Key takeaways
- India's market regulator found 87.81% of first time derivatives traders and 87.69% of repeat traders lost money in the year to March 2026, so experience barely moved the odds.
- Repeat Indian derivatives traders lost an average of ₹135,716 each that year, against ₹58,620 for first timers, so experience enlarged the loss rather than shrinking it.
- In the Finnish trading register, 100 extra trades improved 30 day returns by 3.6 basis points once investor attrition was taken out, roughly 30 basis points a year.
- Ignoring investor attrition makes those learning estimates two to four times larger, and adding investor fixed effects alone cuts them by about 50%.
- Taiwanese day traders with 50 or more days of losses behind them had a 95.3% chance of trading again within a year, against 96.4% for those in profit.
Investors do learn from their mistakes, but mostly by stopping
Do investors learn from their mistakes? The records say yes, a little, and not in the way the question implies. The strongest finding in this literature isn't that experienced investors trade better. It's that the ones who traded worst left, which lifts the average of whoever is still there without anybody getting better at anything.
Amit Seru, Tyler Shumway and Noah Stoffman measured the split in 2010, using the complete Finnish shareholding register from January 1995 to December 2003. It covers nearly 1.3 million individuals and firms and more than 22 million trades by households. Correct for investors dropping out, and an extra 100 trades bought an improvement of 3.6 basis points over a 30 day holding period, or about 30 basis points a year. Skip that correction and the same estimates run "about two to four times higher". Measure experience in years rather than trades and the improvement is, in their word, "negligible".
So the honest answer has two halves. Individual investors do get slightly better by trading, and the effect is real rather than zero. But most of what looks like investor learning in the aggregate numbers is survivorship.
A regulator ran the test on first timers and repeat traders and found no gap
The cleanest modern version of this test comes from a regulator rather than a journal. In August 2026 India's Securities and Exchange Board published its study of individual profit and loss in the equity derivatives market, built on client level data from the top 15 brokers. It splits traders into first timers and "regular traders", meaning anyone who had traded the segment in an earlier year and came back. That split is the question itself: does prior trading experience change the outcome?
| Year to March 2026 | First time traders | Repeat traders |
|---|---|---|
| Share making a net loss | 87.81% | 87.69% |
| Average net loss each | ₹58,620 | ₹135,716 |
| Share of all individual traders | 25% | 75% |
| Share of total net losses | 12% | 88% |
The regulator's own reading is flat: "So prior trading experience did not appear to materially affect the likelihood of incurring losses." The experienced group lost more than twice as much per head, and they traded far harder. Their average annual turnover was about 45.7 times their portfolio value.
The chart above plots the loss making share across the five years to March 2026: 90.2%, 91.5%, 91.1%, 90.9% and then 87.7%. Four years at roughly nine in ten, then a fall of about three points. The regulator attaches its own caveat to that last bar, and it is the same caveat the academic literature spent a decade learning to apply: "The year also saw a 20% drop in the number of active traders, so a lower loss-maker share does not necessarily mean trading outcomes have genuinely improved."
Two different things get called learning, and only one of them is improvement
The Finnish paper is valuable because it separates them. Learning by doing is an investor actually getting better: better selection, and a smaller disposition effect, which is the habit of banking gains while sitting on losses. The median investor in that register was 2.8 times more likely to sell a stock that had risen than one that had fallen.
Learning about ability is a different thing. The investor discovers their ability is poor and stops. Both routes raise the average performance of the surviving pool, and only the first is a person improving. The authors put the balance plainly: "the primary way that low-ability investors learn is by learning to stop trading."
Their selection model shows the exit is performance driven. An investor whose performance sits one standard deviation below the mean is about 15% less likely to keep trading. That matters for how you read every study in this area. A sample of experienced investors is a sample of people who didn't quit, and quitting correlates with being bad at it. The paper's verdict on the earlier work is direct: "By ignoring investor attrition, the existing literature significantly overestimates how quickly investors become better at trading."
The learning estimate shrinks every time it is corrected
Here is the same Finnish data under four specifications. Each measures the market adjusted return on a stock in the 30 days after an individual bought it.
| Specification | Improvement in 30 day return |
|---|---|
| Raw regression, one extra year of trading | 37 bp |
| Add investor fixed effects, one extra year | 22 bp |
| Add the attrition correction, per 100 extra trades | 3.6 bp |
| Add the attrition correction, measured in years | negligible |
The rows aren't strictly comparable, because the last two count experience in trades rather than years. The paper's own cross specification summary is the number to hold on to: unadjusted, the estimates run "about two to four times higher". Adding investor fixed effects alone cuts them by about 50%.
And the raw estimate isn't merely high, it's absurd on its face. Taken at face value it implies that "an investor with eight years of experience should outperform a new investor by about 22% per year". Nobody believes that. It's the shape a learning curve takes when the people who failed have been quietly removed from it. For context on the base rate, a stock bought by a Finnish individual investor returned 48 basis points less than the market over the following 30 days.
Where the losses arrive daily, the learning is still weak
If slow feedback were the binding constraint, day traders would be the fastest learners alive. They get a profit and loss number every afternoon. Brad Barber, Yi-Tsung Lee, Yu-Jane Liu, Terrance Odean and Ke Zhang tested that against the complete transaction record of the Taiwan Stock Exchange from 1992 to 2006. Aggregate day trader performance, net of fees, was negative in each of the 15 years. The share of day traders who were profitable sat "consistently about 5%".
Some learning does show up. More than 75% of all day traders quit within two years, and poor performers quit sooner. But the sensitivity to losing money is small. Here is how often traders stopped, split by whether their past intraday trades had made money.
| Day trading experience | Stopped a month or more, in profit | Stopped a month or more, in loss | Stopped a year or more, in profit | Stopped a year or more, in loss |
|---|---|---|---|---|
| 10 to 49 days | 41.98% | 43.19% | 5.43% | 6.55% |
| 400 or more days | 7.58% | 9.30% | 1.58% | 2.05% |
The authors' summary is the best sentence in the paper: "Traders are sensitive to losses, but not very sensitive." Pooling the experienced groups, previously unprofitable traders with 50 or more days behind them had a 95.3% chance of day trading again within a year. For those who had been profitable the figure was 96.4%. Across the sample as a whole, previously unprofitable traders accounted for 72% of day trading volume. Persistence on that scale fits a model of biased updating, where wins are credited to skill and losses to luck, which is also the engine behind overconfidence in investing.
Taiwan in the 1990s wasn't an unusual market. The Indian regulator's literature review places its own loss rate alongside an earlier UK Financial Conduct Authority analysis, in which 82% of contract for difference clients lost money, and the Australian regulator's figure of 68% for the 2024 financial year.
People grade the outcome, not the decision
There's a mechanism underneath all this, and it explains why repetition fixes so little. James Choi, David Laibson, Brigitte Madrian and Andrew Metrick examined 401(k) records across five large US employers. They found that savers update on their own realised return rather than on anything that predicts future returns. A one standard deviation better 401(k) return in a year raised the saver's contribution rate by 0.13 percentage points of income. A one standard deviation rise in the variance of that return cut it by 0.16 points at once and 0.34 points the following year. The average annual change in contribution rate across the sample was 0.30 points, so these aren't rounding errors.
The trouble is what that signal contains. The authors tested whether a good 401(k) year predicts the next one, and it doesn't. Across the sample the equally weighted participant portfolio produced a three factor alpha of minus 1 basis point a month and a four factor alpha of plus 11, neither distinguishable from zero. Savers were learning diligently from a variable carrying no information. A bumpier return also cut the savings rate by more than a better return raised it, and by more again a year later. Whether that is loss aversion or something else, the trigger is identical: the saver’s own recent experience.
Experience does blunt it, slowly. "Each additional decade of age reduces return chasing by 19% and variance avoidance by 30% relative to the tendencies found in a 21 year old." On the point estimates, return chasing continues until age 74. Income, the usual proxy for financial sophistication, had no significant attenuating effect at all. The Indian derivatives data shows a gradient of the same gentle kind. In the year to March 2026, 89% of traders under 30 lost money, against 81% of those over 60.
The strongest case that investors do learn from their mistakes
The objection to all of the above is serious and it carries numbers. Lei Feng and Mark Seasholes, writing in the Review of Finance in 2005, found that sophistication and trading experience together "eliminate the reluctance to realize losses". The asymmetry is the interesting part. The same forces "reduce the propensity to realize gains by 37%" without removing it. That is a large, measured improvement in a documented bias, attributed to experience.
Gina Nicolosi, Liang Peng and Ning Zhu found something adjacent in US discount brokerage records, 65,118 household months across 2,973 households from March 1991 to November 1996. Investors bought more actively when their own history implied they could pick stocks, and the authors report that "trading experience helps improve portfolio performance". That's individual level evidence of people reading their own record and acting on it.
Two critiques sit on top of that case. The first is the attrition point again. Feng and Seasholes estimated that for sophisticated investors the disposition effect "is essentially attenuated after about 16 trades", and the Finnish authors note that because Feng and Seasholes "do not adjust for heterogeneity and attrition, it is impossible to tell whether their estimates imply that a particular investor who trades 16 times will no longer exhibit the disposition effect or whether they imply that most investors who exhibit the disposition effect will cease trading before they place 16 trades". The second critique is scale. Cutting one bias by 37% among survivors is entirely compatible with a market where nine in ten derivatives traders still lose money.
The aggregate gap hasn't closed, and it isn't proof either way
The figure people usually reach for here is the behavior gap, the difference between what a fund returned and what the average pound or dollar inside it returned. Morningstar's Mind the Gap 2025 study, covering more than 25,000 US funds and exchange traded funds, put the average dollar's return at 7.0% a year for the decade to 31 December 2024 against the funds' own 8.2%. The gap of 1.2 percentage points is, Morningstar writes, "generally consistent in magnitude with gaps we estimated over the 10-year periods ended Dec. 31 of 2020, 2021, 2022, and 2023".
Five overlapping decades, five similar gaps. Read as a learning test that's a null result, and it is the result most often quoted as one. Morningstar is careful about what the figure is, though, and anybody citing it should be too. The methodology states that the estimate "is not a proxy for the average investor's dollar-weighted return". Regular contributions and rebalancing open a gap on their own, which is why the study says "it's not advisable to view this study's findings as a parable of "dumb money" or evidence of individual investors' fallibility". Where the gap does widen is with transacting. Sorted by the volatility of a fund's cash flows, it runs from 0.8 points a year in the calmest quintile to 1.8 in the busiest.
What this evidence cannot tell you
Start with where the data comes from. Finland to 2003, Taiwan to 2006, US brokerage accounts to 1996, US retirement plans to 2000, Indian derivatives accounts to 2026. There's no UK dataset of comparable depth anywhere in this, and the single UK figure available is a regulator's loss rate on contract for difference accounts, which is not a study of learning at all.
Trading records also aren't portfolios. These samples measure people who place orders, often many of them. Someone who holds three funds and adds to them monthly is barely represented, and nothing here settles whether investors do learn from their mistakes in a portfolio they rarely touch.
The attrition correction that does most of the work above carries a limitation the authors state themselves. They cannot observe how fast the quitters would have learned had they carried on. If quitting mostly removes people who would have learned slowly, the corrected estimate is the right one. If it removes people who would have improved, it is too low, and no amount of trading data settles that.
Both the Indian and Taiwanese datasets cover speculative, leveraged instruments, where the base rate of loss is high and the feedback is fast. That's the easiest place to see a mistake and the hardest place to survive one. It isn't obvious that the same arithmetic governs someone rebalancing a two fund portfolio once a year.
What would change the conclusion
A dataset that follows the quitters. Every estimate above is bounded by the same blind spot. Link brokerage records to pension and fund accounts, and you could see whether a failed stock picker who moved to index funds had learned the most valuable lesson on offer or simply left the sample. That distinction is unresolved here and it carries the entire result.
A measured reduction in a bias among people who kept trading. Feng and Seasholes's 37% is the shape of finding that would move this, and what it needs is a replication with an attrition adjustment attached. If the reduction survives that correction, learning by trading is a bigger effect than 3.6 basis points per 100 trades.
Cheaper mistakes. A good deal of the Indian loss rate is transaction costs rather than bad selection. The loss making share there rises from 82.1% before costs to 87.7% after them. Strip the costs out and many of the same decisions stop looking like mistakes, which changes what there is to learn from in the first place.
The variable worth watching isn't your own return history. The Finnish, Taiwanese and Indian records agree it carries very little information about your ability, and the Indian regulator found that having traded before made no material difference to the odds of losing money again. Turnover is the variable most consistently associated with worse outcomes across these studies, and unlike ability it is directly observable.