Portfolio-driven disposition effect: the headline did not survive a stricter test (corrected 17 Oct 2026)

The short version

Corrected 17 Oct 2026: the “2.041 times larger” figure in this section did not survive a stricter test. The text is kept as published; see the correction and the robustness update.

The disposition-effect study found that traders on a public CC0 panel sold winning positions more readily than losing ones. A 2024 Journal of Finance paper by An, Engelberg, Henriksson, Wang and Williams argues that this habit depends on the whole portfolio: it is strong when the portfolio is down and weak when it is up. Averaging over both states, as the first study did, hides that.

So I wrote the test down, fixed a pass mark of 2× in advance, and then ran it on the same panel. On the one pre-registered specification, the gap between selling winners and selling losers was 2.041 times larger on trader-days when the trader’s other holdings were mostly at a loss than when they were mostly at a gain. That clears the 2.00 bar, by 0.04, and the 95% interval [1.862, 2.220] contains 2.00, so the margin is inside the noise. The direction is not in doubt: the interaction is negative in every specification run, including the strictest one, where the ratio is 3.013. Two placebo tests are the reason to think it is behaviour rather than an artefact of how the variable is built.

This is a measurement of a published finding on one public dataset. It is not a trading rule and says nothing about what anyone should buy or sell.

What was tested, and why

The objection. The first study measured the disposition effect unconditionally. An et al. find the effect is concentrated when the investor’s portfolio is at a loss. If that is right, an unconditional measurement averages across exactly the state that matters. This was the strongest objection to the first study, the data was already public, and as far as I could find nobody had run this test on this panel.

The hypothesis, fixed in advance. On the Xueqiu panel, the gap between the proportion of gains realised and the proportion of losses realised is larger on trader-days when the trader’s other holdings are mostly at a loss. In regression terms, the coefficient on gain × PortfolioGain is negative.

The pass mark, fixed in advance.

The data and the period

• Source: Jin, Li & Zhu (2021), PLoS ONE 16(2), doi:10.1371/journal.pone.0246759. Replication files doi:10.7910/DVN/NI6SBJ, CC0 1.0. Main panel sample2.Rdata, Dataverse file 4053809.
• What it is: public portfolios on Xueqiu.com, a Chinese social-trading platform, holding Shanghai and Shenzhen A-shares.
• Size: 2,780,879 trader-stock-lot-days, 4,731 traders, 3,501 stocks, 428 trading dates.
• Period: 24 June 2016 to 27 March 2018.
• Integrity check: the file downloaded on 16 October 2026 has sha256 0ebd88d9719cebbc56246cef5c14dcb5513b2d84d4a97d18ba0000dbadcbe8de, the same hash recorded when the study was first run on 10 October.

Each row carries two yes/no flags: was the position closed that day (sale), and was it in profit that day (gain).

Method

Portfolio state is a count, not a return. The panel has no position weights that could be used, so "the rest of the portfolio is down" is measured as: of the trader’s other positions held that day, what share are in profit. PortfolioGain = 1 if that share is above 0.5 and 0 if below.

Leave-one-out. The focal position is excluded from its own portfolio measure. Without that, the conditioning variable would be built partly from the thing being conditioned on.

Exclusions, all fixed in advance.

Swipe the table sideways for more columns →

The analysis sample covers 4,482 of the 4,731 traders.

The test. A linear probability model, sale ~ gain + PortfolioGain + gain × PortfolioGain, with trader fixed effects and standard errors clustered by trader. The ratio reported is the implied gap when the portfolio is down divided by the gap when it is up.

A column called ret exists in the file and was not used for the primary test: it disagrees with the study’s own gain flag on more than a third of rows, and no codebook settles what it measures. That decision was recorded before any outcome was computed.

Results

The pre-registered test

Swipe the table sideways for more columns →

Ratio: 2.041, 95% CI [1.862, 2.220]. N = 2,175,069 position-days, 4,482 trader clusters.

Applied literally, the pre-registered rule returns replicated: negative, significant, and the ratio is above 2.00. It is above 2.00 by 0.04, and the interval contains 2.00.

Every other specification (not pre-registered)

Swipe the table sideways for more columns →

† At sample means. Specifications 3 and 4 interact with continuous controls, so the ratio has no single value: in specification 3 it is 1.478 at two positions, 1.914 at five (the median) and 2.722 at eight. Specification 6 is the placebo below.

Only specification 1 was pre-registered. Specification 7 was added because a third review pointed out it had never been run. It absorbs everything that is constant within a trader on a given day, including portfolio size, holding period and activity, and it gives the largest ratio of any specification. An et al.'s own comparable regression figure is 3.368 (0.293% at a portfolio loss against 0.087% at a gain, their Table 2 Panel A); 3.013 sits about 1.2 standard errors below it.

The placebos

Parity placebo. The same leave-one-out share, built from a column that cannot matter: whether the trader’s other stock codes end in an even digit.

Swipe the table sideways for more columns →

Uncontrolled, the placebo is 21.8% of the headline. With controls it goes to nothing while the real interaction holds.

Foreign-portfolio placebo. The other positions' gain counts are shuffled across trader-days within the same date, portfolio size and own-gain cell. That keeps everything mechanical about the variable and removes only the fact that it is this trader’s book.

Swipe the table sideways for more columns →

Without trader-day effects, about 12–13% of the uncontrolled headline is structural. Under trader-day fixed effects the placebo ratio falls to roughly 1, while the real variable gives 3.013.

A side finding

PortfolioGain has a positive main effect of +0.024398 (t +16.56) in the pre-registered model: a position was more likely to be closed on a day when the rest of the holdings were up, whether or not the position itself was in profit. That is not the disposition effect, it was not what the test was looking for, and under trader-day fixed effects it is absorbed (+0.000341, t +0.13).

What this does not show

1. The conditioning variable is a count, not a value-weighted return. One holding up 1% and one down 40% count as a 50/50 tie and are dropped. An et al. condition on the size of the gain or loss. This is the central difference between the two studies, and it is not resolved.
2. Neither conditioning variable is independent of the focal position. Leave-one-out reduces the dependence but does not remove it. The ret index used in specification 5 is a portfolio index, so it contains the focal position too: regressed on the position’s own gain flag it gives +0.051066.
3. The pre-registration named a statistic it did not pin down. It required "the PGR − PLR gap at least 2× larger" without saying which gap. Pooled, the ratio is 4.420. Averaged per trader, it is 0.857, which is the opposite direction. Both figures are stated in pdde/README.md; the public script does not print them. The regression ratio used above is a third measure, and the choice of it was made after seeing the numbers. The argument for it, that the pre-registered test is that regression, is reasonable but post hoc.
4. The tie exclusion is not neutral. A tie needs an even number of other holdings split in half, so it can only happen on days with an odd number of positions. All 280,771 excluded tie rows are from such days.
5. At two positions the design is degenerate. In a mixed pair the winner is always labelled "portfolio down" and the loser "portfolio up". That stratum is 20.6% of the analysis sample.
6. One placebo seed is not clean. Under trader-day fixed effects, seed 2 of the foreign-portfolio placebo gives t = −2.22. Its ratio, 1.083, is still close to 1 against the real 3.013.
7. Levels do not travel; only ratios do. Here the outcome is the daily chance of closing a held position, so the coefficients are in a different unit from An et al.'s. Their interaction is −0.206%; this one is −0.024977.
8. Chinese A-shares, one social-trading platform, June 2016 to March 2018. The traders chose to run public portfolios on that platform, so the population is selected. Nothing here is a claim about retail investors in general, or about any firm’s clients.
9. It measures a habit. It tests no remedy for it, and it is not a signal for anyone’s trades.

Robustness update, 17 Oct 2026: does the result survive holding age?

In short: no. The “about twice as strong” figure does not hold up when each trader is compared with themselves at the same holding age. Holding age itself turned out not to be the reason; differences between traders were.

Why we checked. A robustness check of the disposition-effect study found that the habit of selling winners is concentrated in positions held a day or less, and that mixing holding ages distorts the figures. This study compares that habit between two kinds of day. On this panel, positions on days when the rest of the book was mostly down were younger (median 25 days held, against 30). So I wrote a new test and a pass mark down, tried them on made-up data first, and sealed them before running them on this panel. The seal hash is recorded in SEAL-ROBUST-PDDE.md and will be published in the study repository’s commit history.

The made-up data came first, and it matters. In synthetic panels where the rest of the book has no influence on selling at all, only the habit of selling winners and the fact that old positions are sold less, the specification this study used found the effect, in the published direction, in 25, 37 and 30 of 40 panels, depending on the setup. In one setup its typical ratio was 1.72, not far from the 2.04 published here.

What came out.

The pre-registered rule says the result fails: the ratio is not above 1 with any confidence, let alone near 2.

What this changes. Holding the holding age fixed barely moves the figure (2.09, 2.01). What removes it is letting each trader have their own winner–loser gap: done that way, in a check I made after seeing the result, the effect is gone. Traders differ in how strongly they sell winners. If the traders with the strongest habit also spend more days with their book down, a single pooled gap makes it look as if being down strengthens the habit. That is a difference between people, not a change within the same person, and the study’s hypothesis is about the second.

Two caveats that limit this result. First, the stricter test leans against finding the effect: on the made-up panels with no effect, it came out below 1 more often than above. So 1.03 does not show the effect is zero; it shows the published size is not supported. (On made-up panels with a real effect it did find it, every time.) Second, it can use only the trader-and-holding-day groups that contain a winner and a loser on both kinds of day, about a quarter of the rows. Using each position’s gain or loss from the day before, instead of the same day, makes the effect weaker in every version, not stronger.

What still stands. The pre-registered verdict, replicated by 0.04, is recorded as it was. The data, the code and every figure above are published. Like the original, this shows a pattern, not why traders sell, and it is not a signal for anyone’s trades.

Reproduce it

Folder: pdde/ in donttrade/disposition-effect.

pip install pyreadr pandas numpy scipy
python reproduce_pdde.py

Run it from inside pdde/. The script downloads the panel from Harvard Dataverse itself; since 16 October 2026 its download step sends an explicit User-Agent (see Versions).

On 16 October 2026 the public script was re-run from a fresh download. Every figure in the tables above matched its output to the printed precision.

References

• An, L., Engelberg, J., Henriksson, M., Wang, B. & Williams, J. (2024). "The Portfolio-Driven Disposition Effect." Journal of Finance 79(5). doi:10.1111/jofi.13378
• Jin, X., Li, R. & Zhu, Y. (2021). "Could social interaction reduce the disposition effect? Evidence from retail investors in a directed social trading network." PLoS ONE 16(2): e0246759. doi:10.1371/journal.pone.0246759. Data doi:10.7910/DVN/NI6SBJ, CC0 1.0.
• Odean, T. (1998). "Are Investors Reluctant to Realize Their Losses?" Journal of Finance 53(5), 1775–1798.

Methods and negative results, yes. Live signals, never.