What Aristotle can teach us about AI-enabled quantitative investment

{"text":[[{"start":7.65,"text":"At first, equity analysis was marketing. The job of a bank analyst was to encourage brokerage trading while helping to grease wheels in corporate advisory. The conventional way to work it was to recommend buying stocks. Then came the regulatory interventions — May Day, Spitzer, Volcker, Mifid II — and a crisis of purpose from which the industry never recovered. "}],[{"start":31.1,"text":"As a result, when opening a sell-side note, it’s no longer a surprise to read sentences like this:"}],[{"start":null,"text":"

Aristotle’s Poetics, in our view, provides a useful framework for justifying quality premie and for how investment narratives can shake off contrary indicators when those indicators have a logical explanation for being transitory.

"}],[{"start":37.4,"text":"The quote above is from Jefferies’ chemicals sector team, led by analyst Laurence Alexander, and buried under the words is a reasonable point. "}],[{"start":46.3,"text":"A lot of quantitative trading these days uses large language models to create investment narratives. Having been fed an earnings report, a few thousand newswires and social media feeds, some price data, and a heap of literature about behavioural finance, a computer can estimate the strength of investor conviction. The computer can also pattern-match its signals to previous investment cycles, appending labels like “euphoria stage” or “capitulation stage”. Its results are nearly instant, quantifiable, evidence-based and produced without obvious prejudice. So, in theory, the computer should be early to spot when investor sentiment is turning."}],[{"start":83.94999999999999,"text":"Jefferies’ chemicals team runs AI screens to help define current narratives in areas such as data centres, robotaxis and alkene supply. “Understanding where a narrative sits in its lifecycle is often more valuable than understanding the underlying fundamentals — because price moves on the change in investor conviction,” write Alexander et al. "}],[{"start":105.14999999999999,"text":"Unfortunately, LLMs have a habit of getting stuck in recursive loops. Quants complain about how their storytelling machines “tend to lock-in a framework almost obsessively, and how they then seem to bounce back and forth around the scenario grid based on a more narrow range of inputs”, the broker reports. "}],[{"start":121.64999999999999,"text":"Because of this tendency to get lost down rabbit holes, machine-learning systems are bad at guessing whether an investment cycle is over or just stuttering, says Jefferies."}],[{"start":131.14999999999998,"text":"From hereon, we got a bit lost."}],[{"start":133.54999999999998,"text":"Poetics is about how to tell a story. Aristotle outlines how tragedy should excite pity and fear by the actions of its characters, how the foundation of all good comedy is the dick joke, etc. It’s not explicitly about when to buy and sell shares, having been written nearly 2,000 years before their invention, but it does involve narratives. "}],[{"start":154.74999999999997,"text":"Quantitative investing also involves narratives, says Jefferies. Here they are:"}],[{"start":null,"text":""}],[{"start":159.49999999999997,"text":"Stick with us. "}],[{"start":160.69999999999996,"text":"Even to a bot, Aristotle’s rules of structure can separate what’s useful from what’s not. Through an Aristotelian lens, for example, a change in oil prices is an action in a defined world that changes a character’s perspective from ignorance to knowledge. The protagonist can choose how to proceed by signs (such as by checking the Brent price) or, more challengingly, by action and reasoning (such as by checking things downstream of the Brent price). Our hero’s intentions will flip-flop between opposites, resulting in the tragedy of extreme action, before reaching the sense of an ending. This, the broker says, is plot."}],[{"start":199.84999999999997,"text":"“The proper length is such that the sequence of events will admit of a change from bad fortune to good”, Jefferies quotes from Poetics, which it takes to mean that investors should instruct their LLMs to “extend the plot until the initial action matters”."}],[{"start":214.64999999999998,"text":"Character? “Know them by their deeds”, the team quotes, adding in its own words: “Key criteria: purpose, realism (lag effects), incentives, consistency”. Thought, embellishment and diction are treated similarly. Explaining the importance of clarity to diction as it applies to quants, it adds: “Present the type of model, accuracy, and key factors: f(ISM, PMI, OECD LEI): RF, 88% R-2, 80% OOB”."}],[{"start":244.09999999999997,"text":"There’s more! Lots more. But unless we’re mistaken, none of it references Aristotle."}],[{"start":249.59999999999997,"text":"Per Jefferies’ table above, the Genesis phase is when there’s a plausible reason to think the market is wrong, but “the thesis sounds slightly crazy when explained at a dinner”. Validation is when the thesis starts to show up in earnings, industry data and the actions of insiders. Adoption means the story takes hold of the mainstream and valuation multiples are outpacing earnings upgrades. Euphoria is when evidence is no longer needed, because the theme is regarded as inevitable. First Cracks are when disappointing data is being dismissed as transitory, but stocks fail to rally on good news and credit markets are weakening. Exhaustion hits with forced liquidations and the emergence of a competing narrative. "}],[{"start":290.99999999999994,"text":"AI-augmented research “alters every phase of this cycle along three vectors: speed, evidence composition, and failure detection. But it also introduces a specific pathology — thesis lock-in — that distorts the lifecycle in ways traders need to understand,” says the team."}],[{"start":309.94999999999993,"text":"Whereas Genesis to Adoption used to take between six and 18 months, AI-driven systems can get there in weeks. The Euphoria phase becomes shorter but more violent as models reinforce one another’s signals. And because models reposition quickly, First Cracks barely register. "}],[{"start":328.99999999999994,"text":"However, the critical risk with LLM-based analysis is that it aims to please. Once a thesis is encoded in prompts or workflows, the system will keep trying to find supporting evidence. It will rely on sources that previously confirmed the thesis. And it won’t go back over previously labelled data to try other interpretations. This means AI can keep a narrative alive past its expiration date, particularly in the later stages of the cycle, Jefferies says:"}],[{"start":null,"text":"

[T]he AI defends a thesis by reinterpreting negative evidence rather than identifying new positive catalysts, by extending goalposts, or by divergences between related equities. Competing narratives appear prone to being dismissed by AI unless cultivated and encouraged, but the tools also appear prone to missing when there is a separate causal mechanism that also supports the initial narrative (i.e. they are more prone to call a change in direction rather than waves of support with different underlying dynamics).

"}],[{"start":356.24999999999994,"text":"For spotting changes of theme and phase, analysts are better, the analysts conclude. AI research is handy for confirming and tracking trends within a frame, whereas humans “excel at recognising when the frame itself needs to change”."}],[{"start":371.84999999999997,"text":"In case this all seems a bit woolly, here’s where Jefferies’ chemicals team thinks we’re at:"}],[{"start":null,"text":""}],[{"start":377.4,"text":"“The compression of narrative cycles by AI probably will not eliminate the narrative cycle — it likely makes the cost of being late, or being locked into a stale thesis, significantly higher,” says Jefferies. “The asymmetry between early and late entrants is widening. If you’re not in by Phase 2 [Validation], you’re providing exit liquidity to AI-augmented funds,” it adds. "}],[{"start":400.45,"text":"“To perceive is to suffer”, Aristotle also said, anticipating sell-side research. "}],[{"start":415.2,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1781267317_6834.mp3"}