Rigour for strategic decisions
Writing № 02 · Judgement under uncertainty / I
№ 02.4
Writing
Part I / II
Series: Judgement under uncertainty · Part I

Heuristics & biases

When the data run out, decisions run on judgement, and judgement runs on shortcuts that fail in predictable ways. A field guide to the traps, and how to design around them.
Ian Hargreaves Series: Judgement under uncertainty Reading ~11 min Part I of II

Sometimes the data simply don’t exist, or they arrive as words rather than numbers. You have to put a first figure on the share of market a new product will take. You have to price the chance that some event in the news snarls your supply chain. The formal tools for deciding under uncertainty all demand probabilities as inputs, and when there is no dataset to read them off, those probabilities come from a person’s gut.

So the quality of the decision rests on the quality of the gut. How good is the gut? Reliably, predictably bad, not randomly bad, but bad in patterned ways that psychologists have catalogued for half a century. The good news hiding inside that sentence is the word predictable: an error you can name is an error you can design around. And the discipline matters even when you do have data, because judgement still decides how far to trust the numbers you were handed.

Before the list, one warm-up. Sarah ran the brand’s most divisive, most talked-about campaign of the decade. Here are two statements about her, one of which reads as more probable than the other:

The pull (a) Sarah works in marketing.    (b) Sarah works in marketing and personally championed that campaign.

Most people feel (b). But (b) is a subset of (a), every world in which (b) is true is a world in which (a) is true, plus a condition. (b) cannot be more probable than (a); it can only be less. The vivid detail made the story feel more likely while making it mathematically less so. That is the conjunction fallacy, and it is a fair preview of what follows: a shopping list of the shortcuts we all reach for, each one worth recognising before you next put a number on something uncertain.

iavailability

The availability heuristic

We judge how likely something is by how easily examples come to mind. Memorability is not frequency.
Availability · ease of recall

What comes to mind first feels common

The pull Asked how successful their campaigns are, most marketers produce a generous hit rate without breaking stride. Asked to name the ones that flopped, they go quiet: the flops take real effort to retrieve.

The flops didn’t arrive while the hit rate was being set. Wins are vivid, retold, attached to launches and awards; quiet failures are quietly forgotten. Because the wins come to mind more easily, the hit rate feels higher than the ledger would show. Rated again with the flops in the room, the number comes down.

Availability · ease of imagination

Easy to picture reads as likely to happen

The pull The ways this launch could win come effortlessly to mind, the press, the reorders, the case study, a reel that plays itself. An estimate of its success made straight after that reel runs high.

The fluency of the daydream leaks into the estimate: when a future is easy to imagine, we judge it more probable, even though vividness is not evidence. The trap runs the other way too, a risk that is hard to picture (a channel you’ve never seen fail) gets under-weighted precisely because imagination stalls. There is usually no real link between how easily a scenario comes to mind and how often it actually occurs.

Availability · illusory correlation

We ‘see’ patterns that the data don’t support

The pull “Every time we run a price promotion, premium churn spikes.” Pressed for the record, whoever holds that belief rarely knows how many promotions were not followed by a spike, or how many spikes arrived with no promotion at all.

The belief usually rests on two or three vivid coincidences. A promo ran, churn jumped, the pairing was striking, and it lodged. What never lodged: the dozen promos that passed quietly, and the churn months driven by something else entirely. Memorable co-occurrences are available; the non-events are invisible, so a relationship gets “seen” that the full record wouldn’t support, and a perfectly good tactic gets quietly retired on the strength of a coincidence.

The countermeasure is mechanical: fill in the four-box table, promo / no promo against spike / no spike. It does not immunise you, but it puts all four cells where the eye has to pass them, not just the one that caught it. (That discipline is exactly what its quieter cousin, biased assessment of covariation, demands; it’s the last entry on this list.)

iirepresentativeness

The representativeness heuristic

We judge probability by how much something resembles our mental prototype, and ignore the arithmetic underneath.
Representativeness · base-rate neglect

The vivid description drowns out the prior

The pull Asked the chance that this product, beautifully designed, adored in testing, a category first, goes viral within twelve months, people offer a generous number. Asked what share of all products, however good, actually go viral in a year, they offer a far smaller one.

The first number feels high because the description is so representative of a hit. The second, the base rate (the prior probability, in the language of the maths), is tiny, and it should anchor the first far more than it does. A glowing portrait pulls your estimate away from the base rate, which is exactly where it should stay rooted. Keeping the base rate written next to the story is half the countermeasure.

Representativeness · misreading randomness

We expect chance to look patternless

A lead takes ten weeks on average: call it a coin-flip whether any given lead lands short (L, ≤ 10 weeks) or long (G, > 10). Shown two possible runs of the next ten leads, most people judge the first far more likely than the second:

The pull (a) L G L G L G L G L G   or   (b) L L L L L G G G G G

They are equally likely, every specific ten-long sequence has identical probability. The neat alternation merely looks more random, so it feels more probable. Streaks in real data get read as signal (“something’s broken”) when they are exactly what randomness produces.

Representativeness · the gambler’s fallacy

We expect chance to correct itself

Six long leads in a row. Is the next one more likely to land short, because things are “due” to even out? No. The leads have no memory; the probability is whatever it always was. Chance does not keep a running balance and settle up.

Representativeness · regression to the mean

Extremes revert, and the intervention takes the credit

A company posts its worst year in a decade. The board fires the CEO, brings in a turnaround specialist, and performance recovers. The new chief is canonised, the book, the keynote, the circuit. But a catastrophic year is part skill and part luck: a market shock, a one-off write-down. The luck reverts on its own, regardless of who sits in the chair, and an average year would have followed almost any appointment.

The mirror is just as cruel: inherit a peak year, watch performance normalise, and get branded a disappointment for a reversion that was always coming. Leadership changes are triggered by extremes and then credited with the regression. Be suspicious of any before-and-after where the “before” was a record high or low, which questions a good deal of the turnaround-CEO mythology.

Representativeness · the conjunction fallacy

Adding detail makes a story feel likelier and be less likely

Strategy X has a strong track record. So does strategy Y. Which is more probable, that X succeeds, or that X and Y both succeed? “Both” can never beat “X alone,” because every world where both win is a world where X wins, minus the worlds where Y doesn’t. Each extra condition in a plausible-sounding plan makes it read better and bet worse. (This is the Sarah warm-up, in strategy clothes.)

iiianchoring

Anchoring and adjustment

We start from whatever number is in front of us and adjust, almost always too little.
Anchoring · insufficient adjustment

The first number spoken owns the room

“I reckon this lifts share by eight points.” The moment that lands, everyone else is adjusting from eight, and the adjustment is never far enough. Whoever speaks first, or writes the first cell in the model, sets the anchor the whole group drifts around. The defence is procedural: collect estimates privately, before anyone hears anyone else’s.

Anchoring · conjunctive events

We overestimate the odds that everything goes right

A strategy only pays off if A and B and C all happen. You rate them generously: P(A)=.90, P(B)=.85, P(C)=.80. Each looks safe, so the whole plan feels safe, people anchor on the comfortable individual numbers. But all three must hold at once, and independent probabilities multiply: .90 × .85 × .80 = .61, about 61 in 100, closer to a coin toss than the near-certainty the gut reports. The chain is markedly weaker than its weakest link, let alone its strongest. (The chart below.)

Anchoring · disjunctive events

We underestimate the odds that something goes wrong

Now the reverse. A project fails if any of A, B or C goes wrong, and each is unlikely on its own: .04, .02, .05. Negligible, surely. But “any” compounds the other way, the chance nothing breaks is .96 × .98 × .95 = .894, so the chance something breaks is 1 − .894 = .106, better than one in ten. Stack enough rare risks and rare stops being the right word.

Why chains and stacks fool the gut
Fig. 1 · conjunctive vs disjunctive
EVERYTHING must go right .90 × .85 × .80 (A and B and C) Gut ~85% Real 61% the chain is weaker than any single link ANYTHING can go wrong risks .04 · .02 · .05 (A or B or C) Gut ~5% Real 10.6% 1 − (.96 × .98 × .95), stacked small risks aren’t small 0% 50% 100% PROBABILITY →
Two sides of the same multiplication. Independent probabilities don’t add or average, they multiply. Chain several “likely” events with and and the joint odds sink below your worst input; stack several “unlikely” failures with or and the chance of at least one climbs far past your worst input. The gut anchors on a single comfortable number; the arithmetic refuses to. This is precisely the sum Simulate does for you, thousands of times over, instead of leaving it to intuition.

Put both on one real decision and the gut’s double error shows up in a single frame. Take a market-entry bet, scanned the way any strategy review scans one, a row of factors, each judged on its own. It pays off only if a chain of things all hold; it avoids disaster only if none of a separate set of risks fires. Read item by item, both lists look comfortable. Read as a product, neither is.

Worked example · a market-entry bet, read both ways
To pay off, everything must go right (AND)

Four factors, each judged independently likely: the enabling regulation passes (.90), demand holds through the launch window (.85), the channel partner signs (.80), and the platform integration ships on time (.90). Every one a comfortable yes.

.90 × .85 × .80 × .90 = 0.55, a coin-flip, not the “~85%” the individual numbers whisper
To avoid disaster, anything can break it (OR)

Four low-probability failures, any one of which sinks it: a competitor pre-empts (.05), the key hire falls through (.04), a privacy ruling lands badly (.03), a supplier disrupts (.06). Each one shrug-worthy on its own.

1 − (.95 × .96 × .97 × .94) = 0.17, one in six, not the “~5%” each risk feels like

So the same bet is roughly a coin-flip to win and carries a one-in-six chance of an outright sink, two facts that no single line of the strategy announced, and that stay invisible to anyone reading it factor by factor. (Those eight numbers are a PESTLE scan (political, economic, social, technological, legal, environmental) in disguise: regulation and the privacy ruling are the political and legal rows, demand the economic, integration and supplier the technological. The framework lists the factors; it does nothing to combine them, which is exactly where the gut fails and Simulate doesn’t.)

ivoverconfidence

Overconfidence

Asked for a range, we make it far too narrow.
Overconfidence · the too-tight interval

We think we know more than we do

Ask someone for a launch’s likely market share and they’ll give a best guess readily. Ask for a range they’re 90% sure contains the truth, and the range comes back far too tight, the real outcome lands outside it far more than one time in ten. The error isn’t that we know too little; it’s that we believe our knowledge is more precise than it is. The countermeasure is to force the range, then stress it: what would have to be true for the answer to sit outside it? That is exactly what Elicit makes you do, and then it checks whether your three numbers are even internally consistent.

vmore traps

A few more for the shelf

Motivation · believing in desirable outcomes

Wishful thinking

We assign higher probability to the futures we’d prefer. Once a launch has a date, a budget and a team’s reputation riding on it, the forecast quietly bends toward the number the room needs to be true, and the people producing the estimate are rarely the ones who’ll wear it if it’s wrong. The tell: ask what evidence would change the forecast, and notice how little is on offer.

Evidence · biased assessment of covariation

Counting only the hits

To judge whether two things move together, a tactic and a sales bump, say, you need the full two-by-two table: the times both happened, and the times one happened without the other. We over-attend to the confirming cell and under-count the rest, so we “detect” relationships the complete table wouldn’t support. Illusory correlation, with the receipts missing.

Judgement is the only instrument you have when the data run out. Worth knowing how it’s miscalibrated.

The point is not that judgement is useless, it is all you have when the dataset is empty, and a trained gut is a real asset. The point is that the errors are systematic, which means they can be engineered around rather than merely regretted. Gather estimates privately before anyone anchors the room. Write the base rate next to the seductive story. Force a range, then attack it for being too narrow. And when the decision turns on several uncertain quantities combining, hand the multiplication to a machine instead of asking your intuition to do the one thing this whole list says it can’t.

That last move is the bridge to the tools. Elicit turns three honest numbers into a coherent range, and flags when they don’t hang together; Simulate carries those ranges through the conjunctive-and-disjunctive arithmetic to the odds that actually matter. The next piece in this series, on elicitation methods, is about the other half of the job: how to pull a well-calibrated number out of a person in the first place.

The tools that fight backBoth instruments are built around the biases on this page, ranges forced to be honest about their width, base rates kept in view, and the multiplication done for you.

Where this comes from. The three big families, availability, representativeness, anchoring, are the spine of the heuristics and biases programme begun by Amos Tversky and Daniel Kahneman in the 1970s, with overconfidence and the motivational biases added by the forecasting and decision-analysis literature since. This is a practitioner’s field guide framed for marketing decisions, not a survey of the research.