Artificial intelligence is uncovering previously invisible patterns in financial markets, patterns that drive investment performance. But in a world of breakneck technological revolution, the advanced machine learning algorithms that underpin these AI engines can’t be static. Investors need to be confident that they are designed to adapt in response to evolving technology, data and markets.
Get it right, and the reward is returns well in excess of what’s available from a benchmark equity index at only modest additional cost.
Bigger, better, faster
Take Quest AI. Since its launch in July 2023, the strategy has outpaced its benchmark index thanks to the powerful insights it generates into what underlies stock prices. These insights have only been possible thanks to the adaptability of its underlying process, which has kept pace with a big increase in data inputs and technological complexity.
Traditional investors usually rely on a small set of familiar rules about what drives stock returns, rules that are reasonably good at explaining broad, long-term patterns. But in the case of individual companies and over shorter time frames, prices move for many different reasons at once. Some we understand – they can be related to the speed at which information spreads, or to investor psychology or how trading in the market is organised. Other reasons are harder to identify or describe.
That’s where AI comes in. Instead of starting from a few pre‑selected ideas, we feed the system a very rich description of thousands of companies, covering many types of information. The AI then looks back over history to learn which bits of information mattered when, and how they interact. In this way, the tool can capture a balanced mix of well‑understood influences and more subtle patterns that are difficult for human investors to spot.
This complexity, in turn, has necessitated a big expansion of the tech infrastructure behind it. It’s an evolutionary process that doesn’t stop. The latest innovations in AI – and computing more generally – are constantly being assessed for what they might bring to the approach in future, be it large language models (LLMs) or quantum computing. It also means staying ahead of the crowd – most quant investors have yet to deploy AI meaningfully in their processes, according to a recent survey by Bloomberg Research (see Fig. below). https://assets.bbhub.io/promo/sites/33/Bloomberg-Research-Data-Roadshow-Survey-Booklet_2025_FINAL.pdf
When we launched our AI-driven long-only strategy, the process used around 200 features, with each being a data series that assesses a characteristic or indicator about a stock over time. The algorithm analysed their complex interdependencies to isolate the signals that are then used to improve stock picking performance. Since then, the number of features has doubled, with the number of observations increasing more than tenfold to several billions. These expanded features include more refined measures of analyst sentiment – such as how quickly they revise their earnings forecasts, how dispersed these forecasts are across Wall Street, how often target prices are updated – as well as high frequency indicators of trading activity like changes in liquidity during the day.
That massive expansion in processing could only be done by an equally massive increase in computing power. For instance, the number of central processing units (CPU’s) used to compute our model jumped from just under a hundred to more than three thousand, the number of servers more than doubled and the number of graphical processing units (GPU’s) increased four-fold. That’s the computing power equivalent of expanding from 7 to 120 Playstation 5s (the benchmark in complex gaming graphics). Enhancing our GPUs’ capacity substantially shortens the time required to train our models, allowing us to conduct more experiments and make improvements at a much faster pace.
Taking this processing capacity in-house has the added benefit of not being subject to big price rises from cloud companies and other external suppliers as demand for high end computing booms.
Bloomberg survey* of investment research professionals asking "Where are you on your generative AI journey?"
*149 responses, survey conducted from April 2025 to November 2025. Source: Bloomberg.
Language at large
There’s been considerable hype about how the use of large language models (LLMs) can improve stock selection. To be sure, there’s a growing body of literature that shows they can infer sentiment from news to create investment signals. But, so far, we haven’t found them to be useful for our Quest AI strategy. Given our investment horizon, which is rarely more than a few weeks, most of the predictive content contained in news is tightly linked to earnings announcements and analyst commentaries which are already processed by our algorithm. So while there is a considerable body of information to be gleaned from news, our own proprietary signals are currently more effective than sentiment signals generated by LLMs.
We found that the expertise with which we’ve formulated our algorithm and processed our data has enabled us to stay ahead of what LLMs are able to do. LLMs are designed to understand and generate text and can even produce complex mathematical proofs because mathematics is, at its core, a formal language system. Markets, however, aren’t governed by fixed mathematical concepts. Rather, they are driven by adaptive human behaviour, competition, feedback loops, and constraints. There’s a further issue with LLMs: when using them to back-test models they tend to incorporate information that’s not available in the period being back-tested, distorting the results. As a result, while LLMs are powerful tools for analysing information, they are ill suited to predicting stock returns in a way that survives real world constraints.
But we also monitor developments closely and are ready to use LLMs should they prove capable of generating new predictive signals.
And we are using LLMs to help with producing portfolio commentaries and improving the efficiency of code generation, creating presentations and gathering research material, albeit within ring-fenced systems to ensure that proprietary information doesn’t leak into the public sphere.
An agentic future?
There is a major push across the AI universe to develop AI agents – automated systems that can interpret complex instructions and execute them independently. If AI can uncover hidden patterns in the market to unearth investment opportunities, agentic AI can go one step further. It can make the investment process itself more efficient – by keeping control of costs while also implementing AI insights in a consistent manner.
So far, it’s more promise than reality, though industry experts expect it to become functional for some applications over the coming year or two.
For Quest, agentic AI would most likely involve streamlining both reporting and idea generation. Within the financial services industry, there are already projects in train to automate the work of junior investment banking analysts such as generating valuation models, preparing presentations or standardising documents.
Ultimately, we can envision agents that also help with coming up with investment ideas by, among other things, rapidly extracting key insights from new academic research. There will always be a role for humans – not to compete with machines on speed and data processing abilities, but to ensure that autonomous systems operate as they are intended to.
Ch-ch-changes
The attractions of using AI models to optimise investment choices are becoming clearer by the day. But end investors also need to bear in mind the pace of technological change and expansion of data sources affecting these models. Any AI-driven model needs to be adaptable. Which means the people behind the model need to be both nimble and at the cutting edge of technology.
Many of our team’s breakthroughs have come from unexpected sources. But serendipity depends on hard work and flexibility of thinking. Culture is critical – here it means prioritising continuous learning, experimentation and sharing of knowledge.
Flexible processes and deep technical understanding allow us to pivot quickly to new opportunities or to meet challenges when they arise. For instance, we broadened our algorithm’s feature set to incorporate richer analyst estimate revisions and market signals. And at the same time, we re-engineered the operating hardware by tripling the algorithm complexity and moving to dedicated in-house CPUs and GPUs so that the strategy could absorb the additional volume of information without increasing risk or cost to our clients.
This doesn’t just mean reacting to trends but also shaping them. Now, with the advent of AI, we are pivoting again by redefining the human-machine interface and thus turning AI into a seamless extension of human creativity. We embrace the risks associated with early adoption, because the cost of standing still is far higher.
And while AI is our current focus, our approach is technology-agnostic – if we find better tools with which to extract excess returns for investors, we will embrace them. Our team has a proven history of identifying when to shift direction and how to execute those pivots effectively. We have full confidence that we’ll keep doing so as new tools and methods become available – and in doing so we will continue to reward our investors.
