In February 2025, Goldman Sachs told clients the US equity market had become a stock-picker’s market: 74% of the typical S&P 500 stock’s return was being driven by company-specific factors rather than macro forces, against a 20-year average of 58%, and the bank expected that micro-driven environment to persist. Within weeks, sweeping tariff announcements had macro dominating everything again. The call did not survive a month.
That reversal framed the panel “Markets, Macro & Risk – Rethinking Quant in a Macro-Driven World” at the A-Team Group/Eagle Alpha Alternative Data Conference London – moderated by Brendan Furlong, Chief Data Advisory Officer at Eagle Alpha, with Mario Dell’Era, Senior Quantitative Market Risk Manager at EnBW, and Petr Merkuryev, Founder of Medusa Investment Partners. The panel had been convened to ask whether macro has become the dominant driver of markets. Neither practitioner disputed that macro matters; both declined the framing, relocating the real driver elsewhere.
The Argument Was Never Really About Macro
What looks like macro dominance, one panellist argued, is better understood as regime change – and not all of it macro. The past year delivered roughly four regime shifts: a tariff shock, a panic over hyperscaler capital spending that briefly collapsed the mega-caps, a separate sell-off on fears that AI had made the software-as-a-service licensing model obsolete, and a geopolitical conflict. Two of the four were AI-driven, not macro-driven. On this reading, AI now explains close to half of S&P returns – a figure offered as a practitioner estimate rather than a measured one – so the question is not whether macro is in charge, but how often and how violently the underlying regime resets.
A second line of argument reached the same destination from the commodity and energy markets, where the driver is regulation, transition policy and the physical build-out of new generating capacity – slower forces than a tariff headline, but ones that change the rules under which prices form. What both descriptions shared was an emphasis on the speed of change rather than its source. It is the rate at which the environment resets, not the label on the cause, that breaks models calibrated to the previous state.
What Actually Fails Is the Correlation Structure
The sharpest claim of the session concerned what breaks when a regime turns. The individual factors, one panellist maintained, do not stop working. Quality still means long profitable, unleveraged companies and short unprofitable, leveraged ones. Growth still means long the innovators and short the cyclically exposed. The economic meaning of each factor survives. What fails is the correlation structure between factors – and between factors and macro.
The cause offered was crowding. Traditional factors have been democratised through smart-beta ETFs, through the broker baskets that multi-strategy funds use to hedge factor exposures, and through the sheer weight of capital now sitting in the same trades. A decade ago the same multi-factor strategies were less crowded and, when they unwound, less violent. Now the capital arrives from too many directions at once, so when the unwind comes, the diversification that was supposed to sit between factor models collapses, because every factor sells off together.
COVID-19 showed the mechanism twice. As the pandemic hit, quant equity strategies were crowded into the same configuration – long growth, long quality, short volatility – so when the unwind began, every factor fell together and diversification across factor models offered no protection. Market-neutral books also stopped being neutral: a book built for zero beta but tilted towards high-quality names runs a long position, and when every stock falls together that tilt behaves like simple market exposure, producing directional losses the structure was meant to rule out. Then the recovery hit the same book from the other side. When vaccines were announced, the growth and stay-at-home names quants were long had already fallen, while the value, energy and travel names they were short led the rally. The same positioning was punished going down and coming back up.
Causation Arrives Too Late to Trade
If the equity argument was about structure, the risk-management argument was about knowledge. A model’s parameters, one line of reasoning held, measure how well its owner understands the market. When they stop holding, the problem is not just that the model needs recalibrating – its owner has lost knowledge of the market itself, because the market has become something else.
This is where correlation and causation part company. Causation is the weaker of the two and can only be identified after the event. By the time the new driver has been named – which asset, which process took over – the move has happened and the pain has been taken. Recalibration is possible, but only once the market settles into a new stationary period, and the gap between the old regime breaking and the new one stabilising is precisely the interval in which anything can happen and very little can be hedged.
You Cannot Backtest a Regime That Has Never Existed
The same problem recurs, in starker form, in markets undergoing physical transformation. Grid-scale batteries are reshaping power markets in ways that have no historical precedent to test against. The effect is already visible in heavily renewable systems such as California, where intraday power prices have changed shape: the spread between peak and trough has narrowed as flexibility enters the market, yet the average price has not fallen, because the owners of the gas-fired flexibility that remains can charge more for it. A narrower spread but a firmer average was not the intuitive expectation – and there was no past in which to find it. A backtest assumes the future is drawn from the same distribution as the past; when the asset base itself is new, that assumption does not hold, and the training data underpinning both alpha research and risk modelling describes a world that no longer exists.
The proposed response was regime-aware testing, run fast. A stress test encoded against last quarter’s correlations is obsolete before it finishes running; a scenario for a fresh geopolitical shock has to be built in days, not weeks. Building one by hand is slow, because each shock propagates through the book in ways that force the next adjustment, and the back-and-forth can absorb a week. AI can compress that to minutes, generating a scenario grid a human team then reviews. On the long-horizon side, where there are no market prices at all – financing a wind farm against a power-purchase agreement decades out – the same tooling generates the credible scenarios against which exposure and capital reserves are sized.
This remains a human-in-the-loop process. Models hallucinate, and judging in real time whether a move is noise to hold through or a regime change to cut rests on heuristics human practitioners have internalised and machines have not. The interesting claim was that those heuristics can increasingly be made explicit and handed to the machine – not to remove the human, but to let the process run autonomously between the points where judgement is actually required.
The Edge Has Moved From Code to Context
The closing thread pointed at where this leaves the people who build the models. Through the 2010s, the quant’s edge was technical fluency – coding in SQL, R or Python when many around them could not. As AI absorbs that work and natural language becomes a usable interface to backtesting and analysis, the durable edge shifts to domain knowledge: understanding how a market actually reacts, and holding the heuristics that separate noise from signal.
The provocative corollary was that discretionary and fundamental managers may now be better positioned for this blended world than pure quants – provided they stay fluent with data – because they are already practised at reading a messy, causally tangled market rather than only its numerical outputs. Whether that proves right or not, it matches what the audience itself flagged as the most pressing need: not better signals, but the ability to blend systematic and discretionary judgement as regimes turn over faster than any single model can track.
Subscribe to our newsletter
