Recency bias

What recency bias is

Recency bias is a weighting error about time. The evidence is not missing and it has not been misread: an older record exists, is available and is remembered, and it is quietly discounted because something newer arrived. The result is a judgment that tracks the latest data point more closely than the underlying reality it is supposed to measure.

What separates it from its neighbors in the cognitive bias family is the axis it runs along. Confirmation bias sorts evidence by whether it agrees with you. The availability heuristic sorts it by how easily it comes to mind. Recency bias sorts it by arrival time, and it needs no belief, no motive and no emotional stake to operate. It will distort your view of something you are completely indifferent about.

The one sentence version, and what it does to a record

In one sentence, recency bias means the newest observation is treated as the most representative one, when in most cases it is simply one observation among many and carries no special claim to be typical.

The practical effect is easiest to see with numbers. Consider a supplier whose monthly on time delivery rate over 24 months has been steady at about 96 percent, with two bad months at 78 percent, both of which happened to fall in the last quarter. A review meeting held today will describe an unreliable supplier. A review meeting held four months from now, with the same 24 months of data plus four ordinary ones, will describe a reliable one. Nothing about the supplier changed between those two meetings except which observations were nearest the door.

• Small samples get treated as trends, because two recent points feel like a direction.
• Long records get treated as background, because they are not new and nobody circulated them.
• Reversion gets read as change, because an unusual run followed by a normal one looks like a decline.

The evidence: recency in memory and recency in judgment

The firmest experimental evidence sits in memory research, where the pattern is called the serial position effect. Ask people to memorize a list and recall it in any order, and their accuracy forms a U shape: the first few items are recalled well, the last few are recalled best, and the middle is recalled worst. Bennet Murdock reported the curve in detail in 1962, and the two halves have names, primacy for the opening advantage and recency for the closing one.

The most informative follow up is that the recency portion is fragile. Murray Glanzer and Anita Cunitz showed in 1966 that inserting a short filled delay between the list and the recall, occupied by an unrelated task such as counting, wipes out the recency advantage while leaving the primacy advantage largely intact. Two different mechanisms are producing the two ends of the curve, and only one of them survives a pause.

Recency in judgment is a separate literature and a messier one. Robin Hogarth and Hillel Einhorn set out a belief adjustment model in 1992 that predicted when a sequence of evidence produces a recency effect and when it produces a primacy effect, with the answer depending on how the information arrives: short, simple, item by item sequences tend to produce recency, while long sequences evaluated as a whole tend to produce primacy. That is the honest state of it. Order matters, and which end wins depends on the format, so anyone claiming that the last speaker always wins is overstating a real finding.

Where recency bias shows up in decisions people actually make

Recency bias appears wherever a judgment about a long record gets made at a single moment in time, such as annual performance reviews, hiring interviews, supplier assessments, sports selection, quality ratings and investment decisions.

• The performance review. A manager writing a review in December remembers November clearly and March not at all. The employee who delivered steadily all year and had a difficult autumn receives a worse review than the one who was quiet until October, and neither review measures the year.
• The interview panel. Eight candidates are seen over two days. The panel discusses them after the eighth, and the eighth is described in the most detail. Panels that score each candidate immediately after their own interview produce different rankings from panels that wait, which is the cheapest fix available in hiring.
• The forecast. A team that has beaten its target for two quarters forecasts a third. The two quarters are real and the extrapolation is a guess, and the sample supporting it is smaller than anyone says out loud.

What recency bias is confused with, starting with a genuine trend

Recency bias is confused first with a genuine trend, and after that with its own mirror image. Recency bias favors what came last; primacy bias favors what came first. They are the two ends of the same serial position curve, and which one dominates depends on the format of the sequence rather than on the person.

The last row is worth pausing on, because the gambler's fallacy and recency bias look like contradictory claims about human beings and both are well attested. Faced with a run of the same outcome, people sometimes expect it to continue and sometimes expect it to reverse. Which way they go depends on whether they believe the process has a memory: runs produced by a person or a team are read as evidence of form, and runs produced by a mechanism are read as debts owed. Neither reading examines the whole record, and neither reader tends to notice the omission, which is the self-assessment problem described on the Dunning Kruger effect page.

The confusion that costs the most is with a genuine trend, because a real change also announces itself in the most recent data and looks identical at first sight. Calling every reaction to new information recency bias is as bad an error as ignoring the record: an organization that always weights all 24 months equally will be the last to notice that something broke in month 23.

One question separates the two: is the recent swing larger than the normal variation already visible in this record? A supplier whose monthly rate has always moved between 94 and 98 percent and is now at 96 has not changed, however recent the 96 is. The same supplier at 78 for three consecutive months has moved outside its own history, and that is a trend rather than a bias in the reader.

What actually reduces recency bias

What reduces recency bias is a written record kept at the time, in a fixed format, because the bias operates on what you can retrieve and a record does not require retrieval. Six moves:

1. Score each item at the time it happens, before the next one arrives.
2. Plot the full record before discussing the latest point.
3. State the number of observations behind every judgment, including the recent ones.
4. Compare the recent period against the long run average, not against the previous period.
5. Ask whether this same data point would have moved you if it had arrived first.
6. Randomize the order in which candidates, options or reports are reviewed.

Be honest about the limit. Recent data is sometimes genuinely more informative: when a real change has occurred, the newest observations are the only ones that reflect it, and a rule that always weights all periods equally will be slow to notice a real shift. The skill is not to ignore recent information but to ask whether you have enough of it to distinguish a change from a fluctuation, which usually means asking how large the swing is compared with the normal variation in the record. Extrapolating the last three points because that feels like the simplest reading is not a use of Occam's razor. The razor is about choosing between explanations of the same evidence, and it says nothing about assuming that a recent pattern will continue.

The test to run on your next judgment

Ask two questions before you commit to an assessment of anything with a history:

• How does this look over the entire record, and when did I last actually look at it?
• If today's observation had arrived first instead of last, would I be giving it this much weight?

A useful third, for meetings: what is the oldest piece of evidence anyone in this room has mentioned? If the answer is six weeks and the decision concerns five years, the room is not judging the record, it is judging the last email about it.

Recency bias in psychology: serial position effects, not a bias of its own

In psychology, the term of art is the recency effect, and it belongs to memory research rather than to the catalog of cognitive biases. It is one half of the serial position effect, the U shaped recall curve described earlier on this page, and it has been studied since long before "recency bias" entered business and investing writing as a phrase. A reader moving between the two vocabularies should expect the research to be about word lists and recall curves, not about quarterly reviews.

Two traditions therefore have to be kept apart:

• Memory research owns the recency effect proper, where the finding is about the last items in a studied list and is among the most reliably reproduced results in experimental psychology.
• Judgment and decision making owns order effects in belief updating, where recency is one possible outcome and primacy is the other, and where the format of the evidence decides which appears.

What the popular version overstates is how automatic it is. Business writing tends to present recency bias as a fixed law that the most recent data always wins, which the research does not support: Hogarth and Einhorn's 1992 model was built to explain why some sequences produce recency and others produce primacy, and a short filled delay is enough to remove the memory version entirely. The honest summary is that recency is a strong tendency in particular conditions, chiefly when judgments are made item by item and immediately, and that those are exactly the conditions a review meeting creates.

How long recency bias lasts, and what it does in investing

There is no fixed duration, and the two literatures give different answers because they are describing different things. The recency effect in memory is measured in seconds: Glanzer and Cunitz showed that a filled delay of half a minute is enough to remove it. Recency bias in judgment has no such clock. It persists for exactly as long as the newest information remains the most accessible thing in the decision, which can be a quarter, a review cycle or years if nobody opens the full record. That is why the countermeasures above are all about records rather than about waiting.

In investing, recency bias is the standard explanation for performance chasing: money flows toward funds and asset classes after a strong run and away after a weak one, so that the buying decision is made using the smallest and most recent slice of the available history. This site describes the reasoning pattern and offers no financial guidance of any kind. The general point is the one this whole page makes: a track record is data, the last year of it is a sample of that data, and a judgment built on the sample while the record sits unopened is not a judgment about the track record at all.

Where to go next

BIASESCognitive biasWhat it is, with examplesBIASESThe Dunning Kruger effectThe Dunning Kruger effect, and what the study actually showedBIASESConfirmation biasDefinition and examplesEVIDENCEOccam's razorWhat it says and what it does not