Macro Didn't Take Over. Calibration Cycles Broke.

Notes from the Eagle Alpha London panel on what's actually breaking in quant, and why it isn't the factors.

Earlier this week I was on stage at Eagle Alpha's Alternative Data Conference in London, on a panel titled "Markets, Macro & Risk: Rethinking Quant in a Macro-Driven World," alongside Mario Dell'Era of EnBW and chaired by Brendan Furlong of Eagle Alpha. The opening premise was that macro has become the dominant driver of returns across markets (equities, commodities, rates and FX), and that traditional quant models and signals are being tested as a result.

I'd push back on that framing. Macro didn't take over. Calibration cycles broke under faster regime transitions. That distinction matters because it changes what you should do about it.

The cadence problem

In early 2025, Goldman published research showing 74% of S&P 500 returns over the prior six months were micro-driven, well above the 20-year average of 58%. That's the opposite of the consensus framing. Stocks were moving on their own fundamentals.

What happened next is the actual story.

In the 14 months that followed, markets moved through four regime shifts. Liberation Day in April 2025: pure macro, policy-driven. The November 2025 AI capex panic: a narrative shock, $1.35 trillion of mega-cap tech wiped out in weeks on circular financing concerns. The February 2026 SaaS sell-off: another narrative shock, $285 billion erased on Claude Cowork launch day as the market re-priced build-vs-buy economics for software. And the Iran war this spring: back to macro, with oil rallying while gold fell.

Two macro, two pure narrative. All four happening faster than factor and risk models calibrated on years of historical data can adapt to. That's the real diagnosis.

Correlation failure ≠ factor failure

The next point I wanted to land: what's actually breaking isn't the factors themselves. Quality still means investing in profitable, low-leverage companies. Growth still means tech and innovation. Those themes work, and they'll keep working over long horizons.

What breaks, more often and more violently, is the correlation structure the models assume. Correlation between factors. Correlation between factors and macro. Correlation between asset classes.

A few examples.

In March 2020, COVID broke factor diversification. Momentum, Growth, Quality all sold off together. Realised correlations between them spiked far above what the models assumed, because the same names were crowded across all three factor books. Market-neutral quants who'd tilted toward high-quality, low-risk names lost money because in a broad market sell-off, all betas converged to 1.

In 2022, the rate hike cycle broke the bond-equity correlation that multi-asset portfolios had relied on for two decades.

In 2026, the Iran war broke the "gold rallies in wartime" pattern. The historical playbook said oil up, gold up, yields down, equities down. Three of four went the wrong direction. The reason: the market expected the Fed to stay hawkish in response, so nominal yields rose more than inflation, real yields rose, the dollar strengthened. That's what killed gold, not the war itself.

Each of these is a correlation failure, not a factor death. Diagnosing it the wrong way leads you to throw out the factors when what you actually need is to rethink the correlation structure they sit inside.

Crowding makes the unwinds worse

There's a second problem layered on top of correlation failure: factor crowding has accelerated sharply since 2015.

Factor investing has been democratised. Smart-beta ETFs, broker factor baskets, factor-tilted CFDs: what used to be the territory of dedicated quant shops is now the territory of fundamental managers hedging unintended factor exposures and multi-strats hedging tilts at the book level. Research on factor decay (McLean & Pontiff, 2016) suggests backtest alphas can overstate live performance by as much as 50%. When the unwinds come, they're more violent than they were in the 2010s because the crowd is bigger and the doors are smaller.

The multi-strats are an interesting case. Citadel, Millennium, Point72, Balyasny: collectively hundreds of billions of dollars run on roughly the same risk infrastructure (Barra, Axioma). At the platform level, the goal is a book that's factor-neutral in the Barra/Axioma sense. That's well-managed risk by traditional measures.

But by construction, multi-strats tend to hedge against similar types of risk and miss the same hidden themes. Private Credit. AI exposure. When those themes unwind, "perceived" neutrality can break violently. The Iran war was instructive: macro funds that were supposed to benefit from macro events took some of the biggest hits, with reports of energy PMs being cut at multi-strat platforms in the weeks that followed. Strategies built for the regime that just ended get punished hardest when the regime changes.

AI as the scenario-generation layer

Traditional stress testing has the same problem as traditional factor models. Stress scenarios are calibrated on historical data and impose a correlation structure that may not be relevant in the next regime. Take Iran: a traditional war-shock scenario would say rates down, equities down, gold up. All three were wrong.

When I was running risk at a multi-strategy hedge fund, we used to adjust Barra/Axioma stress shocks manually. Once you change one shock, the rest is messed up. Propagation across factors stops being internally consistent. It can take weeks of back-and-forth for a risk team to produce a realistic stress scenario for a new regime. Given the cadence we just discussed, weeks is too slow.

The workflow we're building at Medusa uses AI to do the iterative scaffolding work. You define a theme (say, a Strait of Hormuz closure). The AI generates a set of sub-scenarios with distinct factor profiles. It produces primary shocks and propagates them consistently across the correlation structure, iterating until internally consistent. The risk team reviews and adjusts.

This isn't science fiction. Axioma announced an AI-driven stress-testing capability earlier this year. We're actively building it on the Medusa platform.

What stays human is the judgement layer: which themes to model, which sub-scenarios to weight, which results to actually act on. Risk should be a research tool, not a compliance check.

The missing role: Context Engineer

We didn't have time on stage to fully land this last point, but it's the part that matters most operationally.

The 2007 quant meltdown and the 2009 junk rally gave us "Quantamental": the idea that quants need context and discretionary PMs need quant infrastructure. That term was coined around 2015–2017. Alt data accelerated the convergence after 2015. In the last two years, AI has accelerated it again.

From the discretionary side, every PM is now a power user of data. They don't need to involve a data scientist to look at credit-card panels or web traffic. From the quant side, models can recalibrate faster, niche datasets are economically viable, and signals can be constructed bottom-up the way fundamental analysts construct revenue forecasts. This is the golden age of fundamental research at quant scale. Quantamental 2.0, if you want a name for it.

But making it work requires the discretionary framework to be machine-readable. Not the old idea of encoding the investment process into an algorithm. That scares discretionary PMs and doesn't work. I mean feeding heuristics, regime priors, and override decisions into AI as structured context. Why we broke a risk limit. Why we overrode a trade recommendation. What market signals matter in which regime.

The question is: who does that work?

I call this role the Context Engineer. Part librarian: knows how to organise knowledge and make it machine-readable. Part investor: knows what matters for markets and when, with deep enough investment context that they don't get this wrong. Part engineer: knows how AI works and how to ship systems.

These are not data scientists. Data scientists build models. Context Engineers organise data and feed it into models.

2025 was the year everyone tried to build agents. 2026 is the year we realised context is the bottleneck.

What this means for lean investment teams

If you're running a small investment team, you can't outspend the multi-strats on Barra infrastructure or hire a 20-person risk team to manually recalibrate when the regime shifts. The good news is you don't have to.

The leverage is in AI agents sitting on top of your existing stack: your Excel sheets, your broker files, your Bloomberg feeds, your existing Python. Calibration cycles compress from weeks to days. Stress scenarios get generated against the current regime, not the last one. The discretionary judgement stays where it should, with you.

That's the version of "modernizing quant" that works for a lean fund.

If you want to talk about how regime-aware risk and stress testing can work on your stack, get in touch.