Alpha decay is one of the foundational assumptions in quantitative finance. The empirical literature, beginning with McLean and Pontiff’s 2016 study of 97 anomalies and replicated and refined across multiple subsequent studies, has repeatedly found that returns degrade out-of-sample and degrade further once published. The assumption sits inside almost every institutional model risk framework as a working axiom: a signal that works today will, in some measurable way, work less well tomorrow.
That assumption is precisely what Jacopo Credi, Co-Founder and Chief Technology Officer of Axyon AI, is willing to question. He argues that the decay phenomenon, as commonly assumed, has not shown up empirically in the kind of framework his firm has built – and that the working axiom may be doing more to shape how the industry models its own performance than to describe how its models actually behave.“Alpha decay in models is an assumption pretty much everyone in this space makes,” Credi says, in conversation with Market & Alt Data Insight. “Sometimes I wonder how many people have actually sat down and tested it. Across multiple experiments over the years, I honestly haven’t found strong empirical evidence of a model decay phenomenon, at least for the type of problem we address and the modelling framework we use.”
The claim is narrower than it first sounds. Credi is not saying that decay is a myth. He is arguing that in Axyon’s specific operating environment – large-cap developed markets in the US and Europe traded at monthly-to-quarterly horizons through very large model ensembles – the decay pattern the literature documents has not been the framework’s dominant failure mode.
More room in the literature than is acknowledged
The mainstream finding remains intact. McLean and Pontiff is well-cited for good reason. But the body of research has accumulated qualifications directly relevant to the kind of framework Axyon runs.
The most consequential is geographic. A 2020 study by Jacobs and Müller, examining 241 anomalies across 39 stock markets, found that the United States is the only country with a reliable post-publication decline in long-short returns. None of the 38 non-US markets in the sample produced the same pattern. A second qualification, from Falck, Rej and Thesmar’s 2021 paper in Quantitative Finance, found that decay is more pronounced in large-cap names – because they are cheaper to arbitrage – than in small-cap names.For a firm operating in US and European large caps, those findings cut in opposite directions. The European exposure sits in the half of the literature where post-publication decay is harder to find; the large-cap focus sits in the half where decay should be more visible. The net is meaningfully less than the canonical US-small-cap decay finding would imply, but not nothing.
A third strand of recent ML-finance research has examined whether the construction of the model itself shapes decay outcomes. The finding, still emerging, is that ensemble methods built for partial uncorrelation behave differently from single-model strategies under regime change. Which matters here, because regime-agnostic ensemble design is not the architecture typically tested in the decay literature.
What Axyon is actually doing
“We don’t believe that markets can be solved once and for all with one big, beautiful model, because it’s the nature of the market – it’s a reflexive system that changes depending on what you know about it,” Credi says. “We think of it as a learning game rather than a knowledge game.”
That premise has shaped the firm’s architectural decisions. “We don’t scrape or use particularly exotic sources like satellite imagery or radio-frequency measurements of point-of-sale credit card receipts,” Credi says. “It’s mostly good old data used by traditional quant models: prices, volumes, fundamentals, estimates on those fundamentals, macro indicators, commodities. The alpha we extract is not really in single data items but in how they are treated.”
Credi’s framing pushes back against the dominant narrative that AI-driven alpha generation depends on novel alternative data sources. “The raw material is data, and that’s key, but the raw material itself is not the value added,” Credi says. “The value is everything after that: how you process it, and how the machine uses it to find patterns predictive of future relative returns.”
The processing layer is where the framework’s defence against decay is built. Axyon runs very large ensembles, with constituent models engineered to perform well across many possible market conditions rather than optimised for any single regime. “It’s almost exactly what an asset manager does when building a portfolio,” Credi says. “You don’t select a single best-performing asset; you build a portfolio of many different assets that are at least partially uncorrelated.”
Axyon’s ensembles are built to be regime-invariant by construction. Credi is sceptical of the regime concept itself. “The highest-performing model today might not be the best-performing model tomorrow, but it might well be the best-performing model again five years from now,” he says. “Financial practitioners and researchers would call these market regimes. To me, that’s a human way of making sense of something very complex by assigning a label. Models don’t know they’re in a regime; they’re just statistical learners.”
This implies that if model performance varies cyclically rather than monotonically – better, then worse, then better again, depending on which conditions an ensemble’s constituent models are calibrated to – then what looks like decay over a short window may not be decay at all. The whitepaper Axyon published together with SimCorp in March, with its 10-year backtest on US All Cap equities, bears on this directly. The strategy showed consistent positive returns from 2015 through August 2024, followed by what the paper describes as “a substantial drawdown through the rest of 2024, a short recovery period and then another smaller shortfall through the end of the test period.” On the conventional framing, that drawdown is the empirical signature of decay arriving. On Credi’s framing, it is the ensemble passing through a market environment its constituent models are less calibrated for. The data does not adjudicate between the two interpretations.
Where the black box gets opened, and where it doesn’t
The decay question presupposes that buyers can see enough of what a model is doing to detect decay if it occurs. That puts pressure on the explainability layer, and on the line between what a vendor will disclose and what it considers proprietary.
Axyon’s response is a tool called Axyon Lens, a set of decompositions built on SHAP values that connects model inputs to model outputs. The output that goes to clients is filtered. “There’s a manual step in the process – effectively a manual filter – that masks, for example, the name of a variable or the way it’s constructed, because that’s considered the company’s IP,” Credi says. “We can use placeholders for specific definitions of a variable.”
The level of explainability Axyon Lens offers is, in the firm’s account, sufficient for clients to satisfy internal governance. Feature-level identification is not part of what is disclosed – that part of the framework sits behind the IP boundary.
The crowding question, still open
If the decay argument is about whether a signal degrades over time as conditions change, the crowding question is about whether it degrades because too many people are using it. The two are sometimes conflated; they should not be. Crowding decay is the more relevant risk for a signal distributed at institutional scale through SimCorp One, which provides SimCorp clients with access to Axyon’s predictive analytics inside the platform’s portfolio construction infrastructure.
Axyon’s response is built into the architecture rather than managed through ongoing capacity controls. “The same signal stream, the same information content, can be implemented differently by different clients – rebalanced monthly, quarterly, at different frequencies, across different slices of the stock universe,” Credi says. “The degree of freedom in sizing and slicing the information provided to clients is in itself a barrier to signal degradation.”
SimCorp’s partner ecosystem strategy, which has accelerated through 2026, is in early stages of placing third-party AI signals inside institutional portfolio workflows. The question of what happens to the signal when adoption scales further is, for now, theoretical.
Where this leaves things
Robust empirical evidence of model decay exists across a wide range of strategies, particularly in the US and particularly in the most-arbitraged segments of the market. Credi’s position – that within Axyon’s specific operating envelope, the assumption has not behaved as expected – sits in a different category. The international finding from Jacobs and Müller is consistent with it. The emerging ML-finance literature on ensemble robustness is consistent with it. The methodological scaffolding behind the claim – very large ensembles, regime-agnostic design, slow time horizons in liquid markets – is internally coherent.
A buyer who treats decay as an inevitability will allocate, monitor and exit differently from a buyer who treats it as one possible failure mode among several. The framing is a methodological position that institutional governance frameworks rarely make explicit – and possibly should.
“We don’t work on models – we work on improving the factory of models,” Credi says. “The factory is far from perfect. It’s genuinely like an assembly line, with steps that are automated, data-driven and scalable.”
For institutional buyers, the next five years of out-of-sample performance will determine whether the framework’s defence against decay holds up. For the rest of the AI signal market, the question Credi has raised – whether the decay assumption survives serious scrutiny – is one that has not yet been seriously asked.
Subscribe to our newsletter
