Please be cautious of fraudsters and scammers impersonating Quantedge Capital, and its employees.Learn more
The difference between a tool and a decision-maker — and why it determines who, or what, is in control
Suhaimi Zainul-Abidin
Chief Executive Officer, Quantedge Capital
Xu Wei Chen
Vice-President, Quantitative Research, Quantedge Capital
Few subjects in finance today generate as much simultaneous enthusiasm and unease as artificial intelligence. The same technology is described, sometimes within a single paragraph, as the defining advantage of the coming decade and as a source of systemic risk that no one fully governs. For investors relying on professional investment managers, the unease usually resolves into one practical question: in a moment of market dislocation, who — or what — is actually making the decisions?
It is a reasonable question, and it has a precise answer. But the answer depends on a distinction that much of the industry leaves unexamined: whether artificial intelligence is being used to make investment decisions, or to support the people who make them. These are not two points on a spectrum. They are different activities, with different risk profiles, and conflating them is the source of most of the confusion in the current debate.
The case for AI as an autonomous decision-maker rests on an intuition that a system trained on vast quantities of market data should, in principle, make better investment decisions than a human. The evidence for that intuition is weaker than the enthusiasm suggests, and it is weakest in precisely the conditions that matter most.
The picture is genuinely mixed on short horizons. Some recent benchmarks find that, over a few months of live or simulated trading, most large-language-model agents modestly outperform a passive buy-and-hold position. Taken alone, results like these are easy to over-read.
The more demanding studies are less flattering. A 2025 backtesting framework that evaluated timing-based AI strategies across roughly two decades and more than a hundred securities found that the advantages reported in narrower studies deteriorated substantially once the test was widened and lengthened — a classic symptom of results that owe more to a favourable sample than to a durable edge. The same work found that AI strategies tended to be overly conservative in rising markets, underperforming a simple index, while in the 2008 financial crisis certain agents amplified losses well beyond the market’s own drawdown rather than cushioning them. In one tested case, an AI strategy’s peak-to-trough loss approached 75%, against roughly 50% for the index it was meant to track.
Two further findings deserve attention from anyone tempted to hand a model the keys. In one study, an AI agent that was successfully trained to predict market direction more accurately went on to make less money, not more: it had optimised the proxy goal it was given — classifying the market state — while ignoring the magnitude of returns and the management of risk. In another, agents placed in identical market conditions produced widely divergent decisions, with the larger models favouring high-reward, high-drawdown strategies. Instability of that kind is not a tuning problem to be solved next year. It is a property of systems that are fluent without being accountable.
None of this establishes that artificial intelligence is useless in investing. It establishes something more specific and more useful: that unsupervised AI is least reliable in tail events and regime shifts — the moments when a portfolio’s survival is actually decided. That is a strong argument for treating AI as support for human judgment rather than as a substitute for it. It is the argument on which our entire approach rests.
A common misconception holds that because a quantitative fund is systematic, it must be an automated black box running without human oversight. The truth is closer to the opposite.
A systematic process is one in which every position is dictated by a model that human researchers have designed, debated, stress-tested, and approved before any capital is committed. The rules have been hardwired into the models. They can be examined, audited, and challenged. Far from removing human judgment, a systematic process is a way of encoding it — of making investment decisions explicit, repeatable, and accountable rather than leaving them to the mood of a given morning. Properly understood, systematic is a form of oversight, not an absence of it.
We have operated this way since Quantedge’s founding in 2006. Advanced statistical modelling — the same mathematical foundation on which modern AI is built — has been embedded in our process from the start. What the recent generation of large language models and agentic workflows changed is not the principle but the capacity: these tools allow far more of that modelling to be applied, more quickly, across more of our work. The instruments improved. The governing logic, and the people who own it, did not change.
There are four distinct reasons for unease with AI in finance, often labelled “black box” risk:
• Autonomy — a system that initiates an action before a human has reviewed it.
• Rogue behaviour — an agent pursuing a narrow objective so single-mindedly that it harms the broader portfolio it was meant to serve.
• Hallucination — the tendency of language models to state false information with the same fluency and confidence as true information.
• Accountability gap — the difficulty of assigning institutional responsibility when an automated error contributes to a loss.
Each of these is a legitimate concern. They describe real failure modes, several of which are visible in the research cited above. The appropriate response is not reassurance but architecture.
We address these four concerns through a single operating structure, applied consistently. We separate AI used as a decision-maker from AI used as a workflow lever, and permit only the second. In this arrangement, artificial intelligence functions strictly as an “actor” — drafting code, flagging anomalies in incoming data, summarising the day’s portfolio changes — while experienced professionals remain both the architects and supervisors.
The governing rule is that no AI output is ever live or market-facing on its own. Every workflow that incorporates these tools requires mandatory human sign-off before any output can influence our environment. Whether the tool has helped catch a bug in a new piece of code or assembled a summary of daily risk exposures, a qualified person must vet and approve the result. This single requirement is what intercepts a hallucination or a logical flaw before it can reach anything that matters.
The tools are confined to a sandbox with no access to the core portfolio-management pipeline; they cannot, by accident or design, alter the deterministic rules that define the strategy. We treat prompt engineering as a discipline of requirements specification — defining scope explicitly, enforcing constraints rigidly, and verifying every output — on the principle that a powerful and entirely literal instrument should be told precisely what to do and given no latitude to improvise.
The human sign-off point exists everywhere, and the operative word is everywhere: there is no node in the workflow where the machine is trusted unsupervised.
The most important constraint on AI in investment management is rarely discussed, because it is structural (and perhaps not sufficiently dramatic to warrant attention).
A foundation model is, by construction, trained on the public record and nothing else. It is exceptionally good at compressing and recombining what is already known. But investment edge — the return an investor is actually paid for — is by definition the part of the world that is not yet reflected in public information and therefore not yet in prices. A model trained on what everyone already knows can help analyse the past with great speed. But it cannot originate what no one yet knows.
This is not a temporary limitation awaiting a better model. It is a fact based on the nature of the technology, which defines where AI’s value lies. What AI contributes is leverage on judgment that already exists. A researcher with years of market experience knows how to frame a problem, which variables to isolate, and where a model is most likely to mislead. Supplied with computational leverage and a structured way to test ideas, that researcher works faster and more rigorously. The identical tools in less experienced hands will produce confident errors at greater speed. Leverage is neutral: it amplifies the quality of the judgment it is applied to, in both directions.
At Quantedge, we deploy these tools at specific points in the research pipeline, each chosen because it removes friction from the path to a human decision, without being allowed to make an unsupervised decision itself:
• Synthesis of daily change. Agents aggregate daily portfolio movements from multiple proprietary and public sources, corroborating internal risk exposures against global market activity to produce a single coherent summary of movements across the investment universe.
• Documentation in lockstep with code. AI maintains and updates technical documentation as models change, keeping our institutional memory aligned with the production environment rather than lagging behind it.
• First-pass code review. Before new code reaches the research pipeline, AI traces logic, catches syntax errors, identifies potential bugs, and helps construct the test suites that establish robustness — a first pass that a human reviewer then confirms.
In each case, proprietary data and code are processed within the secure sandbox under protocols ensuring our information is never used to train external models. And in each case the boundaries are explicit: AI is never authorised to make investment or portfolio construction decisions; it cannot reach the deterministic core of the strategy; and its output becomes part of our reality only after a named human has approved it.
Today’s AI and agentic tools are still somewhat narrow; they have no independent objectives to pursue, and the image of a sentient system escaping its leash belongs to fiction. The real risk is still human: that the fluency of the machine persuades the people using it to stop checking its work. This may not be a risk that dominates headlines, but it warrants the greatest vigilance.
Artificial intelligence produces output that looks finished — well-formatted, articulate, and confident — precisely what a disciplined investment process is built to distrust. The fact that AI manufactures such output at scale raises the potential cost of lapses in that discipline. This is why our guardrails are aimed as much at our own people as at the technology. The sandbox, the siloing, and the mandatory sign-off exist not because the software is expected to rebel, but because of the persuasiveness of plausible wrong answers.
The conclusion follows directly. Artificial intelligence has changed what a systematic manager can do, and how quickly, but it has not changed what the work fundamentally is. By delegating repeatable tasks to capable tools, researchers reclaim the time and attention required for the genuinely difficult work — identifying regime shifts, examining new asset classes, designing strategies that hold up under stress. Our edge sits where it has sat since 2006: in a disciplined, systematic process governed by human judgment, and sharpened by tools that make that judgment faster and better tested. The quality of the tools may have changed over the years, but the primacy of human judgment has not.
• M. Li et al., TrustTrade: Human-Inspired Selective Consensus Reduces Decision Uncertainty in LLM Trading Agents, arXiv:2603.22567, 2026.
• StockBench: Can LLM Agents Trade Stocks Profitably in Real-World Markets?, arXiv:2510.02209, 2025.
• The Losing Winner: An LLM Agent that Predicts the Market but Loses Money, NeurIPS 2025.
• W. Li et al., Can LLM-based Financial Investing Strategies Outperform the Market in Long Run? (FINSABER framework), arXiv:2505.07078, 2025.
This article is for general information and does not constitute investment advice.