Recency Bias: Overweighting What Just Happened

Recency bias is weighting what just happened more heavily than its share of the evidence warrants. The mechanism is invisible because the weighting is never stated: somebody picks a window, reads it, and reports the reading as a fact about the thing rather than about the window. Change the window and the fact changes.

Recency bias rests on something that sounds far too small to matter. The stretch of the past a readingWhat one particular window says, as opposed to what the underlying run of numbers is. is taken over is a choice, and almost nobody experiences making it. A change measured over the last quarter and a change measured over the last two years are both honest statements about the same numbers, and on the same afternoon they can point in opposite directions. So a single reading is never a finding on its own, and a reader handed one reading has been handed a conclusion rather than evidence. One invented index below is read eight different ways, and the eight accounts are set side by side.

Is recency bias a mood, or is it a weighting?

Recency bias is not impatience and it is not excitement. Recency bias is arithmetic that nobody wrote down. Any judgement about how something has gone is a weighted average of what is known about it, and recency bias is the case where the newest observations carry far more weight than their share of the evidence would give them. The weights are real, they are doing all the work, and they are never stated out loud. Because they are never stated, nobody argues with them.

The shape is easier to see outside finance, in a kitchen. A cook has worked in the same household for four years, somewhere near fourteen hundred meals. Three weeks ago a dish came out burnt, and last week another one did. Asked how she cooks, anybody in that house will talk about the two burnt dishes. Nobody in the house sat down and decided to put ninety per cent of the weight on the last three weeks. The household answered with whatever sat closest to hand in time, and what came out of their mouths sounded like a description of the cook rather than a description of three weeks. The weighting is invisible precisely because nobody performs it on purpose. Two dishes out of fourteen hundred is 0.1 per cent of the record and, on the evidence of what people actually say, close to all of the answer.

Move it to a market account and nothing about the shape changes. The Palash 100 index, an invented index built for teaching, stands at 127.0 at the end of the eighth quarter against 121.0 at the end of the seventh. Somebody writes that the index is up 5.0 per cent. The sum is right. The sentence leaves out the windowThe stretch of the past a measurement is taken over. One quarter, one year and eight quarters are three different windows over the same numbers.. The window was one quarter out of nine available readings, eight other starting points existed the whole time, and the person choosing between them made a choice. Recency bias in a market account is that missing sentence, not the tone of the one that got written.

Why does nobody notice choosing the window?

The window gets chosen by whatever is convenient, and convenience never feels like a decision. A quarterly statement arrives, so the window is a quarter. An anniversary falls, so the window is a year. A conversation starts with the words since the low, so the window starts at the low. In every one of those cases somebody has fixed the most powerful input in the whole calculation, and in none of them did anybody experience choosing anything. The absence of a decision is what separates recency bias from a lie. A person reporting a window they never noticed picking is not concealing it; they genuinely do not have it in mind.

The four steps below run in the same order every time, and the damage happens at the last one. A window gets picked. The change over it gets computed, correctly. The reading gets reported. And then the window falls out of the sentence. A sentence about how something has gone has no natural slot for it. By the time the reading reaches a reader, the single most important assumption behind it has been deleted from the record.

What do the same nine numbers say when they are read four ways?

Here is the whole of the evidence, and it is nine numbers long. The Palash 100 index stands at 100.0 when it opens. At the eight quarter ends that follow it reads 118.0, 131.0, 112.0, 104.0, 116.0, 124.0, 121.0 and 127.0. The high point of the seriesA run of readings of the same thing, taken at regular points in time. Nine readings here: one at the opening and one at each of eight quarter ends. is Q2 at 131.0 and the low point is Q4 at 104.0. Everything below is computed from that list and nothing is added to it. Read it twice before going on.

Now take the same nine numbers and compute the change to Q8 at 127.0 from four different starting points. From Q7 at 121.0 the index is up 5.0 per cent, and the sum behind it is 6.0 divided by 121.0, or 0.0496. From Q6 at 124.0 it is up 2.4 per cent. From Q4 at 104.0, the low point, it is up 22.1 per cent. And from the opening at 100.0 it is up 27.0 per cent. Every one of those four statements is arithmetically correct, they describe the identical nine numbers, and they are four different accounts of how the same thing has gone. A reader handed only one of them has been given a conclusion dressed as a fact, and the other three were sitting there the whole time.

Can any window over this series report a fall?

Four windows were enough to show that the accounts differ. Eight windows show something sharper. Compute the change to Q8 at 127.0 from every available starting point and the results run like this: from Q7 at 121.0, up 5.0 per cent; from Q6 at 124.0, up 2.4 per cent; from Q5 at 116.0, up 9.5 per cent; from Q4 at 104.0, up 22.1 per cent; from Q3 at 112.0, up 13.4 per cent; from Q2 at 131.0, down 3.1 per cent; from Q1 at 118.0, up 7.6 per cent; and from the opening at 100.0, up 27.0 per cent.

Seven of those eight readings are positive and one is negative, and the negative one starts at the highest reading in the series. The single negative window is the whole hazard in one line. Somebody who bought at the Q2 peak and somebody who bought at the Q4 low are looking at the identical nine numbers and one of them has lost money while the other has made 22.1 per cent. Neither of them is wrong about the arithmetic. Neither of them is describing the series. Both are describing a window, and the window each of them uses was handed to them by an accident of when they happened to arrive.

Line the eight readings up on a single axis and the list turns into a measurement. The lowest is a fall of 3.1 per cent and the highest is a rise of 27.0 per cent, so the eight windows are spread across 30.1 percentage points. The 30.1 point spread is not a property of the Palash 100 index at all; it is a property of the choice somebody made about where to start, and all eight of the readings are correctly computed from the identical nine numbers.

Why is the reading never the finding?

A finding is something that would have come out the same way had somebody else gone looking. A reading is what one chosen window says. The gap between those two ideas is where this whole subject lives, and the reason it stays hidden is that a reading looks exactly like a finding once it is written down as a sentence. The Palash 100 index is up 5.0 per cent over the last quarter. There is nothing in that sentence to object to. The sentence is true, it is complete, and it will survive any check anybody runs on the arithmetic.

The sentence will not survive a second window, and a second window is never applied because nothing in the wording hints that one exists. The error is not in the number; it is in the silent promotion of a reading into a description of the thing. Notice that the sentence never claims to describe the series. The sentence does not deny it either, and a reader with nothing else in front of them will take the reading as the description every single time.

How to Check for Recency Bias in a Market Narrative: what to ask, and in what order?

Scepticism is not a procedure, and a reader who merely distrusts market accounts ends up distrusting the true ones at the same rate as the misleading ones. A fixed order of four questions works instead, put to any account of how a market has gone, in the same order every time. The order matters because each answer decides whether the next question is worth asking at all.

Question one: what window is this reading taken over? Question two: was that window stated by the writer, or did it have to be worked out? Question three: what do the other available windows say, and in particular the longest one? Question four: when was the window chosen, before the data was seen or after? A window picked once the shape of the data was already known was never available to anybody in advance, so an account that cannot answer question four has not produced a rule at all. Run against all four, most accounts survive the first, stumble on the second, and never reach the fourth.

The four questions, run once on the sentence that the index is up 5.0 per cent over the last quarter. The window is one quarter, Q7 to Q8. The window was not stated, so it had to be worked out. The other windows say up 27.0 per cent from the opening and down 3.1 per cent from Q2, and they disagree with each other. The window was chosen after the data was already in view, so nobody could have followed it as a rule in advance. The fourth answer settles the status of the claim.

How to Analyse a Return-Chasing Scenario: what has to be added to the window comparison?

Return chasingMoving money towards whatever has done well over some recent stretch. The stretch chosen is usually what made it look good. is what recency bias looks like once somebody acts on it. Meera Sundaram holds Rs 3,00,000/- in the Vindhya index scheme and reads that something else is up 22.1 per cent. She is not being foolish. She is responding to a number that is arithmetically correct. The analysis has three steps, and the third is the one people leave out.

Step one names the window that produced the attraction. The 22.1 per cent came from a four quarter window starting at Q4 at 104.0, the lowest reading in the whole series. Step two computes the other windows over the same numbers. The eight quarter window says up 27.0 per cent, so the recent four quarters are not the good part; the two quarter window says up 2.4 per cent, so the very recent part is thin; and the window from Q2 at 131.0 says down 3.1 per cent. Step three adds what moving costs. The turnover measurements in the Palash decision log put costs at 4.1 points a year in the busiest quintile against 0.3 points in the quietest, so moving Rs 3,00,000/- at the busiest rate takes out Rs 12,300/- a year against Rs 900/- at the quietest, a difference of Rs 11,400/- a year on a decision whose attraction came from a choice of window.

The third step changes the conversation. The window comparison shows that the attraction is a property of where the reading started. The cost shows what acting on that property removes, in money, every year, whatever the index then does. Devika Rao, the adviser at Palash Advisory Services Private Limited, has both halves in front of her before the meeting starts, and neither half is a view about what any scheme will do. Brad Barber and Terrance Odean reported that heavy trading and lower net returns travel together, in the Journal of Finance in 2000, under the title Trading Is Hazardous to Your Wealth.

Set all five turnover quintiles against the same Rs 3,00,000/- and the cost line runs Rs 900/-, Rs 1,800/-, Rs 4,500/-, Rs 7,500/- and Rs 12,300/- a year, from the quietest quintile to the busiest. The whole of that range is decided by how often somebody acts, and not one rupee of it depends on which window made anything look attractive in the first place.

Recency Bias vs Momentum: what actually separates an error from a documented pattern?

Momentum and recency bias get confused in exactly one direction, so define both sides before putting them next to each other. MomentumA documented pattern in returns, taken separately under market behaviour with its own measurement and its own boundary. is a pattern in returns that has been measured, published and argued over for decades, and it is covered separately under market behaviour with its own literature and its own boundary. A momentum claim fixes a window and a rule first, in writing, and then reports what happened on data the window was not chosen from. Recency bias is not a pattern and is not a claim about returns at all. Recency bias is the omission of a step: a window gets used and never stated.

Now the contrast, and it comes down to one thing. The separating test is when the window was fixed. Momentum has its window declared in advanceFixed and written down before the data was looked at, which is what makes a claim checkable by somebody else afterwards. and reports out of sampleTested on data that was not used to choose the window or the rule in the first place.. Recency bias picks the window after seeing the data and reports the reading as a fact. The consequence is not that one is better than the other. The consequence is that one of them can be shown to be wrong and the other cannot be shown to be anything.

Somebody who says the last twelve months rose and therefore the next twelve will is not doing momentum badly. The claim is not momentum at all. No window was declared before the data, no test on unseen data exists, and there is no version of the claim anybody could check. The pattern itself is set out under market anomalies, with the measurements and the boundary about what may and may not be concluded from them.

Where is weighting the recent more heavily the right thing to do?

Weighting the recent more heavily is often exactly correct, and leaving that out would teach a superstition instead of a check. The condition is simple to state. Where the thing being measured actually changed recently, the older readings are measurements of something that no longer exists. Weight the recent more heavily in that case. Everything else about the judgement follows from that one question.

The street version is immediate. A tea stall outside an office building has taken the same money every week for three years. Last month the building emptied because its main tenant moved out. The three years of takings are now a record of a different stall in a different setting, and the last four weeks are the only readings describing the stall that exists today. Nobody would call it a bias to ignore the three years. The error is never the act of weighting the recent heavily; the error is weighting it heavily without ever asking whether anything changed. Recency bias and correct updating look identical from outside, and the only thing separating them is a question somebody either asked or did not.

Put invented numbers on the stall and the point sharpens. Across the eight weeks before the tenant moved out it took an average of Rs 9,000/- a week, and across the four weeks after it took an average of Rs 3,250/-. The twelve week average of Rs 7,083/- describes no week that has ever happened. The last four weeks describe the only stall that still exists.

How does an adviser put this into a review meeting?

Devika Rao, the adviser at Palash Advisory Services Private Limited, does not run a workshop on recency bias and would get nowhere if she did. She changes one artefact instead. Every review note in the practice now carries four pre-printed lines at the top, and the note cannot be filled in without answering them: the window used for every figure below it, the longest window available computed alongside, when that window was chosen, and what moving would cost in rupees a year. The failure was never that anybody chose a bad window. The failure was that nobody wrote down which window they had used, so putting the window on the note before any number goes on it is the whole intervention.

The same four lines work without a practice behind them. A person deciding alone, with no adviser and no committee, writes the window at the top of whatever they are reading before they let themselves form a view, and computes the longest available window themselves rather than waiting to be shown it. In the Palash decision log, 20 of the 60 investors adopted a written checklist on 4 November, and across quarters five to eight they recorded a written reason on 34 of 41 decisions, being 82.9 per cent, against 19 of 63, being 30.2 per cent, among the other 40. The checklist moved how often a reason got written down, and no more than that. Eight quarters and 60 investors could not support a claim about returns in either direction.

A credit analyst reading a borrower's accounts runs the identical check with different words. Two strong quarters after six weak ones is a window question before it is anything else, and the answer turns entirely on whether something in the business changed or whether two quarters simply came in above the run. Nicholas Barberis, Andrei Shleifer and Robert Vishny set out how a run of recent outcomes gets turned into an expectation of more of the same in A Model of Investor Sentiment in the Journal of Financial Economics in 1998, and Robin Greenwood and Andrei Shleifer showed in Expectations of Returns and Expected Returns in the Review of Financial Studies in 2014 that what people say they expect moves with what has just happened. Both papers set out the mechanism, and neither reports that acting on it pays.

Meera Sundaram, Devika Rao, Palash Advisory Services Private Limited, the Palash decision log, the Palash 100 index, the Vindhya index scheme, the Nilgiri mid-cap scheme, Suvarna Chemicals Limited and Kesari Logistics Limited are invented.
Educational material. Not advice on any investment, tax, budget or market position.