Investors sell their winners at about one and a half times the rate they sell their losers. That's from Terrance Odean's study of 10,000 accounts at a large US discount brokerage between 1987 and 1993. Of the paper gains sitting in those accounts, 14.8% got sold. Of the paper losses, 9.8%.
The habit has a name, the disposition effect, and Odean put a number on it. On one worked example the asymmetry costs about 4.4 percentage points over the following year. What follows is how that figure was built, how well it has travelled, and why a good part of the field now argues that the whole thing has been misread.
Measuring a reluctance
Counting sales alone tells you nothing. Somebody holding nine winners and one loser will sell more winners, because that's what there is to sell. Odean's fix was to measure every sale against the opportunity to make it.
On any day an account sold something, he sorted every position in it into four buckets: realised gain, realised loss, paper gain, paper loss. The proportion of gains realised, which he calls PGR, is realised gains over realised plus paper gains. The proportion of losses realised, PLR, is the same sum for losses.
Aggregated over seven years the counts were 13,883 realised gains against 79,658 paper gains, and 11,930 realised losses against 110,348 paper losses. PGR came out at 0.148 and PLR at 0.098. A one-tailed test rejects equality with a t-statistic above 35.
Odean flags the weakness in that statistic himself. It treats every sale and every unsold position as an independent decision, and they clearly aren't. Two investors can be reacting to the same headline. He notes this inflates the test statistics without biasing the proportions, so he ran a second test assuming independence only across accounts. Average account PGR was 0.57, average account PLR 0.36, and the mean gap of 0.21 still returned a t-statistic of 19.
December runs the other way
The annual figure hides a sharp seasonal flip. From January to November, PGR was 0.152 and PLR 0.094. In December the ordering reverses: PGR falls to 0.108 and PLR rises to 0.128.
That's tax-loss selling, and it's the one month in which investors do the arithmetic. Barber and Odean later revisited the question with brokerage data that separated taxable accounts from tax-deferred ones. Both household types preferred to sell winners rather than losers, in both kinds of account, all year. Only in December did the proportion of losses realised exceed the proportion of gains realised, and only in taxable accounts. They describe the active realisation of gains as arguably the biggest tax mistake many investors make.
That later study also undercuts a tidy explanation. If the pattern were driven by tax planning, it should vanish inside a tax-deferred account. It doesn't.
One caveat on Odean's original sample: he couldn't split taxable from tax-exempt accounts at all. At the start of the period, 20% of the brokerage's accounts were IRA or Keogh accounts and they were responsible for 17.5% of trades. Their presence dilutes a tax-driven pattern rather than manufacturing one.
What the asymmetry cost
Two sets of numbers make the cost concrete.
First, the sizes. The gains investors actually realised averaged a 27.7% return since purchase. The gains they held onto averaged 46.6%. The losses they realised averaged minus 22.8%; the losses they held averaged minus 39.3%. Small winners out, big losers in.
Second, what happened next. Over the 252 trading days after the decision, winners that were sold returned 2.35% in excess of the CRSP value-weighted index. Losers that were held returned minus 1.06%. That 3.4-point gap carries a bootstrapped p-value of 0.001. Over 84 trading days the gap was 1.03 points; over 504 days, 3.58 points. Holding the loser because it was due a bounce didn't work at any of the three horizons.
Odean's worked example ties the two together. An investor holds $1,000 of a stock behaving like the average realised winner and $1,000 of one behaving like the average paper loser, and needs to raise cash from one of them. Selling the winner realises a $217 gain. Selling the loser realises a $647 loss. At a 15% marginal rate the tax swing between those two choices is $130, and he gives it up by selling the winner.
Deferring $130 for a year at 8% is worth about $10. Add the $34 of return he forgoes by keeping the loser rather than the winner. Against a $1,000 base, selling the loser leaves him roughly 4.4% better off over the next year.
The assumptions matter, and Odean states them. Marginal capital-gains rates in his sample ran from zero to 28% plus state taxes, and he picked 15%. The investor needs taxable gains to offset. A habit of regular loss harvesting also runs down the stock of available losses. The 4.4% is an illustration assembled from sample averages, not a measurement of what any real portfolio lost.
It replicates almost everywhere
Taiwan gives the largest test anyone has run. Barber, Lee, Liu and Odean took every trade on the Taiwan Stock Exchange for the five years ending 1999, over a billion trades from nearly four million traders. In aggregate, investors were about twice as likely to sell a stock held at a gain as one held at a loss. Eighty-four per cent of all investors in that market sold winners at a faster rate than losers.
The split by investor type is the interesting part. Individuals, corporations and dealers all showed the reluctance. Mutual funds and foreign investors didn't, and together they accounted for under 5% of trades by value. Whatever this is, professional money mostly doesn't do it.
Finland supplies the tightest controls. Grinblatt and Keloharju tracked the buys, sells and holds of every participant in the Finnish market and ran logit regressions with what they describe as a kitchen sink of extra controls. The disposition effect and tax-loss selling came out as the two major determinants of whether an owner sells. They also found the reluctance intensifies once a loss passes 30%. Their data has a hard limit: cost basis is only known for shares bought inside the sample window, so sales without one are dropped.
The rebalancing objection, and what happened to it
Here's the obvious counter. An investor whose winner has swelled into an oversized position sells some of it. That isn't loss aversion, it's housekeeping, and it follows the same logic as annual rebalancing versus threshold bands. Lakonishok and Smidt raised exactly this point in 1986.
Odean tested it twice. Rebalancing usually means trimming rather than exiting, so he recalculated using only sales of an account's entire position in a stock. If rebalancing drove the result, the gap should collapse. It widened instead: PGR 0.233 against PLR 0.155.
Then he came at it from the other side. Someone selling to rebalance normally buys something with the proceeds. So he counted only sales with no new purchase on the day of sale or in the three weeks after. The gap widened again, to PGR 0.449 against PLR 0.281.
He's honest about the imperfection. Shares bought before 1987 aren't visible in the data, so selling every post-1987 share isn't always the whole holding. But the direction is wrong for the rebalancing story. Stripping out the trades most likely to be rebalancing made the asymmetry larger, not smaller.
A trading-cost version of the objection fared no better. Low-priced stocks cost proportionally more to trade and losers cluster at low prices. Odean partitioned the data by price band and by size of return, which holds cost roughly constant within each cell. Winners were realised at the higher rate in 14 of 15 partitions, significantly so in 13.
The strongest counter-argument
The serious challenge isn't about rebalancing at all. It's about whether the disposition effect measures what its name implies.
Itzhak Ben-David and David Hirshleifer studied 77,037 accounts at a large discount broker from 1990 to 1996 and asked a sharper question. If investors genuinely prefer realising gains to realising losses, there should be a jump in selling probability right at the break-even price. A tiny gain and a tiny loss carry near-identical information about a company. Only a preference over the sign of the outcome would separate them.
They couldn't find the jump. Using a residuals method, there's no evidence of one at short prior holding periods of one to twenty days, which is exactly where the psychology should bite hardest. At longer holding periods the estimates run from 4.1% to 5.7% of the unconditional probability of selling: statistically significant, economically slight. Regression discontinuity estimates run from zero to 6.4% and aren't significant at all.
What they found instead was a V. Selling probability bottoms out at zero profit and rises in both directions. After a single day of holding, the chance of selling a position that had moved 5% or less was 1.57%. If it had moved more than 5%, 3.03%, roughly double. Big losses do get sold. Investors just sell big gains harder. The right branch of the V is steeper than the left, and that asymmetry of slopes, they argue, is what PGR minus PLR has been picking up all along.
Their preferred explanation is belief revision and attention rather than a taste for booking profits. A large move is news, and news prompts a speculator to re-examine a position. Something similar shows up on the buy side: investors add to losers more readily than to winners, a reverse pattern that a preference for realising gains can't produce.
None of this makes the sell asymmetry go away. It's an argument about mechanism. The mechanism decides whether this is a habit that can be trained out or an artefact of how attention works.
The frame is wider than one holding
A second line of work complicates things further. An, Engelberg, Henriksson, Wang and Williams found that the disposition effect on a stock depends on how the rest of the portfolio is doing.
In their sample the unconditional chance of selling a given stock was 16%. Conditional on that stock being at a gain it rose to 23%, about 45% more likely. Then split by portfolio performance. When the whole portfolio sat at a loss, investors were 2.77 times more likely to sell a winner than a loser. When the portfolio sat at a gain, 1.29 times. Almost all of the effect lives in the losing-portfolio case.
They test the usual suspects, including extreme returns, rebalancing, simultaneous transactions and investor sophistication, and none accounts for it. These figures come from the May 2019 working paper. The study was later published in the Journal of Finance in 2024 and I haven't compared the two versions line by line.
Delegation bends it even harder. Chang, Solomon and Westerfield ran the test across 128,829 accounts covering 73,558 households from 1991 to 1996, separating individual stocks from mutual funds. For stocks the familiar result held: in months when an investor sold something, they were 3.91 percentage points more likely to sell a holding at a gain, against a base selling rate of 21.7%. For equity funds the coefficient flips to minus 6.56 points. Same people, same months, opposite sign. Their reading is that selling a losing fund is easier because somebody else picked it. These numbers are from the April 2014 working paper of a study published in 2016.
Where the tax arithmetic lands in the UK
Odean's example is American and old. The structure carries over; the parameters don't.
GOV.UK, read in August 2026, sets the capital gains tax-free allowance at £3,000 for individuals in 2026-27 and £1,500 for trusts. On gains from shares and other chargeable assets the guidance lists 18% within the basic-rate income tax band and 24% above it, applying from 6 April 2026.
Two things follow. The cost of realising a gain early is real but moderate at these rates, and it only bites above the annual exemption. It also doesn't arise inside an ISA or a pension, which is where a large share of UK retail equity sits. That last point cuts against reading the disposition effect as tax planning gone backwards. Tax explains December. It doesn't explain the other eleven months, and it explains nothing at all in a wrapper.
Why it matters beyond one account
Andrea Frazzini traced the price consequence. If holders sit on losers, bad news gets absorbed slowly; if holders rush to sell winners, good news does too. Using mutual fund holdings from 1980 to 2002 to build a reference purchase price for each stock, he found post-announcement drift is worst when the news and the existing capital gain share a sign. A strategy trading that spread produced Fama-French three-factor alphas above 200 basis points a month, gross of costs, with the caveat that fund managers are standing in for all shareholders.
What would change the conclusion
Three things would.
The first is mechanism. If Ben-David and Hirshleifer are right that a V-shaped selling schedule rather than a break-even discontinuity is doing the work, then PGR minus PLR is a summary statistic and not evidence of loss aversion. Interventions aimed at loss aversion would then be aimed at the wrong thing. Their evidence is a null result on a jump, and nulls are weaker than they look. The jump estimates they do recover are small enough that the burden has moved.
The second is generality. The strongest evidence is retail brokerage data from the late 1980s and the 1990s, before online execution, fractional shares and near-zero commissions. Taiwan and Finland extend it across borders, not forward in time. I haven't found a study that reruns Odean's test on a modern UK execution-only book, and one may exist that I simply didn't reach.
The third is the cost estimate. The 3.4-point return gap is the fragile piece. It sits at the horizon where Jegadeesh and Titman document momentum, while DeBondt and Thaler find reversals at three to five years. A sample dominated by a reversal regime could show held losers beating sold winners, and most of the measured cost would go with it.
What survives all three is narrow but solid. The asymmetry is real, it's large, and it turns up wherever anyone measures it. Whether it's a preference, a belief or an attention artefact remains unsettled. It's worth setting alongside the broader question of what bad market timing costs the average investor, where the sell decision is one input rather than the whole number. Pre-commitment devices such as a written investment policy statement work on the behaviour without waiting for the mechanism to be settled. So does holding down the compounding cost of fund fees over thirty years, which is a smaller annual number and a far more certain one.