Anchoring

Anchoring

Anchoring is a cognitive bias in which an estimate of an unknown quantity is pulled toward an initial value — the anchor — even when that anchor is arbitrary, uninformative, or known to be irrelevant. Formalized by Tversky and Kahneman ($1974$), the bias has two components:

1. Initial anchor. The agent starts reasoning from some salient number already on the table — a prior estimate, a price tag, a random number, the question’s own framing.
2. Insufficient adjustment. The agent tries to adjust away from the anchor toward the true value, but stops short — the final estimate is biased toward the anchor relative to a full Bayesian update.

Anchoring is distinct from the conservatism bias in that conservatism describes under-weighting of informative new signals relative to a rational prior, whereas anchoring describes inappropriate weighting of an arbitrary reference value that is not a real prior at all. The two biases produce similar under-reaction in some settings but are driven by different cognitive machinery.

Financial relevance. Anchoring has been invoked to explain:

• Analyst forecast stickiness — earnings forecasts anchor on the previous period’s number and adjust insufficiently toward new information.
• IPO pricing and $52$-week-high effects — investors anchor on salient reference prices and resist revising away from them.
• Target prices in sell-side research — target prices cluster near round numbers and recent trading prices rather than full DCF values.
• Negotiation and valuation — the first number offered in a price negotiation exerts a measurable pull on the final settlement.

Key experimental finding. The sharpest demonstration of anchoring is that randomly generated anchors still distort estimates. Tversky and Kahneman had subjects spin a wheel of fortune to produce a number between $0$ and $100$, then asked them to estimate the percentage of African nations in the UN. Subjects who spun a high number guessed substantially higher than subjects who spun a low number — despite knowing that the wheel was obviously uninformative about the UN. The anchoring effect is robust to warnings, to incentives, and to being told the anchor is arbitrary.

Intuition

The mental operation the brain performs when asked “what is X?” is not a fresh computation from first principles. It is closer to “where is X relative to something” — and whatever “something” happens to be salient in the moment becomes the anchor. In the Mississippi River example from Tversky and Kahneman, a subject asked to estimate the river’s length will silently start from the length of the local river she grew up near, then ask herself whether the Mississippi is longer or shorter, and by how much. She adjusts upward, but not by enough — the adjustment process terminates as soon as she reaches a plausible answer, not when she reaches the true answer. The bias is not in the starting point (any starting point would do for a rational agent, because a Bayesian would adjust all the way to the posterior mean). The bias is in the adjustment, which is systematically truncated.

The central tension is between the computational cost of full reasoning and the convenience of adjusting from a nearby number. Full reasoning requires loading all relevant knowledge into working memory, applying a model, and producing an unanchored estimate. Adjustment-from-anchor requires only an anchor and a direction. The brain picks the cheaper strategy whenever the stakes are not high enough to force the expensive one — and in financial settings, the stakes are often not felt as high enough in the moment, so anchoring persists even among professionals who should know better.

Properties

(Insufficient Adjustment) Adjustment terminates before reaching the Bayesian posterior.

Tversky and Kahneman ($1974$) frame anchoring as a two-stage process: first, the agent selects an anchor; second, she adjusts. Rational updating would adjust until the marginal cost of further adjustment equals its marginal benefit — in a full-Bayesian world, that means adjusting to the posterior mean. Empirically, adjustment stops earlier: as soon as the answer enters a “plausible range” consistent with the new information, the agent terminates the adjustment and reports the nearest edge of that range rather than the center. The final estimate is therefore biased toward the anchor. This is the core operational difference between anchoring and rational updating.

(Arbitrariness Robustness) Anchors work even when known to be arbitrary.

The cleanest evidence for anchoring-as-bias rather than anchoring-as-information-use comes from experiments with manifestly uninformative anchors. In Tversky and Kahneman’s wheel-of-fortune paradigm, subjects who spun a $10$ estimated the share of African UN members at around $25\%$, while subjects who spun a $65$ estimated around $45\%$ — a $20$-point difference driven entirely by a random number. Later replications with social-security-number anchors (Ariely, Loewenstein, Prelec) find similar effects on willingness-to-pay for consumer goods. These results rule out any rational interpretation under which the anchor contains information about the true value.

(Expertise Does Not Eliminate It) Professionals anchor, too.

A frequent objection is that anchoring is a laboratory curiosity that disappears when real money or expertise is on the line. It does not. Real-estate agents shown identical houses with randomized “asking prices” produce appraisal estimates that track the asking price; experienced judges setting criminal sentences anchor on prosecutors’ requests; financial analysts anchor their target prices on previous analysts’ targets. The magnitude of the bias is smaller among professionals than among students but remains statistically and economically meaningful. Expertise trims anchoring but does not remove it.

(Adjacency to Conservatism) Distinct from conservatism in source but similar in effect.

Both anchoring and conservatism produce estimates that move too little in response to new information. The difference is the reference point. Conservatism is an under-update relative to a legitimate Bayesian prior — the agent has a real belief based on data and weights new signals less than Bayes’ rule prescribes. Anchoring is an under-update relative to an arbitrary starting number — the “prior” is not really a belief at all but an accidental salient value. Both can generate under-reaction to public news, and empirical work on analyst forecast stickiness often has trouble distinguishing them. In the Rabin ($1998$) parametric family, conservatism corresponds to $c<1$ (under-weighting new signals); anchoring does not fit cleanly into the Rabin template because it is not a distortion of likelihood ratios but of the adjustment process itself.

(Financial Applications) Target prices, $52$-week highs, and forecast stickiness.

Three empirical regularities in finance have been attributed in part to anchoring:

1. Sell-side analyst target prices cluster near round numbers and near the current price, consistent with analysts anchoring on salient round values and adjusting insufficiently toward their DCF valuations.
2. The $52$-week high effect (George and Hwang, $2004$) shows that stocks trading near their $52$-week high earn positive subsequent returns — interpreted as under-reaction due to investors anchoring on the $52$-week high and resisting further price increases.
3. Analyst earnings forecast stickiness — forecasts revise slowly after earnings surprises and tend to cluster around previous-period numbers, consistent with anchoring on the last published forecast.

Each of these has competing rational or alternative-behavioral explanations, and distinguishing anchoring from conservatism or representativeness in field data is notoriously difficult.

Examples

The Mississippi River

A subject is asked: “Is the Mississippi River longer or shorter than $500$ miles? What is your best estimate of its length?” The subject thinks, “It’s probably longer than $500$, maybe around $1,500$.” She reports $1,500$. The true length is approximately $2,340$ miles. Another subject is asked: “Is the Mississippi River longer or shorter than $5,000$ miles? What is your best estimate?” She thinks, “Probably shorter — maybe $3,500$.” She reports $3,500$. The two subjects see identical information about the world but produce estimates that differ by more than $2,000$ miles, driven entirely by the anchor embedded in the initial yes/no question. Insufficient adjustment from the anchor in each direction produces a wide spread around the true value.

The wheel-of-fortune experiment

Tversky and Kahneman ($1974$) bring subjects into a lab and show them a wheel labeled with numbers $0$ through $100$. Each subject spins the wheel — unbeknownst to them, the wheel is rigged to stop at either $10$ or $65$. After spinning, each subject is asked: “Is the percentage of African nations in the United Nations greater or less than the number on the wheel? What is your best estimate?”

• Subjects who spun $10$ produced a median estimate of $25\%$.
• Subjects who spun $65$ produced a median estimate of $45\%$.

The wheel is transparently uninformative — it is a wheel of fortune, and the subjects know this. Yet the anchor pulls the estimates by $20$ percentage points. The experiment is the foundational evidence that anchoring is a cognitive distortion, not a rational use of information.

Analyst target prices clustering near round numbers

Sell-side equity analysts issue “target prices” — twelve-month-ahead price forecasts for individual stocks. Empirically, the distribution of target prices is not smooth: it has sharp spikes at round values like $$50$, $$75$, $$100$, $$150$. A full DCF valuation would produce target prices with arbitrary cents ($$73.42$, $$129.87$), but observed target prices cluster tightly on round numbers. One explanation is that analysts start from a salient round number — either a recent trading price or a cognitively easy round figure — and adjust away by an amount that is smaller than their model would dictate. The $52$-week high is a particularly salient anchor; target prices within $5\%$ of the $52$-week high are dramatically over-represented in the data.

See Also

• 8.6. Belief-Based Biases — parent hub.
• 8.6.1. Conservatism — closely related: under-updating from a Bayesian prior rather than from an arbitrary anchor.
• 8.6.5. Representativeness Heuristic — a different heuristic for estimation that substitutes similarity for probability.
• 8.6.8. Experience Effects — related: own-experience data serves as a kind of anchor for expectations.
• 8.1. Principles of Behavioral Finance — the framework for classifying anchoring as a psychology-of-beliefs deviation from Bayesian updating.