- The number that was already there
- Why an irrelevant number still moves the answer
- The anchors already sitting in your account
- What happens to the evidence that arrives next
- Why the two together are so durable
- Write the reasoning before you look at the price
- Name in advance what would make you wrong
- What this does not fix
- Where the replacement numbers come from
- Common questions
The number that was already there
An anchor is a number in front of you when you make a judgement that moves the judgement, whether or not it has anything to do with the question. Your purchase price is one. So is the 52-week high, and so is a round figure like ₹1,000.
Two people open the same chart on the same evening. The stock is at ₹840. The first bought at ₹1,200 and has watched it fall all year; the second noticed it at ₹700 three weeks ago and has watched it rise. Neither ₹1,200 nor ₹700 is a fact about the company. Both are facts about when each person happened to start looking.
Ask them whether ₹840 is a reasonable price and you get two answers, delivered with equal confidence, by people who have read the same disclosures. What separates them is not the evidence; it is the direction the price arrived from, and that is in no disclosure.
The classic demonstration is Amos Tversky and Daniel Kahneman's 1974 paper in Science, “Judgment under Uncertainty: Heuristics and Biases”. Participants watched a wheel of fortune stop on a number — they took it to be random; it was rigged to stop on one of two values — and were then asked what share of the member countries of the United Nations were African. The groups anchored at 10 and at 65 gave reported median estimates of 25 and 45. The design's entire purpose was that the number could not possibly carry information about the answer, and it moved the answer regardless.
Why an irrelevant number still moves the answer
Two accounts compete to explain this, and the honest summary is that the effect is easier to demonstrate than it is to account for.
The first, which is Tversky and Kahneman's own reading of their result, is adjustment that stops too early. You treat the anchor as a starting value and move away from it until the number in front of you looks defensible — and defensible arrives before correct, so you come to rest on the anchor's side of the truth. The second is the selective-accessibility account, put forward by Thomas Mussweiler and Fritz Strack: entertaining the anchor at all makes anchor-consistent facts easier to bring to mind, and you then answer from what came to mind. Both have experimental support, for different anchoring designs.
What matters for a portfolio decision is that neither account asks you to believe the anchor. In the wheel experiment nobody thought the wheel knew anything about Africa. The number does not need your assent; it only needs to have been present while you were forming an estimate. That is why “I know my purchase price is irrelevant” is a true sentence that changes very little.
Anchoring is also one of the effects that came through the replication scrutiny of recent years intact: the Many Labs project reported by Klein and colleagues in 2014 reproduced it consistently across many samples. How large it is in any particular setting is a separate question from whether it exists, and the reported sizes vary widely with the design. Treat the mechanism as established and the magnitude as contested — that distinction runs through the rest of this article.
The anchors already sitting in your account
Anchors are not something you go looking for. They are supplied — by the statement, by the chart, by the scheme page — before you have asked a question, which is what makes them hard to notice.
| The number | What it is evidence of | What it gets read as |
|---|---|---|
| Your purchase price | What the market asked on one past day, and what you agreed to pay that day | The level at which the position becomes fair again |
| The 52-week high | The most anyone paid in a year | A level the price ought to return to |
| A round number — ₹1,000, 25,000 on an index | Nothing. Base-10 is a property of our notation, not of the asset | Support, or a psychological barrier |
| The IPO price | What the issuer and its bankers chose to ask | A floor |
| Your portfolio's peak value | The highest mark-to-market it ever printed, for one day | The amount you have lost |
| A trailing 3-year return | One window, whose value depends on where it ended | What the scheme does |
Take the first row seriously, because it is the one that costs money. Whether to keep or sell a holding turns on two things: what the asset is likely to do from here, and what else that money could be doing instead. Your entry price appears in neither. It is a fact about a transaction that has already settled.
The behaviour it produces has a name and a literature. Hersh Shefrin and Meir Statman called it the disposition effect in 1985 — the tendency to realise gains sooner than losses — and Terrance Odean measured it in a large sample of discount-brokerage accounts in a 1998 Journal of Finance paper, reporting that winners were sold at a higher rate than losers. How big the gap is varies by market, period and account type, and no figure we have seen travels across all three. The direction is documented; the magnitude is contested. Why the two directions are weighed unequally in the first place belongs to loss aversion; what concerns us here is only that the entry price is the number the weighing is done against.
There is exactly one place the entry price legitimately re-enters, and it is not the investment case. It is the tax computation: the cost of acquisition and its date fix the gain and the holding period, which is a mechanical matter of capital gains tax and nothing to do with what the asset will do next. Confusing the two is common enough to have a stock phrase — “I'll sell when I'm back to breakeven” is an investment sentence wearing a tax sentence's clothes.
The 52-week high does the same work at market scale. Thomas George and Chuan-Yang Hwang, in a 2004 Journal of Finance paper, found that how close a stock sat to its 52-week high predicted subsequent returns at least as well as standard momentum measures — a price statistic doing work no accounting statement gave it. Malcolm Baker, Xin Pan and Jeffrey Wurgler reported in 2012 that acquisition offer prices cluster around the target's 52-week high. Whether either is anchoring or something else is argued about. That the number gets used is not.
The last row is the one that hides in plain sight. A trailing 3-year figure on a scheme page is a single number whose value depends heavily on where its window happened to end, and it becomes the reference against which every later observation gets judged. Rolling returns exist to remove that accident, and part of why they feel less satisfying is precisely that a distribution gives you nothing to anchor on.
What happens to the evidence that arrives next
An anchor sets the question. What happens to the evidence arriving afterwards decides whether the question is ever revised, and that is a separate mechanism with a cleaner demonstration.
Peter Wason's 1960 experiment is still the clearest one. Participants were shown the sequence 2, 4, 6 and told it followed a rule they had to discover. They could propose triples of their own and were told, each time, whether the triple fitted. The characteristic run went 8, 10, 12, then 20, 22, 24 — triples chosen because they fitted the hypothesis already formed, ascending by two. The actual rule was any ascending sequence, and most participants announced one narrower than that. Finding it required proposing something the hypothesis said would fail.
The instructive part is not that people got it wrong. It is why the strategy felt informative. Every proposal came back “yes”, and a run of yeses is exactly what a correct hypothesis produces. A test that can only return one answer carries no information, and nothing in the experience of running it tells you so.
Joshua Klayman and Young-Won Ha argued in 1987 that positive testing — checking the cases your hypothesis says should work — is a sensible default across most problems, and fails in one specific shape: when your hypothesis is narrower than the truth. There, every confirming case is also consistent with the wider rule you never considered. That is worth restating, because it is the difference between an insult and a mechanism. Confirmation bias is not a taste for being right. It is an efficient search strategy meeting the one problem shape it cannot solve.
The portfolio version writes itself. “This company is executing well” predicts decent quarterly numbers, decent quarterly numbers arrive, and each one is a genuine confirmation. It is also precisely what “the whole sector had two good years” predicts. The evidence does not separate the two hypotheses, and only one of them was ever put to a test.
Why the two together are so durable
Separately, each is one entry on a long list of documented biases, and separately each is manageable. Together they produce something that looks like conviction and is assembled entirely from real material.
The anchor supplies the hypothesis before any work has been done — ₹1,200 says the position is cheap, ₹700 says it has already run. Confirmation then supervises the intake. The research you go and do is real research, the facts you find are true facts, and the file grows. What the file does not do is grow in both directions. Every item in it is accurate; the defect is in the selection.
Call it the one-sided file. It is harder to dislodge than an ordinary error for a reason worth sitting with: there is nothing false in it to correct. Anyone arguing against it has to argue against true statements, which reads as motivated, which is then itself taken as evidence that the file was right.
The same mechanism is why a second opinion sought after a view has formed is a weaker test than it looks. The question carries the anchor. “What do you think of this at ₹840, I'm holding from ₹1,200” has already told the other person where you stand, what you would like to hear, and which number to reason from. It is one of the arguments for settling what you want from outside input before you go and ask for it.
Note what this does not require. Nobody in the sequence has to be stubborn, innumerate or emotional. A careful person following a sensible search strategy, given a number they never asked for, produces the same result. That is the whole reason the fix has to be procedural rather than a resolution to try harder.
Write the reasoning before you look at the price
The order in which you look at things is the part of this you control, and changing it is close to free.
Write the reasoning down before the price is on screen. Not an essay — three or four lines will do: what this holding is meant to do in the portfolio, the handful of facts the case rests on, and the date. Then look at the quote. Once the number is visible it becomes the reference for everything that follows, and a note written afterwards will be a note about the number rather than about the business.
The cost is real and worth stating. This is slower, and on most decisions it changes nothing, because most decisions were not close. Its value concentrates in the few where you were early and wrong, and you do not get to know in advance which ones those are.
The dated note also does a job memory cannot. Baruch Fischhoff and Ruth Beyth reported in 1975 that when people were asked to recall probabilities they had themselves assigned to events before those events resolved, the recalled figures had moved towards what actually happened. A remembered reason is reconstructed, and it is reconstructed to fit the outcome. A written one does not move.
Institutions solve the same problem by putting the rule before the price at the level of the whole portfolio. A rebalancing band — act when an allocation drifts past a stated width — is a decision taken while nothing was happening, specifically so that it is not taken while something is; rebalancing repays being read as behavioural machinery as much as arithmetic. The same lens explains why continuing a SIP through a drawdown is hard: the instalment falls due on exactly the day the anchor is loudest.
Name in advance what would make you wrong
A view that nothing could contradict is not a view about the world. The practical form of that idea — associated with Karl Popper, and older than any of the psychology above — is a question that takes ten seconds to ask: what would I have to see to drop this?
Answering it before the position exists is the whole trick, because a criterion invented afterwards gets written to spare the position. It has to be an observation, dated, and specific enough that it could actually arrive. “If the story changes” is not a criterion. “If the operating margin sits below the level I said it would hold, for two consecutive years” is one.
A price rule is a different instrument, and the two are worth keeping apart. A stop is risk control: it caps what one position can cost you whether or not the reasoning was sound. A falsification criterion is about the reasoning: it names the fact that would retire it. The two come apart in both directions — a position can be stopped out with the thesis intact, and can sit comfortably in profit long after the reason for owning it stopped being true.
For a scheme rather than a stock, the criterion attaches to whatever you claimed when you bought. If the reason was a shallower worst window in its rolling-return distribution than the category, the falsifier is that distribution deteriorating. If it was a cost advantage, the falsifier is the expense ratio moving. Both are checkable on a fixed schedule, which matters more than it sounds: a view that only ever gets reviewed when it is losing is being reviewed by the anchor.
What this does not fix
Three limits, stated plainly, because a section on debiasing that promises more than the evidence supports is doing the thing it warns about.
Knowing about a bias does not remove it. Emily Pronin, Daniel Lin and Lee Ross documented in 2002 what they named the bias blind spot: people rate themselves as less susceptible to standard cognitive biases than others are, including after the bias has been explained to them. The reasonable expectation from reading this is a better chance of noticing the pattern in one specific decision, not immunity.
The debiasing evidence is mixed. Deliberately generating reasons the opposite view might be correct — “consider the opposite”, studied by Charles Lord, Mark Lepper and Elizabeth Preston in 1984 — has support at laboratory scale. Whether an instruction that works inside a twenty-minute experiment transfers to a decision taken over weeks, with your own money, is not something anyone has shown, and we are not going to claim it here.
And the size of these effects outside a laboratory is genuinely disputed. Anchoring on estimates in an experiment replicates well; the distance from there to a claim about what moves a market price runs through a great deal of intervening machinery. The finance papers cited above document what prices and offers did around a reference level — reading that as anchoring in the heads of the people trading is an inference laid over the result, not the result itself, and it is contested. The mechanisms are the reliable part, and the effect sizes are not.
The discipline itself is not free either. Requiring a written falsification criterion for everything you own means fewer decisions taken and slower changes of mind in both directions — including the occasions when quick would have been right. That is the trade being made, and it is worth making with your eyes open rather than by accident.
Where the replacement numbers come from
An anchor holds when the number that would replace it is tedious to compute, which is part of why the trailing 3-year figure survives on scheme pages: the honest alternative is a distribution, and nobody builds a distribution by hand.
FNOTrader's Mutual Funds app runs on the full AMFI NAV history — around 34 million NAV rows — and reports rolling-return distributions for any scheme and period: the worst window, the share of windows below zero, and the spread between them, alongside SIP and lumpsum simulations reporting the internal rate of return across irregularly dated cashflows — XIRR — and maximum drawdown. Those are figures a falsification criterion can be written against, because none of them depends on where a single window happened to end. What each measure does and does not tell you is covered in risk measures.
None of it is required to do the thing this article describes. A dated note in a text file, written before the quote is on screen, is the entire method. The data only decides whether the criterion you wrote can later be checked without an argument about what it meant.
Common questions
What is anchoring bias?
It is the tendency for a number that is present when you form a judgement to pull the judgement towards it, whether or not the number is relevant. Tversky and Kahneman's 1974 demonstration used a rigged wheel of fortune: participants who saw a higher number gave higher estimates of an unrelated quantity. In a portfolio the anchors are your purchase price, the 52-week high, the IPO price and the portfolio's peak value.
What is confirmation bias?
It is the habit of testing a hypothesis by looking for cases that fit it rather than cases that would break it. Wason's 1960 sequence task is the standard demonstration: given 2, 4, 6, most participants proposed triples their own rule predicted would fit, got a run of confirmations, and announced a rule narrower than the real one. The tests were real; they just could not come back 'no'.
Is my purchase price really irrelevant?
To the investment case, yes — the decision to hold or sell turns on what the asset is likely to do from here and what else the money could do instead, and neither of those contains the price you paid. It is not irrelevant to tax: the cost of acquisition and its date fix the gain and the holding period. The error is letting the tax number act as an investment reference, which is what 'I'll sell at breakeven' does.
Why do anchoring and confirmation bias compound?
Because they act at different stages. The anchor supplies a hypothesis before any research has been done, and confirmation bias then governs which subsequent evidence gets weighed. The result is a file of individually true facts, selected in one direction. That is harder to dislodge than an ordinary mistake, since there is nothing false in it to correct.
Does knowing about a bias stop it happening?
The evidence points the other way. Pronin, Lin and Ross documented in 2002 that people rate themselves as less prone to cognitive biases than others, including after the bias is described to them. Reading about anchoring buys a better chance of noticing it in one specific decision, which is why the useful responses are procedural — changing the order in which you look at things — rather than resolutions to be more objective.
Is the 52-week high a useless number?
Not useless, but it is evidence about the past behaviour of prices rather than about the business. George and Hwang's 2004 paper found that nearness to the 52-week high predicted subsequent returns at least as well as standard momentum measures, and Baker, Pan and Wurgler reported in 2012 that acquisition offer prices cluster around it. Whether that is anchoring by market participants or something else is disputed; that the number gets used is not.
How do I write a falsification criterion?
State the observation that would make you drop the view, before you take the position, in terms specific enough that it could arrive and be recognised. 'If the story changes' fails that test. 'If the operating margin sits below the level I said it would hold, for two consecutive years' passes it. Date it and write it down, because a criterion held only in memory gets edited when it triggers.
Is a stop-loss a falsification criterion?
No, and treating it as one confuses two jobs. A stop limits what a single position can cost you regardless of whether the reasoning was sound. A falsification criterion names the fact that would retire the reasoning. They come apart in both directions: a position can be stopped out while the thesis is intact, and can sit in profit long after the reason for owning it stopped being true.
How does this apply to mutual fund schemes rather than stocks?
The anchors are different but the mechanism is identical. The trailing 3-year return is a single window whose value depends on where it ended, and it becomes the reference for every later observation about the scheme. Rolling returns remove that accident by computing every available window instead of one, which is also why they feel less satisfying — a distribution offers nothing to anchor on.
Continue reading
More in Behavioural Finance · App: Mutual Funds · Definitions: glossary · Free tools: calculators · All: every article