For most of modern market history, forecasting “worked” for a banal reason: the world stayed still long enough for your model to be right. In the 1990s, macro regimes—disinflation, globalization, falling real rates—often persisted 5–7 years, which meant you could build a thesis, size it, hedge it, live through a drawdown, and still have the regime intact on the other side. The forecast horizon and the regime horizon overlapped. That overlap was the hidden subsidy behind the entire culture of prediction: you weren’t predicting a chaotic system; you were largely extrapolating a slow-moving one.

Now flip the ratio. In the post-2020 environment, “macro regime duration” compressing to <12 months isn’t a narrative flourish—it’s a mechanical constraint. If a strategy’s research, deployment, and learning loop is 90–180 days (fast by institutional standards), a 12-month regime gives you 2–4 meaningful update cycles before the environment mutates. If alpha half-life is <90 days, you’re effectively running 4 quarters of decay per year; a signal that looked robust over 5 years can be functionally dead by the time your risk committee approves incremental capital. Forecasting becomes a race where the finish line moves faster than your stride.

Once you accept the regime math, the next break is reflexivity math. Markets are no longer “a thing you observe”; they are “a thing that observes you back,” at machine speed. In U.S. equities, estimates routinely place 60%+ of marginal flow in passive, systematic, and algorithmic execution once you aggregate index funds, vol-targeting strategies, CTAs, options hedging, and algorithmic trade scheduling. In that environment, your forecast does not sit outside the system. It enters as flow, shifts microstructure, triggers other models, widens spreads, alters volatility, and therefore changes the very variables you assumed were exogenous. The forecast becomes a causal intervention. The old worldview assumed prediction precedes action; the new one is that prediction is action. Directional correctness is no longer sufficient—you can be “right” and still lose money because your correctness reshapes the path.

The most concrete demonstration wasn’t theoretical. It was the 2010 Flash Crash, when nearly $1 trillion in market value evaporated in minutes and then reappeared, leaving behind official explanations that read like a tidy postmortem of a messy technical anomaly. The deeper lesson was structural: when algorithms respond to algorithms at timescales shorter than human comprehension, causation completes before oversight begins. Prices can recover. Epistemology does not. If the chain of events that moved the market cannot be narrated in real time, you are no longer operating in a predictive domain—you are operating in a control domain. Control domains do not reward forecasts; they reward understanding how interventions propagate.

This is where optimization quietly becomes the accelerant. Classical quant finance is optimization under constraints: maximize Sharpe, minimize drawdown, manage tails, diversify factors. Optimization assumes a largely passive environment—you can push harder on the objective function without the objective function changing shape underneath you. Modern markets punish that assumption. When many participants optimize similarly, they manufacture the same fragilities: crowded exits, liquidity mirages, volatility feedback, correlation spikes precisely when diversification is supposed to save you. In this world, “better optimization” is not an edge; it is a faster convergence to the same failure mode.

You can see the control problem numerically if you treat the market as a dynamical system with a shrinking reaction window. Suppose portfolios share a volatility-triggered de-risking rule. A volatility spike triggers 2× selling across a cluster of similarly managed books; that selling raises realized volatility, triggering another round, and so on. What once unfolded over weeks now completes in hours or minutes, because rebalancing, execution, and signal updates are automated. Insert policy into that loop. Central banks and fiscal authorities react to markets; markets react to policy expectations; the loop closes at machine speed. Forecasting a “terminal rate” becomes less meaningful than modeling the causal surface: how interpretation interacts with positioning, and how feedback turns intent into outcome.

This is why the institutions that actually survive stopped worshipping point forecasts. They ask not “where will EURUSD be,” but how policy actions reshape distributions, how those distributions interact with positioning, and how reflexivity alters impact. They allocate across scenarios rather than anchor to a point estimate, because point estimates are brittle under regime compression. This is not rhetoric. It is adaptation.

At a structural level, the shift investors are responding to can be expressed as a simple inequality. Let the effective regime horizon be R months, your model update cadence be U months, and alpha half-life be H days. A strategy is forecast-dominant only if R ≫ U and H ≫ execution + learning latency. In the old world: R ≈ 60–84 months, U ≈ 3–6, H ≈ 180–720 days for many styles. In the new world: R ≈ <12, U still often 3–6 (institutional reality), H ≈ <90 for many liquid signals. The inequality flips. When it flips, forecasting becomes narrative rather than edge. Causal control—scenario stress, intervention graphs, counterfactual testing—becomes the edge because it adapts to the inequality instead of denying it.

The timing is not philosophical; it is industrial. Simulation costs collapsed. GPU inference costs fell roughly 200× from 2014–2024; GPT-class inference costs dropped 50–100×. This is not merely “AI is cheaper.” It means scenario engines can run continuously instead of episodically. If alpha half-life is <90 days, a quarterly research cadence is already obsolete. You need systems that hypothesize, simulate, stress, and update weekly or daily. At the same time, decision complexity exploded. Real-world portfolio decisions routinely involve 500+ interacting variables—macro, flows, volatility surfaces, positioning, cross-asset correlations, funding markets, liquidity, and policy narratives—while the human working set remains closer to ~7 chunks. Intuition didn’t fail because humans got worse; it failed because the state space outgrew human bandwidth. Causation becomes a compression scheme: a way to reduce hundreds of variables into levers, pathways, and interventions.

So what does it actually mean to say markets are becoming “impossible to predict”? It means the predictive framing is mismatched to the object. A forecast tries to name an outcome. A causal system tries to model how outcomes change when actions occur. In a reflexive, machine-speed market, actions are not noise on top of fundamentals; they are part of the fundamentals. Your trade is an intervention. The crowd’s trade is an intervention. Policy is an intervention. Narratives are interventions. When a system is dominated by interventions, the relevant question stops being “what will happen” and becomes “what happens if.”

That is the causal threshold. You are no longer forecasting weather; you are steering traffic, and every driver has an engine.

Prediction does not disappear—it becomes a sensor. But the coordinating layer changes. In a machine-run market, you don’t win by being “right” about a number. You win by being right about how actions propagate: which levers matter, which pathways amplify, where feedback loops close, and where intervention risk hides. That is what it means to say causation crossed the threshold in finance. The market has become a control problem disguised as a forecasting problem—and the disguise is getting expensive.