What limits a research team when a piece of factor analysis that once consumed a morning of a quant’s time collapses into a single prompt and an answer? A recent collaboration between Axioma by SimCorp and the AI firm EDS brings the issue into focus. The answer it points to is no longer the cost of computation, but the choice of question.
The study itself is narrow by design. Axioma fed roughly four decades of its own style-factor returns – 19 factors running back to the 1982 inception of its US trading-horizon model – into EDS’s AI engine, and asked how those factors behaved in the run-up to 15 historical market peaks. The firm reports two durable signals: Medium-Term Momentum and Liquidity both produced consistently above-average returns as the market approached a top, while Downside Risk and Dividend Yield returns deteriorated. Melissa R. Brown, Senior Director of Applied Research at Axioma by SimCorp, who led the work with EDS Chief Data Scientist Ben Lieblich, tells Market & Alt Data Insight that the result tracks long-held intuition about how tops form.“I’ve been working with factor investing for a long time,” Brown says. “It always seemed hard to prove, but it felt as though factors generally started to fall apart before the market made a big move, one way or the other, up or down. I think you move from a period of rational investing – which is why factor-based investing works – to something more irrational, where people stop buying value, or stop buying momentum. That shifting mindset shows up first in the factors, and then eventually in the market.”
Factor data is well placed to test a claim of that kind. What the AI overlay changes is how the testing gets done.
What Becomes Askable
Lieblich’s framing, published alongside the research, locates the value in what gets attempted rather than in the hours saved: portfolio managers begin asking the questions they previously skipped because the friction was too high. Brown puts the same point in operational terms.
“Even the basic question we asked in the paper – which factors have materially different performance at the onset of a market peak – would be one,” she says. “But there are so many. Okay, then what happens in the next 60 days? That’s where EDS’s system came in: rather than having to pull up all the data again and recalculate everything, the AI can just go and answer it. It’s not only market peaks – it’s ‘is this factor turning around?’, ‘where do we stand in the AI trade?’, or one that really interests me: how many times historically have you had a sector that’s 35 per cent of the market, and what happened subsequently? The questions are limited only by your imagination, and it’s not as though every one takes two days to answer. It’s five minutes.”
For a research function, that is a genuine change in the unit economics of curiosity. The follow-up question – the second-order test that used to be deferred because the first one had already cost a day – becomes cheap enough to run on a hunch. SimCorp itself describes the resulting shift as one from “can we do this?” to “what do we want to ask?”, which relocates the binding constraint from infrastructure to judgement.
It also relocates the risk. When the cost of asking falls to near zero, nothing external stops a researcher running a dozen variations of a question and reporting the one that confirms an existing assumption. The discipline that friction used to impose – the implicit filter that forced a researcher to justify the compute before spending it – has to be supplied deliberately instead. Brown is candid that the study as published is a partial view: it examined only the pre-peak window.
“We only looked at the before,” she says. “But knowing how strong the performance was – at least for liquidity and momentum, it was strong, consistent and so much higher than average – I think you have to assume it’s going to be much worse than average afterwards.” Whether the factors behave differently after a peak than before it – the test that would distinguish a genuine signal from a pattern that simply persists either side of the top – is left for later work.
Why the Data Layer Carries the Weight
If the AI makes questions cheap, what makes the answers trustworthy sits underneath it. Axioma’s case is that the engine is only as good as the factor data it runs on, and that the defensible asset in the collaboration is the four-decade, production-grade model history rather than the AI layer that queries it. So what should a buy-side team actually verify?
“That’s really two questions,” she says. “One is: did the AI get it right, or did it hallucinate something? That was relatively easy to check against the live data we have and know to be correct – that’s the business Axioma is in, so if our data weren’t correct we’d have really big problems. The other area of data integrity is harder. If I’m buying a company because I think its earnings are going to grow, I want to know how they’re going to grow. If I’m buying a stock because it’s cheap, because it has a high value score, maybe I want to investigate a bit more – does it deserve to be cheap, or are investors overlooking something?”
For any buy-side team weighing tools of this kind, checking whether the model hallucinated is the tractable part, resolvable against a known-good dataset. Checking whether the underlying data deserves to be trusted in the first place is the part that does not automate – and it is where the substrate, not the model, determines the quality of the output. SimCorp’s own account of the research makes the dependency explicit: without the data integrity behind the factor history, it could not trust the answers the AI produced.
That positions the underlying data, not the AI, as the scarce input – an argument consistent with the strategy visible across SimCorp’s recent moves, from the Axioma Factor Library launch to its Axyon AI work, where proprietary model output is increasingly treated as a research asset in its own right. For practitioners evaluating AI research tools, the practical implication is to interrogate the data layer with the same rigour as the model: what is the provenance of the factor history, how is it tested, and is it production output or a more generalised training set.
The Limits of the Signal
The research does not claim to time the market, and neither does Brown. The single peak the momentum signal failed to anticipate was the dot-com top of March 2000 – a caveat that carries weight given how tech-concentrated today’s market is – and even the AI involved in the study noted that the patterns it found are better suited to characterising market regimes than to timing them. Asked how she would want portfolio managers to use the work, Brown draws the line firmly on the side of positioning rather than prediction: timing the market is difficult, she says, but an investor who is starting to worry can diversify further, hold more names and bring down overall portfolio risk.
That restraint is also where the data-workflow story and the market-commentary story part company. The findings are interesting as a demonstration of what becomes researchable when friction collapses; on 15 in-sample observations they are not a top-calling tool, and Brown does not present them as one. What carries forward is methodological. The fingerprints, as SimCorp puts it, were always in the data. What changed is how cheap it became to go looking for them – and, with that, where the real work of research now sits.
“The AI check is relatively easy,” Brown says. “The rest is why people hire discretionary managers and fundamental analysts – that’s what they’re there to do.”
Subscribe to our newsletter
