The discipline inversion: how real money rewired 1,647 Manifold traders

In September 2024 Manifold Markets, the largest play-money prediction market on the internet, turned on a real-money variant called sweepcash. Selected markets had dual currencies for six months. The same users. The same UI. The same question catalog. Only a different currency on a subset of contracts. On March 28 2025 Manifold shut the program down.

This report is about what happened in between.

We pulled every bet Manifold has ever registered, 9,115,727 of them in the public bulk dump plus 8,989,431 from the /v0 API to close the twenty-one-month gap to the present, indexed 130,091 markets, re-fit the Convexly Edge Score composite on the resulting cohort of 15,106 users, and then tested whether removing and re-adding real money to the same forecasters changed how they bet. It did.

Specifically: across the 1,647 users who cleared the twenty-resolved-bet floor in both the pre-sweepcash bridge window and the sweepcash era itself, the median within-user concentration (the share of realized profit sitting in the trader's single largest event) fell by 8.9 percentage points when real money arrived. Across the 1,839 users who were active in both sweepcash and the post-sweepcash window, concentration rose by 3.5 percentage points when real money was removed. Per-bet calibration, measured as the user's Brier score net of their own marginal baseline, did not move. What moved was sizing.

The direction is the opposite of the naive intuition. Real money does not make forecasters take bigger swings. It makes them take less concentrated ones.

This is the first within-person causal test of incentive structure on a prediction-market venue that we are aware of. It is also only one of three cross-venue findings in this paper that argue the same thing, in different dialects: incentive regime is a first-order determinant of how the same measurement framework lands on the same data.

Data

The Manifold cohort is constructed from two sources. The public bulk dump provides every bet and every contract from the platform's launch in December 2021 through July 6 2024: 9.1 million bets, 130,091 markets, 56,139 unique users. We parse the bets file with a two-pass disk-sharded ingestion (first pass computes shard assignments by the first two characters of the userId; second pass rewrites bets into 1,296 per-shard files totalling 2.9 GB). The markets file is streamed with jq and reduced to the fields needed for resolution joins. The /v0 /bets?username= endpoint backfills the gap: 8.99 million bets covering July 2024 through April 2026, fetched with a 32-worker thread pool at roughly twenty bets per second under Manifold's documented 500-per-minute per-IP rate limit.

After merging, we apply the V1-M inclusion criteria. Resolved BINARY cpmm-1 markets and resolved MULTIPLE_CHOICE cpmm-multi-1 markets with a specific winning answerId. Non-redemption, non-cancelled bets only. A floor of twenty resolved bets per user. The final cohort is 15,106 users whose aggregated per-user statistics are released alongside this post at /research/v1m/v1m-data-bundle.tar.gz; user identifiers are hashed and raw per-bet records are not redistributed.

The Polymarket comparison cohort is the frozen V1 cohort from Convexly (2026): 8,656 wallets with at least five resolved positions, active through April 18 2026. Per-pillar values and fitted coefficients come from the committed validation report in the Convexly repository.

Kalshi trade-level data is pulled directly from the public /trades endpoint at api.elections.kalshi.com: 2.48 million trades in the working sample, anonymized, priced in USD notional using yes_price * count for yes-taker legs and (1 - yes_price) * count for no-taker legs.

Three pillars, two venues, different coefficients

The Edge Score V3b composite is a three-feature, OLS-fit ranking on signed log PnL, introduced in Convexly (2026). The three features are:

• Posture, the z-score of minus skill_brier. Positive values mark users whose profit is earned despite imprecise calibration.
• Conviction, the z-score of log concentration. Positive values mark portfolio concentration in a single dominant event.
• Discipline, the z-score of minus log position count. Positive values mark fewer, larger bets.

Fitted on Polymarket, the coefficients were +0.79, +2.72, and -1.15. Conviction dominated by a factor of three-to-one. Discipline loaded negatively, which is to say: on a real-money venue where the fitted cohort is a positive-profit leaderboard, users who hold fewer resolved positions earn more signed PnL per unit of raw feature value.

Refit on the 15,106-user Manifold cohort with the V3b-M multi-outcome extension (one-vs-rest Brier for MULTIPLE_CHOICE cpmm-multi-1 markets; backward-compatible with V1 on binary-only data within floating-point precision), the same three features produce three different coefficients.

Three observations, in ascending order of how surprising they were to us.

Posture is smaller on Manifold, same sign. The play-money platform rewards calibration, just not as much as the real-money platform does. This is the least surprising finding in the paper and the easiest to rationalize: Manifold makes per-user calibration plots a first-class product feature, so users who optimize for the visible leaderboard metric would be expected to show a positive posture coefficient.

Conviction collapses by a factor of seventeen. The dominant pillar on Polymarket drops to noise on Manifold. What concentrates PnL on real money barely moves the needle on play money.

Discipline flips sign with a permutation p of 0.0001 across 10,000 shuffles. On Polymarket, more positions predicts lower realized PnL. On Manifold, more positions predicts higher realized PnL. The bootstrap 95% CI bands (-1.21 to -1.09 on Polymarket per the V1 validation report; +0.23 to +0.50 on Manifold V1-M) do not overlap.

We stratified the Manifold cohort by bet count to test whether the discipline flip is an artifact of the long tail of low-activity users. It is not. The coefficient holds the same sign in every activity stratum (+0.07 among 20-50 bet users, +0.08 among 51-200, +0.37 among 201-1,000, +0.37 among 1,000+). In the whale subcohort the posture coefficient swings up to +1.03, comparable in magnitude to Polymarket's +0.79. Conviction in the whale subcohort reaches +0.94, still well below Polymarket's +2.72. The direction of the discipline flip is structurally stable.

The methodology transfers. The coefficients do not.

Fat tails below a mathematical boundary

Hill alpha is a standard tail-index estimator for heavy-tailed distributions. The convention in financial-market literature is to read alpha as “how fat is the tail.” Equities sit around 3.5. Hedge fund returns cluster at 2 to 3. Below 2, the variance of the return distribution is formally undefined. Below 1, even the mean is undefined.

The V1 paper reported alpha = 1.28 on per-user PnL for the 8,656-wallet Polymarket cohort, 95% bootstrap CI (1.20, 1.36). The same estimator applied to the 15,106-user Manifold V1-M cohort returns alpha = 0.86, 95% bootstrap CI (0.80, 0.93). The CIs do not overlap. The play-money venue is measurably more fat-tailed than the real-money venue by this measurement.

Before explaining why this is a surprising finding at all, it is worth acknowledging what alpha = 0.86 means arithmetically. It means the population expected value of Manifold per-user PnL under this empirical tail shape is not well-defined. Any point estimate of average per-user profit in the cohort is an accident of where the sampling stopped. The one-in-a-million outcome is not well-characterized by the one-in-a-thousand one.

The intuition the naive reader brings to this comparison is that real money makes people take bigger risks. Under this intuition, the real-money venue should have fatter per-user tails than the play-money venue. The data argue the opposite. Across 15,106 Manifold users and 8,656 Polymarket wallets, per-user PnL is more fat-tailed on the play-money venue by a gap that survives bootstrap resampling and does not depend on cohort-construction choices we tested.

Two mechanisms are consistent with this direction. First, play-money Manifold has no ruin barrier. A user who loses 99% of their mana continues to trade. A Polymarket wallet that loses 99% of its capital typically stops. This produces different survivorship structure across the two visible cohorts, and that survivorship structure biases the measured tail shape in different directions. Second, the average Manifold balance is small (500 to 5,000 mana is typical) and the average market is thin (100 to 1,000 mana total pool), so a single serious bet can dominate a user's realized PnL. This is the same mechanism visible in the median-concentration inversion: Manifold users concentrate more within their own portfolios (median 0.87 versus Polymarket's 0.51) and the platform as a whole concentrates the heavy-hitter tail in a smaller share of users.

The bet-level Hill alpha narrows the gap but does not close it. On per-bet notional (mana amount on Manifold, USD notional on Kalshi, no Polymarket equivalent available in our data), alpha sits at 0.91 on Manifold against 1.07 on Kalshi (CI 1.068, 1.076; tail n 248,493). At the same aggregation level the venues are closer together than at the per-user level. The difference between the aggregation levels (user-level gap of roughly 0.40 alpha units versus bet-level gap of roughly 0.15 alpha units) implies that approximately half of the user-level divergence between Polymarket and Manifold is not explained by bet-size distribution. It is explained by how bet flow concentrates across users.

The sweepcash natural experiment

The cross-venue coefficient divergence and the cross-venue Hill-alpha gap both carry an obvious confound: the users on Polymarket and the users on Manifold are not the same people. Identity-matching across the two venues was attempted in the paper via a three-tier heuristic (exact username, bio-referenced cross-venue identity, shared linked social accounts) and yielded zero CERTAIN, zero PROBABLE, and two CANDIDATE matches on an 8,698-wallet pilot. A within-person cross-venue test is not possible with the available data.

What is possible is a within-person cross-incentive test, thanks to Manifold's own sweepcash program.

From September 25 2024 through March 28 2025, selected Manifold markets had a real-money variant redeemable for US dollars, operating in parallel with the mana markets on the same platform. Then the program shut down. We partition the Manifold bet stream into four windows by timestamp: bulk (before July 2024), gap_pre_sweepcash (July through September 2024), sweepcash (September 2024 through March 2025), and post_sweepcash (after March 2025). For each window, we compute the same per-user statistics we used in the full-cohort analysis, subject to a ten-resolved-bet floor within the window.

Window-level Hill alpha on per-user PnL:

The sweepcash-era alpha of 1.05 is statistically indistinguishable from the gap-pre-sweepcash bridge alpha of 1.02 and is closer to the Polymarket V1 alpha of 1.28 than to the bulk-window play-money baseline of 0.88. After sweepcash ends, alpha reverts to 0.84, with a CI whose upper bound (0.96) barely touches the sweepcash-era CI's lower bound (0.90). The window-level story is consistent with the hypothesis that real-money incentives, applied to the same platform and in many cases the same users, shift the tail shape of per-user PnL in the direction of real-money venues elsewhere.

The within-user comparison tests the same hypothesis at the individual level. For users active in both pre-sweepcash and sweepcash windows (n = 1,647), we compute the delta of their concentration and skill_brier metrics across windows, user by user. The median within-user deltas:

The concentration delta reverses sign with the incentive regime. When real money arrives, the same user shows 8.9 percentage points less concentration; when real money departs, the same user shows 3.5 percentage points more concentration. The skill_brier deltas are near zero in both directions, which is to say: incentive regime does not cause users to improve or degrade their per-bet forecast accuracy. What changes is sizing. The forecasts stay where they were. The money around them moves.

This is, to our knowledge, the first within-person causal demonstration that incentive structure on a prediction market shifts trader concentration behavior without shifting trader calibration. Servan-Schreiber, Wolfers, Pennock, and Galebach (2004) compared real-money TradeSports to play-money NewsFutures on 208 NFL games and found no market-aggregate accuracy difference, which is the closest prior result in the literature. Their comparison was at the market-aggregate level, across different populations on different platforms. The present comparison is at the individual level, across the same population on the same platform, separated only by whether a subset of contracts resolve for dollars.

What it does not show

The finding here is narrower than the popular summary “real money makes people more careful.”

It is true that real money reduces concentration in a measurable, within-person way on this sample. It is not true that real money improves calibration in our measurement (it doesn't move). It is not true that real money makes users more profitable in a way we can measure (median realized PnL within each window is negative and noisy, as it is on any cohort whose enrollment is conditional on losses being realized). It is not true that users who perform well on one venue will perform well on another. Our identity-match cohort is too small to test cross-venue within-person skill transfer, and that remains a gap in the cross-venue prediction-market literature.

It is also not true that the Edge Score “doesn't transfer” in the strong sense. The measurement framework transfers. We re-ran the same three-pillar decomposition, the same z-scoring procedure against cohort-local moments, the same OLS fit against signed log PnL, the same 10,000-shuffle permutation null, on the same kind of input data, and got sensible, confidence-interval-bounded outputs on both venues. The pillars are well-defined on Manifold. The z-scores are well-defined on Manifold. The OLS fit converges. What does not transfer is the magnitude and in one case the sign of the fitted coefficients. An analyst who ported Polymarket's coefficients directly onto Manifold data would rank users in roughly the opposite order on the discipline axis and would overweight conviction by an order of magnitude. The artifact would look like a model. It would not be one.

Where the methodology lands

Two takeaways follow from the results above.

The first is for anyone who studies skill in prediction markets empirically. Any claim that a skill measure “generalizes” across venues needs to commit, up front, to whether it is claiming framework portability (the same pillars, the same feature definitions, the same statistical apparatus produce well-defined outputs on each venue) or coefficient portability (the same numerical coefficients rank users comparably across venues). On our evidence, framework portability holds and coefficient portability does not. A paper that claims the latter owes the reader a coefficient-stability test. A paper that claims only the former owes the reader a venue-specific refit protocol.

The second is for anyone who operates a prediction market or builds analytics on top of one. The per-venue incentive regime is a first-order determinant of the structural relationships between trader-level features and profit outcomes. It is not a second-order nuance. A tool that scores a Polymarket wallet's skill and then rescores the same user's Kalshi or Manifold activity using the same coefficients is not producing a consistent score; it is producing three different measurements under a shared label. This is a solvable problem (refit per venue, version the coefficients, expose the coefficient set alongside the score) but it has to be solved intentionally, not assumed away.

These findings are not claims about individual trader talent. They are claims about what happens to a measurement when the environment the measurement is computed in changes. The finding that real-money sweepcash produced a within-person concentration delta of minus 8.9 percentage points does not tell you that a particular Manifold user became “more careful.” It tells you that the median of a cohort of 1,647 users moved. That movement is enough to change how a skill-composite scores the same cohort on two different instantiations of the same platform, which is enough to argue that a cross-venue product that ignores the regime is misspecified.