may still employ non-linear processes, the ideologies and guardrails present can be explained. And that is how we address general discomfort with AI making end-point decisions within a landscaped framework: they can’t just do whatever they want. They are following – creatively – human-directed rules within a human-incepted strategy. 

Quantitative trading has been a well-established practice for decades, while AI-based trading is a relatively newer concept. ​In your opinion, what are some of the unique advantages AI brings to the trading landscape that traditional quantitative approaches might lack? 

Somewhat revisiting the answers in question #1,​ I believe the advantages are​ 1) confidence in the integrity of the model, which is supported by on-model execution in real-time (due to the lack of perception bias), and 2) the autonomous learning layer. Quantitative systems rarely bear out replication of the model; well-programmed AI systems (say for example the DIP ETF) execute exactly the model, which increases manager certainty that the test results are likely representative of how the strategy will perform over forward time.  

Quantitative trading often involves the use of historical data and statistical analysis to identify patterns and trends. How does AI-based trading go beyond traditional quantitative methods to capture more complex patterns in financial markets? 

Again, quantitative systems are fixed, by nature. So the quantitative system requires (substantially) the exact same market conditions with the same market participants reacting the same way, to deliver the expected performance. AI is ever-evolving, and so as the patterns the system has trained on themselves evolve, the AI is able to make micrometric adjustments that quantitative systems cannot, in order to continue to match performance. AI systems may stumble when a new condition is presented, just as quantitative systems do, but they will learn the new condition, and the patterns that accompany it, and behave differently the next time that new condition is encountered – unlike the quantitative system, which will fail in the exact same way, over and over again, unless manual intervention occurs. Sophisticated AI systems can also learn to anticipate changes, and act defensively, whereas quantitative systems cannot, because they don’t have a “brain” capable of making any adjustments autonomously. 

Regulations in the financial industry are continually evolving. How does Kaiju navigate the regulatory landscape concerning the use of AI in trading, and how do you ensure compliance with relevant laws and guidelines? 

Thus far there isn’t much in the way of regulation governing the use of AI in trading, but we do expect that to change. The easiest way to remain on the right side is to 1) ensure that all data consumed is paid for (we do not “scrape” the Internet for anything), 2) remember what AI does best is pattern recognition, and not market or market participant manipulation, which are areas that regulation is likely to be most focused on. 

A significant challenge in AI-based trading is the risk of overfitting models to historical data, leading to poor performance in real-world conditions. How does Kaiju address this challenge and ensure the generalizability of its AI models? 

Collectively, we‘re not originally from the financial space; as a diverse group of scientists, we have a lot of experience avoiding over-fitting scenarios that would have been detrimental to work in our previous scientific lives (particle collision theory, in real-world applications for example, can yield some fairly catastrophic results when model over-fitting occurs…). When we started using AI in trading, years ago now, we brought this discipline to the work we do here at Kaiju. We generally adopt an “it can’t be worse than this” outlook with our model validations: forcing terrible executions, slippage that we’ve never personally experienced in real life, counterparty rejections that never happen, and disadvantageous order flow into our core assumptions. We match that with a lower candidate suitability rate than our models show is achievable, and the result is that our real-world performance ranges from on-model, to better-than-on-model when forward tested. Our DIP ETF for example, rebalances a significant portion of its portfolio daily, executes entirely on-model, and has done so without variance since we launched it over 7 months ago.

As technology advances, new AI-based trading strategies and techniques emerge. How does Kaiju​ seek to stay at the forefront of these developments, and what investments do you make in research and development to enhance your trading algorithms? 

Machines are awful at context, and can’t understand nuance at all. So where I see that “balance”, is in letting the machines do what they do best – pattern-based technical trading, flow predation, and what we call Stratified Risk Distribution ® – and letting the humans do what they do best, like global macro.

We’ve been doing this for a while now, so ​I believe ​we have a​n​​ ​​advantage in that we’re constantly improving a substantially mature system, several generations in at this point, rather than dealing with a ground-up proposition. That said, we do have a robust pipeline for evaluating new technologies that might be complementary to, or superior to, current components of our systems, to make sure that we’re able to consistently innovate in the ways we’ve become accustomed to. In our time working in this space, we’ve already seen enormous progress; model retraining that used to take weeks now gets done over a weekend, or for shorter retraining sessions, overnight. In terms of the strategies themselves, we have a current backlog of profitable human-incepted strategies that remain relevant that we are working through, evaluating each for suitability on the private or public fund side, and determining whether any are workable within an ETF wrapper (which of course lets us bring it to the broadest possible investor audience for consideration). Regardless of how powerful the foundational technologies are, and what levels of efficiency we now enjoy having access to, our cycle from strategy evaluation through refinement, all the way to live use in the market, isn’t something we take lightly or cut corners on. Each strategy has about a 12-month timeline it goes through at this point before you’d see it in a fund or launched as an ETF, and while we’re thankful advances in this space are removing some bottlenecks and reducing some of our analysis cycles, there’s a floor below which we’d consider the process “rushed” regardless of how robust or efficient the validation.

The biggest investments we make are in the areas of data sanitisation and services, and of course, talent acquisition. The latter really is the name of the game these days; top talent is very difficult to find, and acquire, so we have formal relationships with the top institutions worldwide, and incubation programs within those institutions that give us access to professionals at a much earlier stage than would normally be the case. 

The human element is still crucial in trading decision-making, especially in times of market unpredictability. How can financial companies strike the right balance between human expertise and AI-driven strategies? 

I’d argue that humans are pretty terrible at making decisions in times of unpredictability; just look at the performance of “top managers” across multiple crises. They may have managed one or two well, but not most of them. That’s an area where​ I believe machines excel because they’re not subject to bias, ego, or fear. Where humans really shine is in understanding context and nuance, and making investment decisions based on those criteria. Machines are awful at context, and can’t understand nuance at all. So where I see that “balance”, is in letting the machines do what they do best – pattern-based technical trading, flow predation, and what we call Stratified Risk Distribution ® – and letting the humans do what they do best, like global macro. An AI system can’t currently intuit what a global leader will or won’t do, what global political response to that action might be, and what the resultant socio-economic fallout might be, locally, regionally, or Internationally, with enough certainty to profit off of it. Humans can, and do, “read’ those things with greater skill. 

Many traders are hesitant to adopt AI-based strategies due to concerns about job displacement. How do you see AI and human traders working together in the future, and how does Kaiju support its human workforce in embracing technological advancements? 

​​I believe ​Technical and quantitative traders will be replaced by AI systems at the execution and portfolio management levels, period. That’s a certainty​ in my view, just as widespread PC adoption was a certainty, and we watched file rooms and floors full of typists evaporate. But I do not see traders becoming obsolete at all – in fact, at Kaiju, they’re a critical component of our AI systems development and strategy implementation. Market mechanics are not mathematical processes whose outcomes can be guaranteed; just because an asset has a theoretical fair market value, does not mean that there will be available buyers or sellers of that asset when required. Traders assist the AI team with the real-world implementation of the systems that work so well in a lab but might work less well (or not at all) in practical application. Then you have counterparty relationships. While basic equities trading in liquid securities is already almost entirely machine-driven (at the institutional level), that is not the case for derivatives of almost any kind. While the machine can generate the execution instructions, traders still have to carry out (and at times modify) those instructions in order to achieve the desired outcome.  In a machine vs. machine world, you’d just end up with endless stalemates, while humans are able to negotiate the path through, and out the other side. Finally, I’ve yet to see a machine-incepted strategy (outside of black box systems) which wasn’t awful. In my experience, the best partnership is human-incepted strategies, refined and perfected (and sometimes executed) by machines – and that’s what we do here at Kaiju. 

Looking ahead, what do you foresee as the most significant challenges and opportunities in the realm of AI-based trading, and how is Kaiju positioning itself to address these effectively? 

From my perspective, the most concerning challenges fast approaching involve questionable ethics. AI is a phenomenal mimic (as has been seen extensively with GPT), and my fear is that unscrupulous professionals may use AI to influence retail investment decisions, and then predate on those decisions via internal systems; hounds to the hunters, if you will. Large complex systems capable of that would be very difficult to identify, and then triage, if they could be identified at all. It’s not the machines that we should be fearing (regardless of how popular a trope that is in science fiction); it’s people using machines to predate on other people, via channels that the machines can access. Remember that at the end of the day, the machines follow the instructions we give them. Even a black box trading system (which has the least controls it must abide by) cannot spontaneously decide it would rather sing in Spanish… My hope is that global resources are directed into identifying these types of predatory systems, and neutralising them (and we can do that, already).  

​​In terms of opportunity, that’s where the doom-and-gloom of how awful people can be when given a subjugative tool and left unchecked turns to optimism. AI represents enormous empowerment to many who do not currently have it. An AI risk evaluation system, for example, could give any investor valuable, accurate portfolio risk assessments on a regular basis, and even a broad base of options for remediation of any deficiencies, which could keep them safe in times of instability or crisis. Investors could use AI to model their own strategies or evaluate the levels of certainty associated with an investment they are planning on making. In short, responsible application of AI would lead to more efficient markets, with better informed, more capable participants – and that’s to everyone’s benefit. Humans will always be a part of this; it’ll never be just “who has the best machine”, and I think the assistive capabilities of AI are vastly more exciting than the science fiction of true Artificial General Intelligence. We need help as humans, not an extinction mechanism (because left alone, we appear to already have that largely covered…). ​ 

Important Risks 

The Fund is distributed by Quasar Distributors, LLC. Exchange Traded Concepts, LLC (the “Adviser”) serves as the Fund’s investment adviser. Kaiju ETF Advisors (the “Sub-Adviser”) serves as the Fund’s investment sub-adviser.  

Investing involves risk, including loss of principal. The Fund is subject to numerous risks including but not limited to: Equity Risk, Large Cap Risk, Management Risk, and Trading Risk. The Fund is actively managed and may not meet its investment objective based on the Sub-Adviser’s success or failure to implement investment strategies for the Fund. The Fund’s principal investment strategies are dependent on the Sub-Adviser’s understanding of artificial intelligence. The Fund relies heavily on a proprietary artificial intelligence selection model as well as data and information supplied by third parties that are utilized by such a model. Specifically, the Fund relies on the Kaiju Algorithm to implement its principal investment strategies. To the extent the model does not perform as designed or as intended, the Fund’s strategy may not be successfully implemented and the Fund may lose value. A “value” style of investing could produce poor performance results relative to other funds, even in a rising market, if the methodology used by the Fund to determine a company’s “value” or prospects for exceeding earnings expectations or market conditions is wrong. In addition, “value stocks” can continue to be undervalued by the market for long periods of time. The Fund is expected to actively and frequently trade securities or other instruments in its portfolio to carry out its investment strategies. A high portfolio turnover rate increases transaction costs, which may increase the Fund’s expenses. Frequent trading may also cause adverse tax consequences for investors in the Fund due to an increase in short-term capital gains. The fund is new, with a limited operating history. 

Executive Profile 

Ryan Pannell, CEO and Global Chair of Kaiju Worldwide, an asset manager democratizing access to AI-powered investment products and issuer of the BTD Capital Fund (Ticker: DIP). Ryan was a pioneer in cryptographic web-based messaging technologies, studied theoretical physics, and quantum mechanics, and enjoyed a successful career in film and television before becoming a professional trader. This unorthodox mix of markets and the arts gives Ryan a unique perspective on investment management and the novel applications of AI.