The world that didn't happen

Uncertainty, causality and decision intelligence

1. The world that didn't happen#

Why high-stakes decisions need causal, counterfactual and self-aware intelligence, and why language models are not it

Black swans, Pearl's ladder of causation and action-to-outcome decisions, mapped to markets, CEOs and chief investment officers, and turned into fifteen requirements for AI that decides under uncertainty.

Summary

2. In brief#

Markets, boardrooms and investment committees are complex, partially observable worlds that change as people act in them. Deciding well there means reasoning about uncertainty that cannot be removed, causes that cannot be seen directly, and futures and pasts that never happened. That requires a kind of intelligence most AI does not have.

Introduction

3. Deciding in a world that could have been different#

Every high-stakes decision, from a CEO's acquisition to a trader's position or a pension fund's hedge, is taken in a world that is uncertain in more than one way, whose causes are only partly visible, that responds to the action itself, and in which the most dangerous paths are the ones that have not happened yet. The data such a decision rests on is never wrong in itself. It is simply the record of one history, and every backtest, risk report and review that confirms a set-up confirms it against that one history.

This paper sets out what such decisions demand. Sections 1 to 3 describe the uncertainty: its forms, black swans, the realities that did not happen, and Pearl's ladder of causation. Sections 4 and 5 describe the decision: an action that changes the state of the world in a specific context, and the four-step intelligence flow, Perceive, Diagnose, Predict, Recommend, we use to reason about it. Section 6 shows where today's finance paradigms fall short: mean-variance, the bell curve, asset buckets, bull and bear labels, few dimensions, three scenarios, correlation, numbers only and straight lines. Section 7 turns to AI and sets out fifteen properties that intelligence in uncertain, complex environments requires. Section 8 explains why large language models, however fluent, are the wrong foundation for this task.

01

Uncertainty

4. The forms of uncertainty, and the black swan#

Calling everything “risk” hides the fact that different kinds of uncertainty call for different responses. Mixing them up is one of the most expensive mistakes in high-stakes decisions.

Aleatoric uncertainty is chance: randomness in the world that more data describes better but never removes. Epistemic uncertainty is ignorance: what we do not know yet but could learn. Unknowable uncertainty is where neither the outcomes nor their probabilities are known (Knight 1921); Dellermann (2017; 2020) calls this unknowable risk, because the data needed to estimate it does not yet exist.

For an AI system this taxonomy has a second dimension: what the uncertainty is about. There is uncertainty about the state of the world (what is true now, much of it unobserved), about the mechanisms of the world (which causes drive which effects, and whether they still hold after a regime change), and about the model itself (whether its parameters are right and whether the current situation lies inside the range it was built for). A system that reports one confidence number blends all of these, and the decision maker cannot tell whether to hedge, investigate or step back (Kendall and Gal 2017; Hüllermeier and Waegeman 2021).

Black swans#

Nassim Taleb (2007) defined a black swan as an event that lies outside regular expectations, has an extreme impact and is explained convincingly only after the fact. The stock market crash of October 1987, the collapse of Long-Term Capital Management in 1998, the global financial crisis of 2008 and the pandemic shock of 2020 all fit. Two features matter for decision systems.

The first is the turkey problem. A turkey fed every day for a thousand days grows more confident with each meal that the farmer means well; its confidence peaks on the day before Thanksgiving (Taleb 2007). Strategies that sell volatility, run carry trades or use leverage have the same shape: long runs of small, steady gains that make the model look validated, followed by one loss that erases them. The data confirmed the model right up to the moment the model failed.

The second is fat tails. Bell-curve models drastically understate how often extreme moves occur. At the same volatility, a normal distribution expects a daily move of five standard deviations about once in seven thousand years; a modestly fat-tailed distribution expects it roughly once a year. Real markets behave far more like the second.

02

The paths not taken

5. The realities of the world that didn't happen#

Every price series, set of accounts and track record is the record of one history. Many others were possible, were just as consistent with the underlying causes, and left no data.

A crisis that nearly occurred, a merger that nearly failed, a rate decision that could have gone the other way: these are real parts of the risk a decision maker faced, even though they never appear in a database. Taleb (2001) called them alternative histories. Statistically, history is a sample of size one from the distribution of possible worlds, and decision makers routinely treat that one sample as if it were the distribution.

The consequence for AI is direct. A model trained only on the realised path learns its luck and calls it structure. To judge decisions by their quality rather than their outcome, a system must be able to generate the histories that did not happen, from a model of the causes, and evaluate decisions across all of them. That is a causal and counterfactual task, which is where Pearl's ladder comes in.

03

Causality

6. Pearl's ladder: seeing, doing, imagining#

Judea Pearl distinguishes three levels of causal reasoning. Each level answers questions that the levels below cannot, no matter how much data they have (Pearl 2009; Pearl and Mackenzie 2018).

Rung 1, association, is about patterns in observed data: when X is seen, how likely is Y? Almost all of statistics and machine learning, including large language models, lives here. It is powerful for recognising regularities, and blind to why they hold.

Rung 2, intervention, asks what happens if we act: if we raise prices, add duration or acquire the target. The answer differs from rung 1 because acting changes the system. Firms that raised prices in the past may have done so because demand was strong, so the observed association between price rises and revenue says little about the effect of our own price rise. Confounding is the rule in business and markets, not the exception.

Rung 3, counterfactuals, asks what would have happened had something been different, given what actually did happen. Was the fund's return skill or luck? Would the deal have failed without the cost cuts? Would the drawdown have been smaller with a different hedge? These are the questions behind attribution, accountability and learning from experience, and they require a causal model of the world combined with the facts of the specific case.

Why causality matters when only one path is observed#

A dataset is one draw from the worlds that could have happened. In that draw, many variables move together: some because one causes the other, others because of the particular conditions of that period. A correlation model cannot tell the two apart, and it gives weight to whatever predicted best in the one history it saw. A causal model separates them. Causal mechanisms are invariant: the effect of a cause on its outcome holds across environments, while incidental links change from one world to the next (Peters, Bühlmann and Meinshausen 2016; Schölkopf et al. 2021). Economists know the same point as the Lucas critique: relations estimated under one regime break when the regime, or the policy, changes (Lucas 1976).

That is why a causal model is more robust than correlation-based and traditional quantitative models. They are tuned to the one path that was observed. It is anchored in the mechanisms that generate all the paths, including those that did not happen. The figure shows the trade-off: in the recorded history, the correlation model looks better; in the worlds that did not happen, it is the one that fails.

Why this must be addressed, and how#

04

The decision

7. Action to outcome: a decision changes the world#

Most AI is built to answer “what is this?” or “what will happen?”. A decision answers a different question: “if I do this, in this situation, what will happen to the things I care about?”

Lorien Pratt, a pioneer of decision intelligence, frames a decision as a chain from actions (the levers a decision maker controls) through intermediates to outcomes, influenced along the way by externals that are not under the decision maker's control, and judged against goals (Pratt 2019). Her causal decision graph makes that chain explicit, so that data, models and human expertise can each be attached to the link where they belong.

Three properties separate an action-to-outcome decision from a diagnosis or a prediction.

Question

Example

Step in the intelligence flow

Output

What it needs

Effect on the world

Pearl's rung

Needs context, goals, constraints

Can be judged by accuracy alone

Typical failure

In a hedging decision

05

The intelligence flow

8. Four steps from context to action#

A decision is an intervention in a specific context, so it cannot be produced in one step. Our intelligence flow has four steps, Perceive, Diagnose, Predict and Recommend, each answering its own question. Each step is one component of every analysis.

01

Perceive · Perceive the context · What should we look at? The financial universe is far too vast to analyse in its entirety. No amount of data or compute will capture every ticker or variable, so more data is not the answer; decision-centric scoping is. Defining the context starts from the transaction or valuation decision itself: establishing the underlying ontology, selecting the true peer set, and isolating the specific signals that bear directly on the outcome.

02

Diagnose · Causal diagnosis of context and mechanisms · What are we looking at, and how do the mechanisms work? Two diagnoses. First, the state: what situation is this, what is driving current outcomes, which exposures are really the same bet under different labels. Second, the mechanisms: which causes drive which effects, how strongly, through which intermediates, and whether those relations hold across regimes or flip. This is the step from Pearl's first rung towards the second: from what moves together to what causes what.

03

Predict · Counterfactual prediction · What will most likely happen, what could happen, and what would happen? Three questions, three kinds of output. What will most likely happen is a calibrated forecast, as a distribution. What could happen is the full range, including tails, black swan stress paths and imagined futures that have no precedent in the data. What would happen is the counterfactual: the outcome under action A versus action B, and what would have happened had we acted differently in the past, which is Pearl's second and third rungs.

04

Recommend · Contextual optimisation · What should we do, given our context, constraints and goals? Search the space of possible actions for the one that best serves the decision maker's goals across the simulated futures, within hard constraints such as leverage limits, liquidity, mandate and risk budget. The answer is contextual: the best action for this firm, in this regime, with these goals. Trade-offs are made explicit, so people can choose, own and explain the decision.

06

The legacy toolkit

9. Where today's finance paradigms fall short#

Most of the analytical toolkit in finance was built for a world of scarce computing power. Each tool was a sensible simplification at the time. Together they assume away exactly the features of the world described in this paper: fat tails, causes, regimes, high dimensionality and paths that did not happen.

Symmetric risk and the bell curve#

Mean-variance optimisation, the foundation of modern portfolio theory (Markowitz 1952), measures risk as variance. Variance treats a 10% gain and a 10% loss as equally bad and is blind to the shape of the distribution. Two assets with identical mean and volatility can carry very different risks of ruin. Combined with the assumption of normally distributed returns, it understates the tails that black swans live in (Section 1).

Labels instead of drivers#

Portfolios are organised in asset-class buckets, and companies in sector and style labels. Similarity is judged by the label. But labels are administrative; risk is driven by underlying factors such as global growth, real rates, liquidity and inflation. Five buckets can be one bet.

Bull or bear, instead of nested regimes#

Market commentary reduces the state of the world to one word. Real decisions face several cycles at once, each with its own rhythm: the economic cycle, the sector cycle and the lifecycle of the company itself. The same earnings miss means something different for a maturing company in a sector trough during a late expansion than for a growth company in an upswing. Regime awareness means modelling these layers and their interaction, not a single switch.

Few dimensions, three scenarios, straight lines#

Screens and models compress firms into a handful of ratios, the future into a base, bull and bear case, and relationships into straight lines. Each compression is convenient and each loses exactly the information that matters most under uncertainty: the features that make two similar-looking firms different, the probability mass beyond the bear case, and the thresholds where small changes trigger large effects.

Correlation, numbers only, and a static past#

Three more simplifications complete the toolkit. Each looks harmless on its own, and each fails in the situations that matter most.

Correlation

Correlation-based tools assume that past co-movement will persist and is informative about cause; Section 3 showed why neither holds. The relation between stocks and bonds is the standard case. It was positive through much of the 1980s and 1990s, negative for two decades after 2000, and positive again in 2022, when stocks and bonds fell together. A hedge learned in one regime fails in the next.

Numbers only

Tools that only read numbers ignore the text, voice, networks and human judgment in which much of the signal arrives: the tone of an earnings call, a changed clause in a loan agreement, who lends to whom, an operator's doubt about an integration plan. A change in the world often shows up in these sources first and in the reported ratios last.

A static past

Models fitted once to the past are, by construction, tuned to a world that has already moved on. The error does not arrive in one break. It grows a little in every period since the fit, and stays invisible until a regime change exposes it.

07

The shift to AI

10. Intelligence in uncertain, complex worlds needs a different approach#

If decisions are interventions in uncertain, partially observable, causally structured worlds, then AI that supports them must be built for that, not adapted from systems built to complete text or classify images.

Fifteen properties follow from the argument of this paper. They fall into three parts: how the system sees the world, what it knows about itself, and how it decides, learns and imagines. Each serves one or more steps of the intelligence flow.

• INTELLIGENCE STEP · PROPERTIES IT DEPENDS ON MOST
• 01 Perceive context · Partial observability, multimodality, humans as a source, memory, fast adaptation
• 02 Causal diagnosis · Causal modelling, awareness of uncertainty about mechanisms, explainability
• 03 Counterfactual prediction · Probabilistic, causal and counterfactual, awareness of the model's own uncertainty, metacognition, imagination
• 04 Contextual optimisation · Vast state spaces, goal, context and constraint optimisation, planning, explainability, recursive self-improvement

Part 1 · Seeing the world as it is#

The first six properties decide what the system can represent and what it learns from: a world that is uncertain, causal, only partly visible and vast, observed through many kinds of evidence, including people.

Figure 23 A partially observable world. Data grows inside the inner circles; the outer circle stays largely hidden. Schematic.

Part 2 · Knowing itself#

The next two properties turn the system on itself: what it is uncertain about, and whether its own confidence is earned.

Part 3 · Deciding, learning and imagining#

The last seven properties turn a model of the world into action, and action into experience: deciding for the decision maker's goals, remembering, adapting, and imagining the data that does not exist yet.

09 · GOAL, CONTEXT AND CONSTRAINT OPTIMISED

Best for you, here, now

The objective is the decision maker's goals, risk preferences and hard constraints in the current context, not generic accuracy or a benchmark.

12 · MEMORY

Experience that accumulates

Every decision, its context, reasoning, alternatives and outcome is stored and retrievable. Institutional memory stops walking out of the door.

10 · EXPLAINABLE

Drivers, scenarios, trade-offs

Every recommendation traces to the causes, scenarios and trade-offs behind it, so people can challenge it, own it and answer for it to boards, clients and regulators.

13 · RECURSIVELY SELF- IMPROVING

Learns from outcomes and counterfactuals

The system improves its world model, its weighting of human and machine sources, and its own reasoning from results, under governance that keeps every change auditable.

11 · PLANNING

Decision lifecycles, not single moves

Decisions are sequential: stage an investment, keep an option, rebalance later. The system plans over a learned model of the world, as game-playing systems do (Schrittwieser et al. 2020), while respecting that the rules change.

14 · FAST ADAPTATION

Learning in context

Regimes change faster than models can be retrained. In-context learning lets a model adapt from a handful of new observations at inference time (Hollmann et al. 2025).

Figure 25 Fast adaptation: a regime change seen by a static model and by a model that learns in context.

Applied to the uncertainty described in Section 1, imagination does different work for each form. For chance, it fills the tails that the realised history under-samples. For ignorance, it shows which unknown matters most and so which evidence to buy first. For novelty, it cannot produce probabilities, but it can produce stress tests and early-warning signals for situations that have never happened, so that the first time is not a complete surprise.

08

A fair question

11. Why language models are pseudo-reasoning, and not the right foundation#

Large language models are extraordinary at reading, summarising and conversing. They belong in the toolkit as an interface. They are not a decision engine, because what looks like reasoning is a different process.

A language model is trained to predict the next token of text. When it produces a chain of reasoning, it generates text that has the form of reasoning seen in its training data. Often that text is right. But the process is association on Pearl's first rung, applied to words rather than to the world: there is no explicit model of causes, no calibrated distribution over outcomes, no search over actions against your constraints. The result is reasoning-shaped output, which is why we call it pseudo-reasoning.

Two objectives, and two rungs of the ladder#

The clearest way to see the difference is to ask what each system is built to optimise. A language model is trained to maximise the likelihood of the next token given the text before it. A decision system is built to choose the action with the best expected outcome under a posterior over the states the world could be in, given a mandate and constraints. These are not two routes to the same place. The first objective is satisfied by text that reads as the training corpus reads; the second is satisfied only by an action that performs well across the worlds that could happen. A model can score perfectly on the first and have nothing inside it that corresponds to the second.

In the vocabulary of Section 3, a language model works on the first rung of Pearl's ladder, and it works there on words rather than on the world. It learns which statements tend to follow which. The second and third rungs, intervention and counterfactuals, require a structural causal model: variables, the mechanisms that connect them, and the ability to hold one variable fixed and recompute the rest. Without such a model there is no do-operator (Pearl 2009), so the question “what happens if we act” cannot be evaluated; it can only be answered with the text that usually follows such a question. The same holds for “what would have happened had we not acted”, which Section 2 showed is the question every high-stakes decision turns on.

Five consequences follow, each the mirror image of a property in Section 7.

A sample from a text distribution is not a posterior over outcomes. When a language model answers, it draws one sequence from a distribution over text. That distribution is not a distribution over outcomes in your world, and it does not separate chance from ignorance (Properties 1 and 7). Asked for a probability, the model returns a number that reads like a probability; it was not computed from evidence about the world and it is not calibrated against anything in it.

The prompt is treated as the whole world. A decision system holds a belief over hidden state and updates it as evidence arrives (Property 3). A language model has no state beyond its context window: what is not in the prompt does not exist, what is in the prompt is taken at face value, and nothing is carried forward when the next observation contradicts the last. There is no partial observability and no belief state, only a document.

No search over actions under constraints. A recommendation is the result of evaluating alternatives against goals and hard constraints and planning over the decision lifecycle (Properties 9 and 11). Generation is not search: the model produces the most probable continuation, not the action with the best expected outcome within your mandate. Planning needs a model to plan in and a verifier to plan against, and the language model on its own has neither (Kambhampati et al. 2024).

Frozen in a moving world. The model's knowledge is fixed at training time. Regimes that arrive afterwards are invisible until the next retraining, which is far too slow for Property 14; and regimes that arrived before training leak into any historical test, which is the look-ahead problem in the evidence below.

Chain of thought is text produced alongside the answer, not the computation behind it. The intermediate steps are generated by the same next-token process as the conclusion (Wei et al. 2022). They can help the model and sometimes improve accuracy, but they are not an audit trail: the stated reasons are not guaranteed to be the operative ones, which is the opposite of Property 10. And because training and evaluation reward a confident answer over an admitted gap, the model is shaped to guess rather than to know what it does not know, which is the opposite of Property 8.

What the evidence shows#

Brittle under small changes. When the numbers in grade-school maths problems are changed, or an irrelevant sentence is added, accuracy of leading models drops markedly, consistent with pattern matching rather than formal reasoning (Mirzadeh et al. 2025).

Collapse with complexity. Models trained to “think” longer still fail beyond a certain problem complexity, and can reduce their reasoning effort as problems get harder (Shojaee et al. 2025).

Weak at causal inference. Asked to infer causation from correlational statements, models perform close to chance and do not generalise (Jin et al. 2024); their apparent causal knowledge largely repeats causal facts stated in text (Zečević et al. 2023).

Weak at planning. On planning benchmarks, models produce few executable plans on their own and need external verifiers to be useful (Kambhampati et al. 2024).

Explanations that are not the reason. Stated chains of thought can systematically misrepresent what actually drove the answer (Turpin et al. 2023).

Confident errors. Training and evaluation reward guessing over admitting uncertainty, which makes fluent, wrong answers a structural feature (Kalai et al. 2025).

Look-ahead bias in finance. Pretrained models can carry knowledge of later events into historical tests, making backtests look better than they are (Glasserman and Lin 2023).

Where language models belong in the flow#

Mapped onto the four steps of Section 5, the role is precise. In Perceive, language models read: filings, transcripts, contracts, news and expert notes become structured inputs, entities, drivers and claims with their sources, that feed the perceived context. In Recommend, they explain: the drivers, scenarios and trade-offs behind an action are turned into language a committee can challenge and a board can understand. They do not belong in Diagnose or Predict, where causes are identified and outcomes are simulated under intervention. Those steps run on an explicit causal model, calibrated distributions and a search over actions, and they must give the same answer twice. The boundary is the one drawn throughout this paper: language models work on text about the world; the decision is computed over a model of the world.

Conclusion

12. Deciding for all the worlds that could happen#

High-stakes decisions are made once, in a world that will be observed only once. That is exactly why they must be reasoned about across all the worlds that could happen.

The uncertainty is plural: chance, ignorance and the unknowable, with black swans hiding in the tails. The data is a single path of history, with the paths not taken missing and the data a novel decision needs not yet created. The decision itself is an intervention that changes the world, in a context, under constraints, for goals, and it needs the full intelligence flow: perceive the context, diagnose causes and mechanisms, predict what will, could and would happen, and recommend the action that best serves the goals within the constraints.

Intelligence equal to that task is probabilistic, causal and counterfactual; it knows that it sees only part of the world and how uncertain it is about the world, its mechanisms and itself; it uses humans as biased but valuable sources; it optimises for real goals and explains itself; and it remembers, adapts, plans and imagines. Language models can speak for such a system. They cannot be it.

A decision maker does not need a better forecast of the next shock. They need a system that reasons across the worlds in which it happens, sees the loop before it fires, and sizes the buffer for it.

We observe one world. We decide for all the worlds that could have happened, and all that still can.#

© vencortex®. This paper sets out a view on decision making under uncertainty. Figures marked as schematic, simulated or illustrative contain no performance claims. The portfolio state-space figure is computed for this series as log10[C(3000,50) × C(1049,49)]. It is not investment advice.