A $36,000 P&L gap between trading-time cohorts

In the March 2026 snapshot, traders active from 8-12 UTC averaged +$21,786 in cumulative P&L. Traders active from 16-20 UTC averaged -$14,099. The $35,885 difference compares cohorts; it does not prove that time of day caused the result.

The cohorts come from time-bucketed performance data on every trader who traded at least 5 markets within a given window, with at least 50 traders per bucket. 6 four-hour windows cover the full day, and 4 of them are profitable on average. The profitable windows cluster between 0 and 12 UTC, and the losing windows occupy 12-20 UTC. The spread between the best and worst window is about $36,000.

Win rates across the windows are close: 53.1% in the 8-12 UTC bucket and 54.1% in the 0-4 UTC bucket. The snapshot does not have the controls needed to attribute the P&L difference to spread width, counterparty mix, liquidity depth, trader selection, or anything else.

So time of day is a cohort descriptor here, not a trading prescription. A trader who changes hours may not reproduce the historical cohort average.

Every number comes from production data across 8,000 to 15,000 Polymarket wallets, depending on the metric. None of it is simulated.

The time-of-day analysis covers 6,421 to 7,539 traders per bucket: every wallet that traded at least 5 markets within a given four-hour UTC window. The largest bucket (12-16 UTC) holds 7,539 traders and the smallest (20-24 UTC) 6,421. They are not cherry-picked winners: every active wallet that meets the activity threshold is in, whether it made or lost money.

The duration analysis spans 2,814 to 9,029 traders depending on the band. The whole-cent fill, longshot, and specialization analyses use a stricter filter of 20+ markets traded, which leaves 606 to 4,605 traders per cohort. The winners-versus-losers comparison requires 10+ markets and captures 10,581 traders: 4,981 profitable and 5,600 unprofitable. The grade-tier analysis uses 0xinsider's grades (0xinsider.com/leaderboard) across the S through D tiers.

Two things matter about this dataset. It is large enough that single outliers do not drive the averages: a trader making $10 million in the 8-12 window could skew a 100-trader sample, but not a 7,298-trader one. And it includes losers. Much trading analysis studies the winners and drops everyone else; this dataset keeps every qualifying wallet, including the ones that lost money and stopped trading.

Every metric here comes from 0xinsider's analytics pipeline, which turns raw on-chain Polymarket activity into structured trader metrics. One caveat: the time-of-day, duration, and whole-cent-fill cohorts below come from the original study run and have not been recomputed, because the table that produced them has been superseded. Treat them as a March 2026 snapshot. Two things have been re-queried against production, on July 12, 2026: the grade-tier figures, and the specialization analysis, which no longer supports the original conclusion. Any trader's metrics open from their row on 0xinsider.com/leaderboard.

The full breakdown: 8-12 UTC, 7,298 traders, 53.1% win rate, +$21,786 average P&L. 4-8 UTC, 6,870 traders, 54.0%, +$14,148. 0-4 UTC, 6,463 traders, 54.1%, +$9,829. 20-24 UTC, 6,421 traders, 53.9%, +$3,080. 12-16 UTC, 7,539 traders, 53.7%, -$1,795. 16-20 UTC, 7,180 traders, 53.8%, -$14,099.

The four cohorts from 0-12 UTC and 20-24 UTC had positive average P&L; the 12-16 and 16-20 UTC cohorts had negative averages. Trader count does not explain that ordering: the 8-12 UTC cohort had 7,298 traders, while the largest cohort, 12-16 UTC, had 7,539.

The snapshot did not measure trader location, experience, retail participation, spread width, or competition by cohort. Claims that one group was more skilled, less crowded, or trading in better conditions would be hypotheses, not findings from this dataset.

The original duration table has been superseded, so the current code cannot show whether these bands measured position holding time, a first-to-last trading window, or something else. The same trader can appear in more than one band. The figures below are averages across traders assigned to a historical cohort, not per-position results.

The 1-hour-to-1-day cohort had the highest average P&L in the snapshot: 7,856 traders, a 62.9% win rate, +$256,691 average P&L, and +$472 expectancy per trade across the cohort. The under-1-hour cohort held 9,029 traders, with a 26.2% win rate and -$148,027 average P&L.

The other cohorts: 1-3 days, 5,516 traders, 61.1% win rate, +$27,278 average P&L, and +$283 expectancy; 3-7 days, 4,521 traders, 58.9%, +$4,867, and +$157; 1-4 weeks, 4,733 traders, 59.8%, -$19,146, and +$699; over 4 weeks, 2,814 traders, 59.9%, -$178,487, and -$2,424.

These historical cohorts show an association. They do not identify the cause of the differences or say how long to hold a position, and the missing source-table contract rules out a stronger conclusion.

Every Polymarket fill has a maker and a taker, but this study cannot see that role for third-party wallets. The public Data API has no liquidity-side field, and the authenticated CLOB trade history covers only the account that makes the request. The cohorts below use whole-cent fill share instead. That price pattern may correlate with order behavior, but this dataset does not measure the relationship.

The study splits 8,545 traders with 20 or more markets by the share of their fills at whole-cent prices. In the over-70% cohort, 1,649 traders averaged $9,156 in P&L, with a median of -$17 and a 49.8% win rate. Those figures describe the price-pattern cohort. They do not show whether its members provided liquidity or earned the spread.

The 50-70% cohort holds 1,880 traders. It averaged $52,715 in P&L, with a +$182 median and a 51.3% win rate. That average is nearly 6 times the over-70% cohort's and more than 3 times the under-30% cohort's.

The 30-50% cohort holds 1,987 traders, averaging $20,764 in P&L with a $0 median and a 50.8% win rate. The under-30% cohort holds 3,029 traders, averaging $16,079 with a $0 median and a 51.7% win rate. These results do not show whether order roles, fees, or spread costs caused the differences.

The median is the clearest result in this breakdown. Only the 50-70% cohort has a meaningfully positive median, at +$182; the over-70% cohort sits at -$17 and the two lower bands at $0. So the center of that one distribution was positive. The superseded source does not preserve the aggregation method or a profitable-trader count.

The March 2026 snapshot shows an association between the 50-70% whole-cent-fill band and stronger P&L. It does not prove a maker-taker mix or prescribe an order strategy. Your own whole-cent fill share is on your 0xinsider profile page.

Longshots here are outcomes trading below 20¢ a share. A 5¢ contract that resolves to $1 returns 20 times the stake, and that payoff is what draws traders to them. The cohort results run the other way. Heavy longshot traders (50%+ of trades in sub-20¢ contracts) have the lowest win rate of the four longshot cohorts, 46.3%, and an average P&L of only $9,941 across 671 traders.

Avoiding longshots entirely did not do best either. Traders who put under 10% of their activity into longshots, the "mostly favorites" cohort, number 4,605 and averaged $23,172 in P&L with a 54.0% win rate. Two cohorts did better and are effectively tied: the 10-30% group averaged $34,633 (1,995 traders, 49.9% win rate) and the 30-50% group $33,075 (606 traders, 47.4%). A $1,558 gap across cohorts of this size is noise, and we will not build a recommendation on it.

The two got to a near-identical total by different routes. The 30-50% group earned $23,315 from its longshot trades, the most of any cohort, while the 10-30% group earned only $4,426 from longshots and the rest from the other side of its book. The pattern is an association: cohorts with 10% to 50% of trades in longshots averaged more than cohorts under 10% or over 50%.

The heavy longshot group shows the other side. Its win rate is 7.7 percentage points below the mostly-favorites group. Its longshot trades made $13,771 on average, but the rest of its book lost money, leaving $9,941 overall. The data cannot say why.

One known bias fits the pattern. People overweight low-probability, high-payoff outcomes, which Kahneman and Tversky described as probability weighting in prospect theory. Our calibration study found that bias on Polymarket's cheapest contracts: positions bought near 5¢ won 3.5% of the time (0xinsider.com/research/how-efficient-are-prediction-markets).

The honest read is a range, not a number: somewhere between 10% and 50% of trades in contracts below 20¢, with the two cohorts inside that band less than $1,600 apart in average P&L. The limit is clearer. Above 50%, average P&L drops to $9,941 and the win rate to 46.3%. An earlier version of this article named 30-50% as the single optimal allocation; the cohort averages do not support that precision.

An earlier version of this article reported that moderately specialized traders were the top performers, and told you to concentrate 50-80% of your activity in a few categories. That conclusion does not survive its own numbers, so I am retracting it out loud instead of quietly deleting it.

The original cohorts, by Herfindahl index of concentration, as average P&L: diversified (below 30%) $21,859; moderate focus (30-50%) $23,545; specialized (50-80%) $27,158; ultra-specialized (above 80%) $25,090. Read that way it is a tidy inverted U and a neat recommendation. Now read the medians of the same four cohorts: -$2, $3, $7, and $0.

A cohort whose median member has made $7 does not have an edge, and the best-looking cohort sits inside the noise. A handful of outliers carried the averages in each bucket, and averaging P&L across a long-tailed distribution will manufacture a gradient out of noise every time. Re-queried against production on July 12, 2026, the picture is unchanged: specialists (above 70% concentration) show a median P&L of -$42 and generalists (below 20%) -$31. Both are about zero, with generalists marginally ahead.

The honest conclusion is that category concentration, measured this way, does not predict profitability. Focus may still pay off for you personally, because it is hard to price a market you do not understand. But it does not separate winners from losers at the population level, and this article no longer claims that it does. Our guide to comparing traders explains what a claim like this needs before it can stand (0xinsider.com/research/what-separates-winning-polymarket-traders).

The dataset splits into 4,981 profitable and 5,600 unprofitable traders (10+ markets each). These groups are sorted by their results, so some differences below are effects of winning rather than causes. Profitable traders averaged $13,811 in volume per market against $7,189 for unprofitable ones, 92% more capital behind each position.

The activity gap is wider. Profitable traders averaged 1,356 markets traded against 488, a 2.8x difference, and 226 open positions against 91. Their biggest win averaged $41,755, against $8,386 for the unprofitable group, a 5x gap that reflects sizing as much as anything else.

The win rate split is 58.5% against 41.0%, a 17.5-point gap. At even odds, 100 trades at 58.5% produce 58 or 59 wins, and at 41.0% produce 41. Over 1,000 markets the first trader compounds gains and the second compounds losses. The lower market count of the unprofitable group is as much a symptom of losing as a cause: wallets that lose run out of capital or stop.

The grade tiers from 0xinsider's ranking (0xinsider.com/leaderboard) add another view. As of July 12, 2026, there were 266 S-grade wallets, 3,166 A-grade, 6,998 B-grade, and 34,967 C-grade, out of 100,898 graded. The figures are medians, because outliers distort the averages. The median S-grade wallet spent a 0.720 share of measured days within 5% of its positive running P&L peak, against 0.348 at C-grade and 0.033 at D-grade. That metric does not measure how jagged the curve is: a wallet whose cumulative P&L never rises above zero scores 0. The median S-grade calibration edge is +3.46%; at D-grade it is -0.03%.

The grade tiers also differ on whole-cent fill share: a 50.0% median for S, A, and B-grade wallets against 35.7% for C-grade. The measure records fill prices, not maker or taker roles, so it cannot explain why the tiers differ.

Taken together, the original study associates stronger results with 8-12 UTC activity, 1-hour-to-1-day duration, a 50-70% whole-cent-fill share, and 10-50% longshot exposure. None of those associations guarantees the next trade. A +$472 cohort-wide expectancy per trade in the duration bucket is not a forecast for one position, and copying a group's measured features does not prove you will reproduce its returns.

Every metric here is available for your own wallet. Search your Polymarket address or username at 0xinsider.com/polymarket to see your time-of-day distribution, duration breakdown, whole-cent fill share, category concentration, and longshot allocation. The data updates as new on-chain activity is processed.

Then ask where you fall. Are your fills concentrated in one UTC window? Which duration bands appear in your activity? What share of your fills land on a whole-cent price? How much of your activity is in longshots? Your profile reports those measures; it does not infer whether you posted or crossed an order.

I built 0xinsider's analytics pipeline to surface these patterns, and the grade-tier metrics are there too. The 266 S-grade wallets carry a median calibration edge of +3.46%: their positions resolved true about 3.5 points more often than the price they paid. That is the benchmark. If your calibration edge is negative, as it is at D-grade (-0.03%), the market prices outcomes at least as well as you do, and better timing or execution is unlikely to make up the difference. Calibration comes first.