Chapter 7
How do you decide well under uncertainty — and how would you know?
On November 2, 2021, Zillow's CEO Rich Barton did something executives almost never do: he stood in front of investors and killed his own flagship bet. Zillow Offers — the algorithm-driven business that bought homes directly, renovated them, and resold them — would be shut down entirely. "We've determined the unpredictability in forecasting home prices far exceeds what we anticipated," Barton said; the company, he told CNBC that evening, simply could not predict where prices would be in six months "within a narrow margin of error."12 The price of admitting it was brutal and public: roughly 7,000 houses to unload, about 2,000 employees — a quarter of the workforce — laid off, an $881 million loss on the business for the year, and a stock that cratered on the news.235 Weeks earlier, Zillow's model had still been buying homes at full speed; the housing market's pandemic whipsaw had turned the algorithm's confident price forecasts into systematically overpriced inventory.34
Zillow's chief rival made the opposite call. Opendoor had pioneered iBuying in 2014 and, when Zillow quit, chose to stay — the market was now less crowded, the mission unchanged. What followed was punishing: the company has never had a profitable year; when interest rates spiked in 2022 the model's economics collapsed; revenue fell roughly 55 percent in 2023 and another 26 percent in 2024; and by June 2025 the stock — once $34 — traded at 51 cents, facing delisting.78 Then came a twist no decision theory predicted: retail traders on social media adopted Opendoor as 2025's biggest meme stock, driving shares up more than 1,000 percent to nearly $11, financing a new CEO, returning co-founders, and a fresh war chest — even as the underlying business lost another $1.3 billion that year.678
So who decided well? The obvious answers keep flipping. Zillow looks wise — it stopped the bleeding — until you notice it also entered the business late, scaled fastest into the most treacherous market, and lost nearly a billion dollars learning what Barton then announced. Opendoor looks like a textbook case of throwing good money after bad — until a meme army hands it a second life that quitting would have foreclosed, and you remember that persistence through unprofitable years is also the biography of Amazon. The uncomfortable possibility this chapter takes seriously: you cannot grade these decisions by their outcomes at all, because in an uncertain world good decisions lose and bad decisions win all the time. Then how do you grade them?
The central question: how do you decide well under uncertainty — and how would you even know? The tools: the rational ideal and Herbert Simon's demolition of it; the heuristics and biases that bend judgment in predictable directions; prospect theory — how framing and loss aversion reshape choices; escalation of commitment, the science of not being able to quit; the conditions under which intuition deserves trust; and decision hygiene — the process disciplines that improve choices when no one can see the future. Zillow and Opendoor then return as a matched pair for the hardest judgment call in management: when stopping is wisdom and when it is surrender.
The rational model of decision-making — define the problem, identify all criteria and alternatives, evaluate each, choose the value-maximizing option — is the implicit standard behind every spreadsheet and business case. Herbert Simon's foundational insight, which helped earn a Nobel Prize, is that real decision-makers cannot follow it: information is incomplete, computation is costly, attention is scarce. Humans operate under bounded rationality, constructing simplified models of problems and satisficing — searching until an option clears a threshold of "good enough," then stopping.9
Bounded rationality is not an insult; it is an engineering spec. Satisficing is often optimal once search costs are counted, and the research on maximizing dispositions finds that relentless optimizers frequently end up with objectively better options and subjectively worse satisfaction. The managerial translation: match decision effort to decision stakes and reversibility — reserve full analysis for consequential, irreversible choices, and make the reversible ones fast. The model's limit is that it describes the capacity constraint on judgment but not its systematic distortions: Simon showed we simplify; the next section shows we simplify in predictable, exploitable directions.
Amos Tversky and Daniel Kahneman's 1974 paper in Science reframed decision research: under uncertainty, people rely on a small set of heuristics — mental shortcuts that are usually efficient and occasionally produce severe, systematic errors.10
• Availability — judging frequency and probability by ease of recall: vivid, recent, personal events loom larger than statistics.
• Representativeness — judging by resemblance to a prototype while neglecting base rates and sample size.
• Anchoring — estimates gravitate toward an initial number, even an arbitrary one; first offers and first forecasts set the terrain.
• Overconfidence and the planning fallacy — intervals too narrow, timelines too short, budgets too small; the inside view (this project's story) crowds out the outside view (how projects like this generally go).16
These are not quirks of careless people; experts show them in their domains, which is what makes bias a management problem rather than a personnel problem. The defenses are procedural: consult base rates before narratives (outside view), generate estimates before seeing others' anchors, and treat confident precision about the future — a home's price in six months, a project's cost next year — as a claim requiring calibration evidence, not a credential. Zillow's forecasting engine, note, was overconfidence industrialized: point predictions with narrow assumed error, scaled with real dollars.2 The caution on the other side: the biases literature can be over-applied — heuristics exist because they usually work, "debiasing" is hard, and post-hoc bias-spotting is itself an exercise in hindsight bias. Use the catalog to design decision processes, not to explain every failure after the fact.
Kahneman and Tversky's prospect theory replaced the rational model's stable preferences with three empirical regularities: people evaluate outcomes as gains and losses relative to a reference point, not as final states; losses loom larger than equivalent gains (loss aversion, roughly two-to-one); and risk attitudes flip with the frame — risk-averse when protecting gains, risk-seeking when trying to escape losses.11 The framing corollary is the theory's most managerial finding: logically identical choices described as gains versus losses produce different decisions. The implications thread through this chapter's cases and your career: sunk positions convert decision-makers into gamblers (a company "down" $500 million is in the loss frame, where doubling down feels like the prudent path back to even); negotiations and change proposals succeed or fail on whether they are framed against the right reference point; and the person who sets the frame has quietly made half the decision. Limits: prospect theory is a theory of individual choices under described risk; organizational decisions add politics, accountability, and repeated play, which can either amplify loss-framed gambling (careers are on the line) or dampen it (portfolios diversify).
Escalation of commitment is the tendency to invest additional resources in a failing course of action. Barry Staw's foundational experiment — "knee-deep in the big muddy" — showed the mechanism cleanly: participants who had personally chosen an investment that then failed subsequently allocated more money to it than participants inheriting the same situation, especially when they felt responsible for the initial choice.12
Escalation is overdetermined — four engines push the same direction. Self-justification: quitting concedes the original decision was wrong. Prospect theory: the loss frame makes the recovery gamble attractive.11 Sunk costs: resources already spent, which rationality says to ignore, dominate attention. Organizational amplifiers: audiences, rivals waiting to say told-you-so, and incentive systems that punish visible reversal more than quiet continued bleeding. The defenses are structural, and each shows up in the Zillow case: separate the deciders (have different people, or explicit pre-set triggers, decide continuation than decided entry); define kill criteria in advance, when you are still in the gain frame; and reframe exits as information victories rather than admissions. The essential caution: persistence is not a bias — most ventures require surviving dark stretches, and a discipline that always quits at the first loss would have killed every great company. The escalation literature's real lesson is not "quit sooner"; it is that the continue/quit decision should be made by the evidence and the pre-commitments, not by the sunk costs and the frame.
The field's great dispute — Kahneman's bias program versus Gary Klein's studies of expert firefighters and nurses making brilliant split-second calls — ended in a rare adversarial collaboration whose conclusion you can carry for life. Intuition is pattern recognition: rapid matching of the current situation to stored experience. It becomes trustworthy exactly when, and only when, two conditions hold: the environment offers stable, valid cues (regularities that actually predict outcomes), and the person has had prolonged practice with rapid, unambiguous feedback.13 Chess, firefighting, anesthesiology, and driving qualify — "kind" learning environments. Stock picking, long-range forecasting, hiring interviews, and predicting home prices through a pandemic do not — "wicked" environments where cues shift and feedback is slow, noisy, or absent, so confidence accumulates without accuracy. Two corollaries matter for the AI era you will manage in. First, the same conditions govern algorithms: a model trained on yesterday's regularities is an intuition machine, and when the environment's cue structure breaks — as housing's did in 2020–21 — the model fails exactly as a human expert would, but at scale and with more decimal places.213 Second, the subjective feeling of confidence is the same in valid and invalid intuitions; the question to ask is never "how sure do I feel?" but "was this environment capable of teaching me?"
If biases cannot be willed away and outcomes cannot be foreseen, quality must live in the process — a conclusion the field now backs with converging tools. Judge decisions by process, not outcomes: in any environment with luck, outcome-grading teaches the wrong lessons (punishing good bets that lost, promoting reckless bets that won); the discipline is to evaluate what was knowable and how it was weighed at decision time. The premortem (Klein): before committing, the team assumes the project has failed spectacularly a year from now and writes the history of why — legitimizing dissent that groupthink (Week 5) suppresses and surfacing risks the plan's authors are motivated not to see.14 The outside view: anchor forecasts on reference-class base rates before adjusting for this case's story.16 Noise reduction: Kahneman, Sibony, and Sunstein document that beyond bias, organizational judgment suffers noise — unwanted variability, with different judges (or the same judge on different days) reaching different conclusions from identical facts; remedies include structuring judgments into independent components, aggregating independent estimates, and delaying holistic verdicts.15 Limits and costs: hygiene is friction — reversible, low-stakes decisions deserve speed, not ceremony — and process can become theater that launders predetermined conclusions. The test of real hygiene is simple: does the process sometimes change the answer?
The pairing is close to a natural experiment on the chapter's hardest question. Same industry, same business model, same brutal environment; one firm exits, one escalates — sorry, persists; the word you choose is the analysis. If escalation theory fully governed, Zillow's exit is wisdom and Opendoor's persistence is pathology. If option-value logic governed, Opendoor preserved a lottery ticket Zillow tore up. The cases force the discipline Section 2.6 preaches: grade the decision processes with what was knowable at the time, and refuse to let the outcomes — including a meme rally nobody could have forecast — do the grading.
The entry deserves as much scrutiny as the exit. Zillow entered iBuying in 2019, years behind Opendoor, and in 2021 scaled aggressively into a market its own model could not read — doubling purchases quarter over quarter while pandemic dynamics were breaking the historical cue structure the algorithm had learned.48 This is Section 2.5's lesson in industrial form: a low-validity, regime-shifting environment defeats pattern recognition, human or machine, and Zillow's forecasting confidence — point estimates precise enough to bet a balance sheet on — was overconfidence with an API.213 By October 2021 the company held roughly 7,000 homes, many bought above what the market would return, and had already paused purchases, publicly citing contractor capacity.3
The exit, by contrast, is a clinic in anti-escalation. Every force in Section 2.4 pushed toward continuing: Barton's personal sponsorship (responsibility), $6 billion of Offers revenue that made the business look like growth (sunk momentum and frame), the humiliation of public reversal weeks after reassuring investors. The board and Barton overrode all of it within one quarter of the disconfirming evidence: full wind-down, a 25 percent workforce reduction, an $881 million annual loss taken openly, and a stated reason that named the epistemic failure itself — forecasting unpredictability "far exceeds what we anticipated."125 Barton framed the exit as learning, not defeat, and redirected the company to its profitable core.1 Honest scoring: the entry and scaling process fails the outside-view test (base rates on predicting prices through a regime shift were knowable); the exit process is close to the textbook — evidence-triggered, fast, complete, and reframed. One company, one year, both halves of the chapter.
Opendoor's continuation decision must be graded at several dates, and the grades differ. In 2021, staying as the pioneer with the deepest pricing experience while a rival's exit reduced competition was a defensible bet, not obvious escalation — Zillow's failure was partly self-inflicted scaling, and Opendoor's tighter spreads had survived where Zillow's had not.48 Through 2022–2024, the case darkens: rates spiked, the model's economics inverted, revenue halved and halved again (down ~55 percent in 2023, ~26 percent in 2024), losses continued a streak unbroken since the 2014 founding, and the stock fell from $34 toward delisting at 51 cents — a period in which every escalation engine (identity as the iBuying company, sunk reputation, loss-frame gambling for resurrection) was running, and no public kill criteria ever surfaced.678 Then 2025: a retail meme army, not the business, revalued the company — shares up more than 1,000 percent to nearly $11, a new CEO from Shopify, co-founders returning, fresh capital — while the operating results worsened (a $1.3 billion loss on declining revenue).678 The analytical trap is glowing in the dark: if you grade the 2022–2024 persistence by the 2025 rally, you are outcome-grading a lottery ticket. The rally rescued the company's financing; it is no evidence the continuation decision was sound when made — and the business the decision was about still loses money.6
| Zillow Offers | Opendoor | |
| Entry process | Late entrant; scaled fastest into regime shift; inside view + algorithmic overconfidence | Pioneer (2014); model built through multiple cycles |
| Response to disconfirming evidence | Full exit within one quarter; loss taken openly; reason stated in epistemic terms | Continuation through 2022–24 collapse; no public kill criteria |
| Escalation engines present | All present — and overridden | All present — identity, sunk reputation, loss frame |
| Cost / outcome to date | $881M 2021 loss; 25% layoffs; core business retained and profitable | No profitable year since 2014; $1.3B 2025 loss; meme rally refinanced the company |
| What outcome-grading would conclude | 'Quitter missed the meme rally' | 'Persistence vindicated' |
| What process-grading concludes | Poor entry process; exemplary exit process | Defensible 2021 stay; escalation-shaped 2022–24; verdict still open |
Three disciplined readings. First, the pair separates the two halves of decision competence that everyday language fuses: Zillow decided badly then well — the same leadership that overconfidently scaled also executed one of the cleanest evidence-triggered exits in recent corporate history, which is a reminder that hygiene is a practice, not a personality.112 Second, Opendoor is the field's best current classroom on why escalation is hard to diagnose from outside: persistence and escalation produce identical behavior and differ only in process — whether continuation was chosen against pre-set criteria and fresh base rates, or by identity, sunk costs, and the loss frame. Absent visible kill criteria, observers cannot tell, and often neither can the deciders.12 Third, the meme rally is this chapter's gift: an outcome so obviously disconnected from decision quality that it inoculates you against outcome-grading everywhere subtler — the promotion that followed a lucky bet, the prudent call that got someone fired. Decisions are bets; grade the bet, not the roll.15
Return to the central question: how do you decide well under uncertainty, and how would you know? Four conclusions travel.
Respect the spec. You are a bounded-rationality machine; so is your organization; so are its algorithms. Match analytical effort to stakes and reversibility, satisfice deliberately on the small stuff, and reserve full hygiene for the bets you cannot take back.9
Fight the frame, consult the base rate. Anchors, availability, and the gain/loss frame decide by stealth; the outside view and pre-anchored estimates are the working antidotes. Whoever sets the reference point has made half the decision — sometimes let that be you, and always notice when it isn't.101116
Pre-commit your exits. Kill criteria written in the gain frame are the only reliable defense against the loss-frame gambler you will become once you are down. Separate the deciding of continuation from the deciding of entry, and reframe stopped projects as purchased information.1112
Trust intuition — yours or a model's — only where the environment could teach it, and grade bets by process. Ask of any confident judgment: were the cues stable, was the feedback fast and honest? And ask of any outcome, before you learn from it: how much of this was the roll, not the bet? The premortem, independent estimates, and process-grading are how organizations answer those questions on purpose rather than in hindsight.131415
Notes appear as superscript numbers in the text and correspond to the numbered sources above. DOIs are provided where available; classic books are cited to their original publishers.