Are large language models good at making investments?

By Matthew Parish, Associate Editor

Tuesday 12 May 2026

The emergence of large language models has provoked a peculiar kind of intellectual temptation. Machines that appear capable of writing essays, passing examinations, generating software code and synthesising complex information from vast quantities of text naturally invite a further question: if these systems can process so much information so quickly, might they also be trusted to choose our financial investments? The proposition is alluring. Investment analysis is, after all, fundamentally an exercise in information processing under uncertainty. One reads company reports, evaluates management statements, interprets geopolitical conditions, compares valuations, and attempts to anticipate future behaviour in markets composed of millions of human decisions. A sufficiently capable machine, one might think, ought eventually to outperform the flawed intuitions of human beings.

Yet the issue is more complicated than the marketing literature of artificial intelligence companies often suggests. Large language models are not financial prophets. They are probabilistic text engines trained to predict plausible sequences of language. Their strengths and weaknesses do not align neatly with the requirements of prudent capital allocation. To use them wisely in finance requires a sober understanding of what these systems actually are, what they can realistically achieve, and where their dangers lie.

The first and most obvious attraction of large language models in investment analysis is speed. Human analysts are constrained by time and concentration. Even exceptionally intelligent portfolio managers cannot read every annual report, central bank statement, sanctions announcement, commodity projection, engineering paper or legal filing relevant to global financial markets. Large language models can ingest extraordinary quantities of material in moments. They can summarise trends, compare documents, identify recurring themes and synthesise broad fields of knowledge into coherent narratives.

For institutional investors drowning in information, this capability has genuine value. A hedge fund manager following semiconductor supply chains might ask an artificial intelligence system to summarise Taiwanese export restrictions, Chinese rare earth mineral policies, American industrial subsidies and earnings calls from multiple technology firms simultaneously. A sovereign wealth fund might use language models to monitor political instability across dozens of emerging economies. In such contexts, large language models function not as autonomous decision-makers but as accelerants for human cognition.

Moreover financial markets themselves are increasingly narrative-driven. The value of many modern assets depends less upon present cash flow than upon future expectations. Technology companies, biotechnology firms, defence contractors and artificial intelligence startups are all heavily influenced by sentiment, speculation and projected future dominance. Because large language models are exceptionally good at identifying linguistic patterns and prevailing narratives, they may possess unusual usefulness in detecting shifts in market psychology.

Indeed some quantitative trading firms already employ artificial intelligence systems to analyse sentiment across news reports, earnings calls, social media discussions and geopolitical commentary. If millions of investors suddenly become optimistic about a sector, such sentiment can itself become economically consequential. Markets are reflexive. Expectations alter behaviour, and altered behaviour changes economic outcomes. In such circumstances, a system adept at monitoring language may offer substantial advantages.

However these strengths conceal profound structural weaknesses.

Large language models do not understand truth in the human sense. They do not possess reasoning grounded in physical reality, institutional memory or lived economic experience. They generate outputs that resemble convincing analysis because they are trained upon immense quantities of human-produced text. But finance is not merely a linguistic exercise. Financial markets involve power, deception, panic, incentives, regulation, irrationality and events that have never occurred before. A machine trained upon historical language may struggle precisely when genuine novelty appears.

This creates a fundamental danger. Large language models are often persuasive even when they are wrong. Their tone is frequently authoritative, polished and coherent. In ordinary conversation this may be harmless. In finance it may become catastrophic. Investors who mistake fluency for accuracy risk placing undue confidence in systems incapable of genuine comprehension.

One may compare the issue to a highly articulate university student who has read every economics textbook ever written but never managed a business, negotiated a contract, experienced a banking crisis or watched a military invasion disrupt commodity flows in real time. Such a student might sound extraordinarily convincing while remaining profoundly naive about how institutions behave under stress. Large language models possess an analogous weakness.

This problem becomes especially acute during periods of market discontinuity. Financial history is shaped disproportionately by rare events: wars, banking collapses, pandemics, revolutions, technological disruptions and sovereign defaults. These moments are precisely when historical correlations break down. Human beings with practical experience sometimes recognise such ruptures intuitively because they understand the fragility of institutions and the psychology of panic. Large language models, by contrast, remain dependent upon patterns extracted from historical text. They may therefore become dangerously unreliable when reality departs from precedent.

The war in Ukraine provides an illustrative example. Prior to the full-scale Russian invasion in February 2022, many conventional economic assumptions concerning European energy dependency, defence spending and sanctions resilience proved spectacularly mistaken. Investors who relied heavily upon stable pre-war assumptions suffered serious misjudgements. Human analysts with direct regional knowledge, understanding of Russian political culture or awareness of military realities often perceive risks that formal economic models overlooked.

Large language models trained predominantly upon historical financial literature before the invasion would likely have struggled to grasp the scale of discontinuity that followed. Defence manufacturers, energy infrastructure firms, logistics companies and agricultural commodity producers all experienced dramatic repricing because geopolitical reality abruptly overwhelmed conventional market assumptions. Such moments demonstrate the limitations of systems that derive their apparent intelligence from textual continuity rather than lived strategic understanding.

Another difficulty lies in incentives. Financial markets are adversarial systems. Public information is often manipulated intentionally. Corporate executives present optimistic narratives. Governments conceal weakness. Speculators promote assets they already own. Social media is polluted by coordinated misinformation campaigns. Even financial journalism may reflect ideological or commercial biases. Large language models trained upon public discourse inevitably absorb portions of these distortions.

An investor using artificial intelligence naively may therefore consume an elegant synthesis of manipulated information without recognising the underlying corruption of the data itself. The problem is not merely technical but epistemological. If the training material contains propaganda, exaggeration or fraud, the outputs may reproduce these weaknesses with remarkable sophistication.

There is also the question of herd behaviour. If large numbers of investors begin using similar language models trained upon similar datasets, financial markets may become more homogenised rather than more intelligent. Investors may converge upon identical narratives and strategies, amplifying bubbles and volatility. Financial crises often emerge precisely because too many participants begin thinking alike. Artificial intelligence could intensify this danger dramatically.

One can imagine a future in which thousands of investment funds rely upon related artificial intelligence systems analysing the same earnings statements, geopolitical developments and macroeconomic indicators. If those systems simultaneously interpret conditions as favourable or dangerous, capital flows may become violently synchronised. Rather than stabilising markets, artificial intelligence could produce sudden cascades of buying or selling with destabilising consequences.

Yet it would be mistaken to conclude that large language models are therefore useless in finance. The prudent conclusion is subtler. These systems may be extraordinarily effective assistants while remaining poor masters.

A disciplined investor may use large language models to challenge assumptions, accelerate research and broaden awareness of unfamiliar sectors or jurisdictions. An analyst considering investments in Eastern European infrastructure, for example, might use artificial intelligence to summarise regulatory frameworks, identify relevant political actors and compare historical precedents across multiple countries. A venture capitalist exploring drone manufacturing or artificial intelligence security might use language models to map technological ecosystems and identify emerging companies.

In such contexts the human retains strategic judgement while the machine expands informational reach. This is likely the most rational equilibrium. Artificial intelligence becomes analogous to a calculator for language-intensive reasoning: immensely useful but fundamentally subordinate to human responsibility.

The deeper philosophical issue concerns the nature of investment itself. Successful investing has never been purely a matter of information. Markets reward temperament as much as intellect. Patience, courage, scepticism and emotional discipline matter enormously. During crises, investors must tolerate uncertainty, resist panic and make decisions amid incomplete information. These are not merely computational problems. They are psychological and moral ones.

A large language model cannot experience fear of bankruptcy, responsibility toward employees, loyalty to a nation or intuition about political legitimacy. It cannot visit a factory, observe corruption within a ministry, detect the charisma of a founder or sense the morale of a military population during wartime. Human beings possess forms of tacit knowledge accumulated through lived experience that remain extraordinarily difficult to encode statistically.

Indeed, many of the greatest investors in history succeeded precisely because they perceived realities invisible to prevailing quantitative models. They understood human weakness, institutional fragility or technological change before consensus recognised it. Genuine investment insight often emerges not from averaging existing information but from perceiving what others fail to see. Large language models, built fundamentally upon statistical aggregation of existing language, may therefore struggle to generate genuinely original strategic insight.

There is a final irony here. Financial markets increasingly reward authenticity, scarcity and trust precisely because information has become abundant and cheap. If artificial intelligence systems make generic analytical capability universally accessible, then uniquely human judgment may become more valuable rather than less. Investors able to combine technological tools with practical experience, geopolitical literacy and emotional discipline may outperform both purely human traditionalists and purely machine-driven speculators.

The prudent use of large language models in investment therefore resembles the prudent use of any powerful instrument. One should neither worship nor fear them. They are tools — sophisticated, useful and dangerous. They can dramatically improve research efficiency and broaden analytical capacity. They can identify patterns invisible to exhausted human analysts. But they also hallucinate, oversimplify and inherit the biases of their training data.

To entrust one’s financial future entirely to a probabilistic language engine would be reckless. To ignore such systems entirely would likely become increasingly impractical. The wisest course lies somewhere between technological utopianism and reactionary scepticism: use artificial intelligence extensively, but never surrender independent judgement to it.

Markets ultimately remain human institutions. They are governed not only by mathematics and information, but also by ambition, fear, pride, violence, law, politics and trust. Large language models may assist us in navigating that world, but they do not yet truly understand it.