Recency bias, and the table that always ends today

What recency bias actually claims

Recency bias is the tendency to give recent observations more weight than older ones of equal relevance. The reason is not that the recent evidence is better. It is that the recent evidence is easier to bring to mind, and ease of retrieval gets read as strength of evidence.

That substitution has a name and a paper behind it. Amos Tversky and Daniel Kahneman described the availability heuristic in Availability: A Heuristic for Judging Frequency and Probability (Cognitive Psychology, 1973): when asked how common or how likely something is, people answer by checking how readily instances come to mind. Most of the time that works, because common things genuinely are easier to recall than rare ones. The shortcut fails whenever something other than frequency is driving the ease.

Recency is exactly such a something. The advantage recent material holds in recall is old news in psychology and predates any of this being applied to money — Bennet Murdock's The serial position effect of free recall (Journal of Experimental Psychology, 1962) is the standard demonstration that items encountered last are recalled at a higher rate than those in the middle. That demonstration used word lists in a laboratory, not years of market history — what carries across is the shape of the effect, not its size. So the last year comes to mind first partly because of where it sits in the sequence, which is a property of memory rather than a property of the evidence.

Two tiers of claim are worth separating before going further. That recall is easier for recent material is about as settled as psychology gets. Any specific laboratory coefficient from the heuristics literature is a different matter, and the field's replication record since the Open Science Collaboration's 2015 reproducibility study in Science counsels treating single estimates carefully. Nothing below rests on a magnitude; the whole argument runs on direction plus arithmetic.

And one thing recency bias does not claim, because the popular version usually gets this backwards. It is not the view that recent information is less useful than old information. A change of fund manager, a change of mandate, a regulatory change to a category — these are recent and decisive, and their relevance comes from what changed, not when. The bias is the substitution of one for the other: letting the date of an observation stand in for its bearing on the question.

A short record moves a long way on one observation

Before any psychology enters, there is an arithmetic reason the recent period dominates an estimate: there is not much else in the estimate.

Take five annual observations, four of them 8% and one of them 60% — figures chosen to make the arithmetic visible, not drawn from any scheme. Their simple average is 92 ÷ 5, or 18.4%. One year out of five has moved that average 10.4 percentage points away from what the other four did. Nothing has gone wrong; that is simply what a mean over far fewer independent observations does when one of them is unusual.

The awkward part is that the confidence attached to an average does not shrink in step with the number of observations behind it. Tversky and Kahneman made this the subject of Belief in the Law of Small Numbers (Psychological Bulletin, 1971): people expect a small sample to resemble the population it came from far more closely than sampling variation allows. A five-year record is a small sample of the thing it is being used to describe, and it does not present itself as one.

Which is where the two halves meet. The short window carries genuinely less information, and the recent window is the one that comes to mind most readily. The estimate ends up resting hardest on the evidence that supports it least.

Why one good year appears in every column

Here is the mechanical consequence that does the damage, and it is visible on any scheme page in the country.

A trailing return figure is measured backwards from today. The 1-year, 3-year and 5-year numbers therefore share an end point and differ only in where they start, which means a single strong episode near the end of the record sits inside all three at once. Take a scheme flat for four years and then up 60% in the fifth — round inputs chosen so the arithmetic can be checked, not a description of anything real. The total factor is 1.6, so the 3-year compound annual figure is 1.6^1/3, about 17% a year, and the 5-year figure is 1.6^1/5, about 9.9% a year.

Three rows, three horizons, three green numbers — and one event, three columns. A reader scanning that table experiences agreement across short, medium and long horizons, which is normally a good reason to take a signal seriously. The agreement is manufactured by the shared end date. The three figures are not three tests of the scheme; they are one episode, divided by three different lengths of time.

This is precisely the defect rolling returns exist to remove. Instead of one 3-year period ending today, they compute every 3-year period available — starting in each month of the record — and show the distribution. The end date stops being privileged, and the four flat years get their own windows instead of being averaged into somebody else's.

The related habit of treating your own purchase price as the reference level is a different bias with the same shape, and it has its own article. Both are a number that arrived by accident being used as though it were chosen.

The record is longest at the moment the run is finished

Now put the table and the memory together, because the combination produces a specific and expensive timing pattern.

A trailing table ends today. So the moment at which a strong stretch is fully inside the 1-year, 3-year and 5-year windows, with nothing newer diluting it, is the moment the stretch has just finished. The most persuasive version of the record exists only after the return has been earned. That is true by construction, not coincidence — it follows from how the windows are drawn, and it would be true of any series measured this way.

What happens next has been measured, and the two standard references point the same way. Erik Sirri and Peter Tufano's Costly Search and Mutual Fund Flows (Journal of Finance, 1998) found that flows respond to past performance and do so asymmetrically, with strong past performers attracting disproportionate inflows. Robin Greenwood and Andrei Shleifer's Expectations of Returns and Expected Returns (Review of Financial Studies, 2014) went upstream of the flows to what investors say they expect, and found survey measures of expected returns moving with returns already earned.

No magnitude is quoted from either, and that is deliberate: how large these effects are depends on the dataset and the specification, and it is contested. The direction is all the argument needs. Both papers also use datasets from outside India, and we are not citing an Indian equivalent, because we do not have one to cite.

There is also a theoretical account of why a run gets extrapolated: Nicholas Barberis, Andrei Shleifer and Robert Vishny's A Model of Investor Sentiment (Journal of Financial Economics, 1998) builds a model in which a sequence of similar outcomes leads to the belief that a trend is in place. Read that as what it is. It is a model of a mechanism, offered as an explanation, not a measurement of how any real investor decided anything.

Now the sentence this section exists to not write. None of the above says a strong stretch is about to end, or that a category with a long record is due a fall. Whether the next period resembles the last is not a question this article can answer, and an article that answers it is selling you something. The claim is narrower and entirely about timing of evidence: the record reads at its most convincing after the return is in the past, so money committed on the strength of the record is being committed to the period that follows it, not to the period it describes.

The clearest place to watch this operate is a fund built around a theme that already has a record — the launch is possible because the record exists, and the record exists because the run happened. Sectoral and thematic schemes are treated separately for that reason. The crowd effect that runs alongside it, where other people's buying becomes the evidence, is a different mechanism that happens to fire at the same moment.

The same table, run backwards

The exit side is the identical mechanism with the sign flipped, and it gets written about far less because leaving quietly looks like prudence.

A weak stretch is fully inside the trailing windows at the moment it has just finished, so a holding's record looks worst exactly when the recent period is the whole of what the table can see. The sentence that follows is usually some version of “this scheme is not performing” — and the honest question is whether anything about the scheme changed, or only which twelve months are inside the window being quoted.

Which produces a fact about trailing numbers that is worth carrying around, because it sounds like a trick and is not. A holding's 3-year figure can improve substantially with no new information whatsoever, purely because a bad quarter has aged out of the window on a date fixed by the calendar. Nothing was learned. The rear boundary of the window moved past an episode, and the number rose. That is an accident of the calendar presented in the same typeface as a result.

The mistake underneath all of this is specific enough to recognise in yourself: reviewing a holding on the window that happens to end today, and reading the window's contents as the holding's character. The review is not wrong to happen. It is wrong to be timed by the calendar and scored by the calendar at once, since both the decision to look and the number you find are being set by the same arbitrary date.

What this does not license, again, is the mirror conclusion. “It has fallen, so it is cheap” is not an inference this article supports either, and the reluctance to realise a decline supplies plenty of reasons to hold a poor holding that have nothing to do with the holding.

Expecting a reversal is the same error with the sign flipped

There is a predictable overcorrection to all of this, and it deserves naming before anyone arrives at it independently.

If a long run of good outcomes is inflating the estimate, the tempting conclusion is that a long run must be nearly over. That is the gambler's fallacy: treating independent outcomes as though they owed a correction, the way a coin that has landed heads five times is imagined to be somehow due tails. Both errors read the same run and reach opposite conclusions, and both are reading the run as though its length were information about the next period.

So recency bias is a claim about how an estimate gets built, not a signal about what happens next. Correcting it does not hand you a direction. It hands you a wider set of evidence and a more honest sense of how little a five-year record can settle — which is less satisfying and considerably more useful.

Two questions do the practical work, and neither requires a forecast. First: what would the older evidence have to look like for this conclusion to change? If you cannot answer that, the older evidence was never in the estimate to begin with. Second: which twelve-month period is doing most of the work in the number being quoted? For a trailing figure taken after a sharp move, the answer is one period, and it is the most recent one.

The wider point is that biases operating on which evidence reaches you cannot be fixed by resolving to weigh evidence better, because the weighing happens after the selection. That argument belongs to the pillar on behavioural biases, along with what actually helps, which is deciding the rule at a moment when no window is ending.

What the thought sounds like, and which window it is using

None of this is diagnosable from outside, and it is not meant to be. What can be inspected is the window a sentence is quietly using, which is usually audible in the sentence itself.

Every sentence on the left is a normal thing to think, and none of them is careless. They are all doing the same structural thing: letting the period that is easiest to recall define what normal looks like, and then measuring everything else against it.

The single tell that costs nothing to check is whether the number being quoted has an end date attached to it in your own head. “Up 17% a year” is not a fact about a scheme until you can say over which three years, and whether moving those three years back by six months would leave the sentence standing. If it would not, the sentence was about the calendar.

Looking at a record that does not end today

Everything above reduces to one practical question: whether the number in front of you was computed from a single end date or from all of them.

FNOTrader's Mutual Funds app runs schemes against the full AMFI NAV history — around 34 million NAV rows — and reports invested amount against value, the annualised return on each, the worst peak-to-trough fall along the path, and the rolling-return distribution across every available start date rather than the one that happens to end this week.

That last output is the one this article points at. A distribution across start dates answers a different question from a trailing figure: not what this holding period returned, but what the range of outcomes looked like depending on when you began — and the spread between the best and worst windows is a direct measure of how much of any single number was the choice of date. The gap between what a scheme returned and what its investors earned is measured separately and treated in its own article.

Past performance is a record of what happened, not an indication of what will. FNOTrader is not a SEBI-registered investment adviser and does not give investment advice.

Common questions

Continue reading

More in Behavioural Finance · App: Mutual Funds · Definitions: glossary · Free tools: calculators · All: every article