Informational and behavioural edge under continuous pricing
A prediction market settles zero-sum at resolution, which is widely read as proof that skill in it is unavailable. Zero-sum constrains the aggregate and says nothing about how that aggregate is distributed. The contract trades continuously for weeks before resolution, at odds that stream in real time, which turns a single draw into a long sequence of decisions and lets a small per-decision advantage compound the way it does in poker or in market making. That time dimension changes what is being traded: the object in a trader’s hands is closer to an option with a deadline than to a single settled bet. It has no underlying to replicate against, and its price path is generated instead by streaming odds, sentiment, and staked collective judgment, which is a mixture of populations moving at once. Three sources of variation follow, and each is tradeable: the structural (spread, depth, fees, time to resolution), the informational (who knows what, and when), and the decompositional (which population is currently moving the price). The consequence is the argument of this paper. A trader’s edge does not require being right at resolution, because entry timing, exit timing, and position management against a moving price are all live sources of return, and every one of them is a question about how a person trades. Most of what a trader needs to improve those three is already visible in their own order flow, and reading it back to them is worth more than any signal a platform could sell. This paper defines a trading persona as a working model of a specific trader estimated from their trades: what they trade, at what size, with what timing, how they respond to a loss, and whether their confidence tracks their accuracy. The signature is concrete: order size against the trader’s own distribution, the latency between a realized loss and the next order, the share of orders that take liquidity where a resting one would have filled, holding period against the contract’s remaining life, exposure concentrated in one correlated cluster, and the distance between implied and realized accuracy. It then puts the persona to two practical uses. The first is informational: for any given position a small number of facts are decisive (resolution wording and its edge cases, time to resolution against the trader’s holding pattern, depth at touch against their typical order size, cross-venue divergence, and the settlement source), and which few are decisive depends on the persona. The second is a recommender that reads the persona and fits the advice to it: a hedge for the impulsive sizer, a resting order for the habitual spread-crosser, a size derived from measured accuracy for the overconfident, a pause for the chaser. The paper states the claims as falsifiable propositions, says what data would refute each, and is explicit that they are hypotheses held by an engineer who has built the pipeline and lacks the account-level sample needed to test them.
A prediction market settles zero-sum. At resolution the contract pays one unit if the described event occurred and nothing otherwise, so the sum of all payouts equals the sum of all positions less whatever the venue and the makers take. That accounting is exact, and it describes a single instant. Everything before that instant is a different market. A contract can be bought and sold at any point in its life, at odds that stream continuously, and what a trader actually holds is a claim whose price moves for days or weeks before the moment when the zero-sum identity binds. The time dimension changes what is being traded. The tradeable object is closer to an option with a deadline than to a single settled bet, and the properties that follow from that are the subject of this paper.
That identity is the foundation of a familiar objection, and this paper accepts the arithmetic while disagreeing about what follows from it. The arithmetic is settled. A prediction market is zero-sum at settlement and negative-sum once fees are counted, so a participant earns only by taking from a counterparty. There is no drift and no beta in an event contract, nothing a patient holder is paid for carrying, and a basket bought and held to resolution returns zero at best before costs. The standard treatment reads the price as an information aggregator and the trade as a transfer (Wolfers and Zitzewitz), and on those terms it is correct. What the arithmetic does not establish, though it is routinely read that way in popular commentary, is that skill in the market is unavailable. Zero-sum constrains the aggregate. It says nothing about the distribution of that aggregate across participants.
What continuous pricing changes is the frequency of decisions, not the size of the pot. A contract bought once and held to resolution is a single draw: one decision, one outcome, and a result dominated by whether the event happened. The same contract quoted continuously for nine weeks is a sequence of decisions, because the position can be entered, sized, reduced, added to, and exited at prices that keep moving. The settlement identity binds the sum of terminal payouts and says nothing about the path of prices at which each participant moved through the position on the way there. The pot is fixed. The number of decisions that determine who holds which share of it is not.
That difference is what separates a lottery from a repeated game, and the consequence is a change in the shape of the outcome distribution rather than its total. A small per-decision advantage that is invisible in one draw compounds when it is applied hundreds of times, so results across participants spread out and concentrate: a minority accumulates most of the surplus and the median participant sits below the mean. That concentration is the signature of every zero-sum game that supports professionals. Poker is the cleanest case, since it is strictly zero-sum before the rake, negative-sum after it, and unambiguously a game of skill over a long enough sequence of hands. Market making is the same structure with a different surface. Nobody concludes from the settlement arithmetic of poker that the game cannot be played well, and the arithmetic of an event contract carries the same information: it constrains the aggregate and leaves the distribution to be decided by how each participant plays their sequence.
Calling the object an option with a deadline is a claim about its shape and not about its machinery, and the place where the comparison stops is worth drawing out honestly. An option’s price rests on an underlying: a sophisticated, continuously observable spot price with its own depth, its own history, and its own volatility surface. That underlying is exactly what makes Black-Scholes-style replication possible (Black and Scholes). Hold the option, hedge it with the underlying, rebalance as the price moves, and the payoff can be manufactured out of the hedge. An event contract has no such underlying. There is no continuously traded thing whose path the contract is a function of, so the replication argument has nothing to stand on. Its price path is generated instead by streaming odds, by sentiment, and by staked collective judgment, which is a mixture of populations moving at once where an underlying is a single quantity. The mechanics of an edge on an event contract are therefore particular to it, and an options desk earns its living by different machinery. What survives the crossing is the part the comparison was asked to carry: a position held through time is a sequence of separate decisions, and the settlement identity governs only the last of them.
That is what makes the price path structurally interesting. A mixture has parts, the parts move at different moments and for different reasons, and a trader who can name the parts is reading more information than a single spot price carries. That event prices aggregate information well is settled (Wolfers and Zitzewitz give the standard account), and it is a statement about where the price ends up. The path it takes to get there is the trader’s working surface. I set out the decomposition itself in earlier work (Pak), where an event price is read as the combined judgment of expressive bettors, hedgers, arbitrageurs, and informed flow, each entering the book with a different urgency and a different price sensitivity. Three sources of variation follow from a price path of that kind, and all three are tradeable.
The structural. Spread, depth, fees, and time to resolution. These are properties of the book and the clock, they are observable at every moment, and they are the most reliable of the three because they require no view on the event at all. A spread is a cost to whoever takes it and a return to whoever posts into it. Depth sets what a given size can actually be executed at. Time to resolution governs how much of the path remains and how long capital is committed to hold it.
The informational. Who knows what, and when. Information reaches an event market in discrete lumps: a filing, a poll, a ruling, a wording clarification. The trader who has read the resolution criteria and knows the settlement source is early on a price that has not moved yet, and being early on a path is a source of return that is settled long before the event is.
The decompositional. Which population is currently moving the price. A move driven by expressive flow ahead of a televised event carries a different expectation from an identical move driven by informed size, and the two leave different traces in order flow. Reading which population is on the bid is a live edge that has nothing to do with holding a view on the outcome.
The consequence is the backbone of this paper: a trader’s edge does not require being right at resolution. A position entered at 0.42 and exited at 0.51 pays the same whichever way the event eventually goes. Entry timing, exit timing, and position management against a moving price are all live sources of return, and each of them is a question about execution over a horizon. That is precisely the ground a trading persona covers, because persona work describes how a person trades and stays silent on what they predict. A forecast is a claim about the world. A persona is a claim about the trader, and on a market with a live price path the second is the one a platform can actually help with.
Against those three sources, a trader’s realized result separates into two parts with very different product implications. One part is set by the instrument. Fees and the venue take are explicit. The spread is implicit and larger than most traders estimate, because a marketable order pays half of it on the way in and half again on the way out. Adverse selection is the third and least visible: the resting orders a taker lifts are disproportionately those an informed participant was content to leave standing, so the realized cost of a fill can exceed the quoted spread. Glosten and Milgrom give the mechanism by which the quoted spread prices that risk, and Kyle gives the version where private information reaches the price only through informed order flow hiding inside everyone else’s. Call this bundle the structural component. It is a property of trading an event contract at all, and the return available in it comes from being on the side of it that collects: posting where the trader was crossing, sizing to the depth actually on the book, and choosing a horizon that matches the capital committed.
The second part is behavioural. It is what the specific way a person trades adds to the structural baseline: the timing of orders relative to their own state, the size chosen relative to the edge available, the order types used, the concentration carried, and the relationship between how sure they feel and how often they turn out to be right. This is the component that responds to information about itself, which is what makes it the subject of a product. The empirical case that it is worth real money comes largely from adjacent markets. Odean shows that individual investors trade far more actively than any reasonable model of informed trading would justify. Barber, Lee, Liu, and Odean, working with the complete Taiwanese trading record, attribute the largest single share of the performance gap between individual and institutional traders to aggressive orders, which is to say to a choice of order type that every trader makes on every single ticket and can change today. Thaler and Ziemba, and later Snowberg and Wolfers, document a century of longshot mispricing in parimutuel betting, the closest available analogue to an event contract, and a mispricing is two-sided: the same bias that leaves the longshot expensive leaves the favourite cheap.
Only one of the two is addressable by a product. Nothing a platform builds removes the spread it charges or the informed flow on the other side of the book. What it can do is put a specific trader on the better side of choices they are already making many times a week: capturing the spread on the entries where a resting order would have filled, sizing the order after a loss the way they size every other order, taking the longshot exposure at a horizon where the proportional spread is smaller. Two honest statements bound the ambition. Improving the behavioural component changes the distribution of a trader’s outcomes and the length of time their account stays in the game, and no system can promise anyone money. And some behaviour that looks like error is consumption: a trader who buys a longshot on a team they love has bought entertainment at a stated price, and a system that treats that as a defect will be wrong about them. The rest of this paper addresses the behavioural component, and keeps the consumption case visible throughout.
A trading persona is a working model of how a specific trader behaves, estimated from their own order flow, expressed as a small set of parameters with explicit uncertainty, and updated as trades arrive. It answers a short list of operational questions. What market classes does this trader take positions in, and how has that set changed? What is their typical order size, and what is the shape of the distribution around it? When do they place orders relative to the resolution date? Do they cross the spread or post? What happens to their next order after a realized loss: does it come sooner, larger, or in a different market class? Do they add to positions that have moved against them? And, over enough resolved positions, does the confidence implied by their sizing and the prices they pay track how often they turn out to be right?
Three properties define the construct. It is behavioural: every parameter is estimated from observed actions, so the model claims to describe what the trader does and stays silent on why. It is time-varying: a persona is a current estimate with a decay, since traders change after a big win, after a drawdown, and after a period away. And it is actionable by construction: a parameter earns a place in the persona only if some concrete intervention depends on its value. A dimension that no recommendation reads is a dimension that should not be estimated.
The natural comparison is a risk score of the kind a broker assigns at onboarding: a scalar, self-reported, fixed for years, used to gate access to products. That score answers a compliance question about what a person may be sold. A persona answers an operational question about what would help this person in the next five minutes, and it draws its answer from behaviour. That sourcing matters. Grinblatt and Keloharju matched a national trading record against psychological assessments and driving records, and found that sensation seeking and overconfidence predict trading activity directly, which is evidence that a trait measured outside the account shows up inside it. A persona runs the inference the other way, reading the trait from the record rather than the record from the trait.
Everything in the persona has to be computable from the venue’s own record: an ordered sequence of orders and fills per account, each carrying a timestamp, a market identifier, a side, a size, a limit price where one exists, an order type, the prevailing quote at submission, and eventually a resolution and a payout. The list below is the working set, with an honest note on how much weight each signal carries.
Order size relative to the trader’s own history. The absolute size of an order says almost nothing, since a $50 order is enormous for one account and rounding error for another. The z-score of an order against the trader’s own trailing size distribution is stable, interpretable, and the single most useful primitive in the persona. Most of the patterns in §4 are statements about when that z-score spikes. Order size read against a distribution is the same primitive the stealth-trading literature uses to locate informed flow in the medium-size band; here it is read for a different purpose, to find the sizes a person chose under pressure.
Time between a realized loss and the next order. A resolution that pays zero is an unambiguous, timestamped event. The elapsed time to the account’s next order, compared against that account’s baseline inter-order time, is a clean measurement with a clear behavioural referent. It is also one of the few signals where the counterfactual is easy to construct, because the same account provides its own control in the periods following a win.
Spread crossing. Whether an order was marketable at submission is recorded exactly. Aggregated into a rate (share of orders that took liquidity) and a price (the distance from the mid at submission, summed over fills), it converts directly into currency, and the same arithmetic run the other way gives the sum a trader would have collected by posting into the spread on the entries where a resting order would have filled. This is the signal with the shortest path from measurement to a number the trader can act on.
Holding period relative to the resolution date. Event contracts have a known terminal date, which makes holding period far more interpretable than it is in equities. The useful form is fractional: the share of the contract’s remaining life the trader held it. A trader who consistently enters at 80 percent of remaining life and exits at 20 percent has a different persona from one who enters at 10 percent and holds to settlement, and the two are paid by different things: the first is trading the path, the second is trading the outcome (§4).
Concentration. The share of open risk in the single largest position, and in the largest correlated cluster, computed at the moment each new order is placed. Correlation across event contracts is coarser than in equities but not absent: several markets on the same election, the same series, or the same regulatory decision move together, and a trader who believes they hold six positions may hold one.
Behaviour near expiry. The mix of prices paid in the final window before resolution, especially the share of volume at prices below roughly 0.10. This is where longshot preference becomes measurable per account, one level below the aggregate, and it is the account-level version of what Kumar found for lottery-type stocks.
Adds to losing positions. An order that increases an existing position at a price worse than the trader’s average entry, recorded with the mark at the time. The counterpart, selling winners early, is the classic disposition effect named by Shefrin and Statman and measured in brokerage records by Odean, and on an event contract the two are asymmetric in an interesting way, since a contract that has moved from 0.60 to 0.25 has a bounded remaining downside that makes the add feel cheaper than it is.
Cancel and replace patterns. These look diagnostic and mostly are not. High cancel rates mix genuine indecision with client-side reconnects, mobile app behaviour, market-making activity, and scripted order management. The rate alone separates automated from manual flow reasonably well, which is useful for the classification step in §8, and it is weak evidence about a human trader’s state of mind.
Anything inferred from a single trade. One large order after one loss is a coincidence at any reasonable significance level. The patterns in §4 are properties of distributions and need repeated instances before they mean anything, which is the whole subject of §7.
Limit-order behaviour read naively. Posting is cheaper than crossing in explicit terms and carries its own adverse selection, since a resting order fills exactly when someone wants the other side. Linnainmaa shows that limit orders alter the measured disposition effect and other inferences about investor behaviour substantially, because an unfilled limit order is a decision that leaves no trace in the fill record. A persona that scores posting as unambiguously good will overstate the gain, and Harris gives the standard practitioner treatment of the trade-off between the two order types.
Inactivity. A gap in the record is ambiguous between deliberate patience, loss of interest, and trading elsewhere. Multi-venue traders in particular will show a persona built from partial flow, and the partial view is systematically biased toward whichever venue happens to be cheapest for the strategy they run there.
The companion paper to this one decomposes an event price into the populations that set it: expressive flow that is buying an interest in the outcome, informed flow that is pricing something it knows, and the arbitrage and hedging flow that sits between them (Pak). A persona is the same decomposition performed on a single account. The question is which of those motives a given order belongs to, and the useful finding is that the same trader supplies all of them at different moments, so the mixture is estimable within an account rather than only across a book.
Three families cover the patterns worth acting on. They are separated by what the trader is deciding, because that is what determines the intervention: execution is what a position costs to enter and leave regardless of the view behind it, sizing is how much is committed relative to what the record supports, and state is what changes in the minutes after an outcome lands. The third column of Table 1 states what is available to a trader who changes each, with the mechanism drawn from adjacent markets and the magnitude unmeasured in this one. The magnitudes are the weakest part of the paper, since they come from equities and racetracks, and whether they transfer is exactly what §8 proposes to test.
| Family | Signature in order flow | What changing it is worth |
|---|---|---|
| Execution crossing the spread when posting would have filled |
High share of marketable orders in markets where the trader’s own holding period is long. Crossing to enter a position held for three weeks is the diagnostic case; crossing to exit before a resolution deadline is defensible. | The best-evidenced item in this paper and the only one measurable without any outcome data. Distance from mid at submission, summed across fills, is the amount paid; replaying each marketable order as a resting order against the recorded book gives the recoverable share. Barber, Lee, Liu, and Odean trace virtually all individual trading losses in a complete national audit trail to aggressive orders, while the same investors’ passive orders are profitable at short horizons. The gain is the half-spread avoided, net of the orders that would never have filled and net of the adverse selection a resting order accepts (Linnainmaa, Harris). On thin event books the half-spread is a material share of a low-priced contract, which is what makes order type the highest-yield single change available. |
| Sizing stake and concentration against the record |
Two signals. Implied confidence recovered by inverting a fractional-Kelly rule (Box 1) and compared against realized frequency. And the share of open risk in the largest position and the largest correlated cluster, computed as each order is placed. | Miscalibration is expensive because it is multiplicative: a trader who believes 0.75 when their record supports 0.60 carries the gap in the direction and again in the stake. Overconfidence is the oldest result in the calibration literature (Lichtenstein, Fischhoff, and Phillips; senses distinguished by Moore and Healy) and Odean’s standard explanation for heavy trading. Sizing to measured accuracy is the one adjustment available without changing a single view. Concentration is a separate matter and needs a separate justification: an event book has no market portfolio and no systematic factor, so the equity-market result that idiosyncratic risk earns no premium does not carry across. What concentration costs here is survival. Correlation in event markets often runs through the shared resolution source rather than the subject matter, so an account holding six positions may hold one, and naming the shared factor lets the trader size that exposure deliberately rather than discover it at settlement. |
| State decisions taken in the window after an outcome |
Time to the next order after a zero resolution falls below the account’s baseline and the size z-score of that order rises, often with a shift into an unfamiliar market class. Position increases at prices below the trader’s own average entry. Rising share of sub-0.10 buys in the final window before resolution. | The weakest evidence and the strongest mechanism. Thaler and Johnson showed experimentally that prior losses raise risk-taking when breaking even is in reach, and Imas shows the effect is specific to realized losses, which is exactly the event a resolution creates and therefore exactly the moment a venue can anticipate. Adding to losers is the disposition effect (Shefrin and Statman; measured in brokerage records by Odean) resting on the reference-dependence of Kahneman and Tversky. On an event contract the add is sharper, because the price moved when information arrived, so the trader is taking the other side of the flow Kyle identifies as the channel by which private knowledge reaches the price. Near-expiry longshot buying is the account-level form of the favorite-longshot bias surveyed by Thaler and Ziemba and weighed by Snowberg and Wolfers, and identifiable per account in the way Kumar shows for lottery-type stocks. The bias is durable partly because it resists arbitrage, so the gain available here is the spread and the horizon: the same exposure carried further from expiry pays a proportionally narrower spread. |
Evidence quality is not uniform across those three rows and the paper should not pretend otherwise. Execution rests on a direct measurement in a complete audit trail. Sizing rests on a large and consistent calibration literature plus an estimation procedure specified below. State rests on laboratory results and equity-market records, and its transfer to event contracts is an open question rather than a finding. A product built on this should ship the rows in that order.
Sizing deserves a figure, because it is the family with the most direct route from a measurement to a recommendation, and because the measurement is the one most easily done wrong. A trader who buys at 0.60 has revealed that their subjective probability is at or above 0.60, and nothing more. The price paid is a bound, not an estimate, so a calibration curve built from prices alone is not identified. Size closes the gap under a declared assumption. If a trader stakes a fraction of their bankroll according to a fractional-Kelly rule, then stake and price together pin the belief, and the Kelly fraction itself can be estimated from the trader’s whole record rather than assumed. Box 1 gives the inversion. The resulting curve is a conditional object: it measures calibration given that the trader sizes in a stable proportion to their edge, and a trader whose sizing is erratic shows up as a poor fit to the model, which is itself a persona parameter worth having.
The two directions of miscalibration look identical to the trader and take opposite advice. Confidence running ahead of accuracy is corrected by sizing down. Accuracy running ahead of confidence means size is being left unused. Both are improvements and only one of them involves trading less. Calibration is also trainable, which is the case for showing a trader their own curve at all: forecasters given repeated scored feedback on their own probability estimates improve measurably against untrained controls (Mellers et al.). This is also the one place where event markets are a better instrument than equities, since the underlying quantity resolves to a fact on a known date and calibration can be measured directly rather than inferred through a risk-neutral wedge (Pak).
Every trading product ships a feed. Prices, headlines, volume, related markets, other people’s positions, sentiment counts. Most of it is decoration, in a precise sense: revising the fact would not change what the trader does. A useful working definition is that a fact is decisive for a given position if a plausible revision of it either flips the sign of expected value or changes the recommended size materially. Everything else is context, and context has a cost, because a screen full of it hides the two or three items that would have mattered.
The standard picture of where a professional edge comes from is an information edge. Professionals win by knowing sooner, by knowing more precisely, or by knowing something the price has not absorbed yet, and the apparatus of research desks, data subscriptions, and expert networks exists to buy those three. That is information arbitrage, and Grossman and Stiglitz give the reason it persists: information costs money to acquire, prices can therefore reflect only what somebody was paid to put into them, and the return to acquisition is what covers the acquiring. A retail trader is seldom in that race on equal terms. On a prediction market the race can be starker still, because resolution often turns on a single fact that somebody involved knows early: a filing already drafted, a decision already taken in a room, a count already tallied.
Using information well is a separate skill from winning that race, and it is the skill available here. Most of what a retail trader is missing on a given position is public. It is the exact resolution wording and its edge cases, the settlement source and how it differs from the price on the screen, the depth available at the size they intend to trade, the time remaining measured against their own typical holding period, and the divergence between two venues quoting the same event. Every one of those is published, free, and consulted by almost nobody, which is what makes it a real and reachable edge. Genuine private information is neither reachable nor free.
For an event contract, that decisive set is short, and it sits almost entirely in the structural and informational sources of §1, with the news feed a long way behind. Each of the five is worth stating in full, since what makes it decisive is specific.
Resolution wording and its edge cases. The contract pays on the words. The event as the wording defines it and the event as the trader imagines it can come apart, and that gap is where the largest surprises live: what counts as an official announcement, which source is authoritative if two disagree, what happens if the event occurs after the deadline, whether a partial or contested outcome resolves yes, no, or void. A trader holding a position whose thesis depends on an edge case that the wording handles differently is holding a position they have misunderstood, and no amount of price data will tell them.
Time to resolution against the trader’s own holding pattern. Time is the deadline in the option with a deadline, and it is the one input a trader always has and rarely prices. A trader whose measured median hold is four days entering a contract that resolves in nine weeks is running a strategy they have never actually executed, and the relevant fact for them is the mismatch itself, which the resolution date in isolation never shows. This is the clearest case where the persona selects the fact: the same date is decisive for one trader and irrelevant for another.
Depth at touch against their typical order size. An order larger than the resting depth will walk the book, and on a thin event market the walk can cost several percent. The generic display shows the best bid and offer; the decisive version shows how far this trader’s usual size would move the price, which is a persona-conditioned rendering of the same order book.
Cross-venue divergence. When the same event trades at materially different prices on two venues, one of three things is true: the wordings differ, the settlement sources differ, or there is an arbitrage that capital constraints have left open. All three are decisive, and the first two are usually the answer. Divergence is therefore best surfaced as a prompt to read the wording, and it earns its place on the screen on that ground alone.
The settlement source. Who adjudicates, on what evidence, on what timetable, with what history of disputes. A price is a belief about the event compounded with a belief about the adjudicator, and traders routinely price only the first. For contracts resolved by a committee or an oracle with a track record, that record is a fact with direct bearing on expected value.
The product implication is a narrow one and worth stating plainly: for a given trader and a given position, select the two or three facts from this set whose revision would most change that trader’s decision, and show those. The selection is where the persona does its work, because which of the five is decisive follows from how the trader trades. A trader who habitually holds to resolution is paid on the wording and the settlement source, since those two settle what the contract becomes at the only instant that matters to them. A trader who exits early is paid on depth and cross-venue divergence, since those two set the prices actually available on the way out. Surfacing all five to everyone is a feed. Surfacing the two that would change this trader’s decision is the product. A trader with a short measured holding period gets the time mismatch first; a trader whose orders routinely exceed depth at touch gets the depth rendering first; a trader who has been burned by an edge case before gets the wording first. The ranking is a judgment call and it is testable, because a fact that is genuinely decisive should change behaviour when shown and should not when withheld, which is the design in §8.
A recommender that ignores the persona has one thing to optimize and will find it. Shown a trader who just took a loss, an engagement-maximizing system will surface the market most likely to draw another order, because that is what its objective rewards. The system is doing exactly what it was asked to do. Kleinberg, Mullainathan, and Raghavan give the general form of the problem: when users have inconsistent preferences across time, engagement optimization systematically serves the impulsive self at the expense of the reflective one. Barber and Odean showed the mechanism live in markets, where attention-grabbing stocks draw disproportionate individual buying, and Barber, Huang, Odean, and Schwarz show the same dynamic amplified by app design on a modern retail platform.
A persona-aware recommender has a better option available to it. Reading the same trader, it can recognize the post-loss state and offer the action that fits it. Table 2 is the mapping, and it is deliberately conservative: every intervention listed is a normal trading action that the trader could have chosen themselves, and each one is aimed at a source of return from §1 that survives regardless of how the event resolves.
| Persona signal | Recommended intervention | Scored on |
|---|---|---|
| Size z-score spikes after losses; short post-loss latency | Offer a partial hedge or an opposing leg sized to cap the drawdown, and present it while the size is still being chosen | Realized drawdown over the following month against a matched holdout |
| High marketable-order share in markets the trader holds for weeks | Default the order ticket to a resting order at or inside the touch, with the expected fill probability and the spread captured shown in currency | Spread paid per unit of exposure, and fill rate, so the gain is net of orders that never fill |
| Confidence implied at entry runs above realized accuracy | A recommended size derived from the trader’s own measured accuracy, with a fractional-Kelly cap as the natural form; the same rule raises the size of an underconfident trader who has been trading below their record | Return per unit of variance, and survival time in the account |
| Very short latency after a realized loss, order in an unfamiliar market class | Offer the wait: hold the ticket, show the trader their own post-loss statistics, and let the order through unchanged if they still want it. The same market is usually still there in an hour, at a price the trader will have had time to read | Realized outcome of post-loss orders, delayed versus immediate |
| Rising share of sub-0.10 buys in the final window before resolution | Show the trader their own historical realized rate on that price band next to the implied rate, and offer the same exposure at a longer horizon where the spread is proportionally smaller | Realized return on near-expiry longshot positions |
| Largest correlated cluster exceeds a threshold share of open risk | Name the shared factor explicitly (same election, same series, same settlement source) and propose the position that spreads the exposure | Portfolio variance and worst-week outcome |
| Sparse history, persona still wide (§7) | Recommend nothing behavioural. Surface only the structural facts from §5, which are correct for everyone | Nothing. This is the abstention case, and abstaining is the correct output |
The incentive conflict here should be named out loud. Platform revenue on most venues scales with volume: fees per contract, spread capture, or both. Trader interest scales with expected value net of costs, which for most traders means trading more selectively and paying the spread less often. A recommender that reads the persona sits exactly on that seam, and the fork in Figure 3 is real: the same model that identifies a trader as impulsive after a loss can be pointed at either branch, and the code for the two is nearly identical. The objective the system is scored on is what separates them.
The practical resolution is to score the recommender on the trader’s outcomes and to make that scoring the primary metric, with a place on the main dashboard. Concretely: a randomized holdout that never receives the intervention, a pre-registered primary endpoint expressed in the trader’s currency (net return, drawdown, or spread paid per unit of exposure), and a published gap between the treated and untreated groups. This is a stricter standard than most consumer software holds itself to, and regulators already apply it to advice in adjacent contexts: Regulation Best Interest requires a broker-dealer recommending a securities transaction to act in the retail customer’s best interest and to keep its own interest from coming first. A venue that builds this will have to accept that some of its recommendations reduce its own near-term revenue, and the argument for doing it anyway is that a trader who is getting better at the game plays it for longer and at larger size.
The estimation problem is dominated by scarcity. A typical retail account on an event venue may resolve a few dozen positions in a year, and several persona parameters need repeated instances of a specific conditioning event (a loss, an expiry window, an add) before they carry information. Any method that produces a confident persona from ten trades is producing noise with a label on it.
The workable approach is hierarchical. Start each new trader at a population prior estimated from the venue’s whole book, then let their own trades pull the estimate away from it in proportion to how much evidence they have supplied. This is standard partial pooling (Gelman and Hill), and its useful property here is that the shrinkage is automatic: a trader with eight trades stays close to the population, a trader with four hundred is described almost entirely by their own record, and nobody has to pick a hard cutoff.
Three practical rules follow from working with this. First, abstention is a first-class output. When the interval still covers the population value, the correct behaviour is to say nothing about that dimension and fall back to the structural facts of §5, which are correct for every trader regardless of persona. A system that always has something personalized to say is a system that is guessing.
Second, the parameters mature at very different rates. Spread crossing is estimable from tens of orders because every order is an observation. Post-loss behaviour needs tens of losses. Calibration needs enough resolved positions to fill several confidence buckets, which realistically means a hundred or more, and it is the last parameter to become usable. A product should expect to ship the cheap dimensions first and let the expensive ones arrive over months.
Third, personas decay. Weight recent trades more heavily and let the estimate widen during inactivity. A trader returning after three months away is closer to a cold start than their raw trade count suggests, and treating them otherwise produces the specific failure of confidently addressing a person who no longer exists.
The cold-start case is the majority of accounts on any growing venue, and it has a clean answer: the persona is the population prior, so ship the product that is right on average. Show the wording, the time to resolution, and the depth at touch, and default the ticket to a resting order. Those need no personalization to justify.
This section is the point of the paper. Everything above is a design, and a design of this kind is worth exactly as much as the evidence that its central claims hold. Four propositions follow. Each is stated so that a venue with account-level order flow can test it, and each is paired with the result that would refute it. They are currently hypotheses held by an engineer who has built the estimation pipeline and does not have the sample needed to test them, which is the reason for writing them down in this form.
Claim. Among traders with comparable activity levels, market mix, and account age, net return per unit of exposure varies with the chasing parameter (δi significantly negative or κi significantly positive in Box 1), the accounts without a detectable one do better, and the gap survives a control for turnover.
Confirms. A panel of accounts with at least one full year of flow, chasing parameters estimated on the first half and returns measured on the second, showing a monotone relationship between the chasing parameter and out-of-sample net return after conditioning on turnover, average position size, and market mix. The within-account version is stronger: the same trader’s post-loss orders underperforming their own post-win orders removes most cross-sectional confounds.
Refutes. The gap disappearing once turnover is controlled, which would mean chasing is a proxy for trading a lot and carries no information of its own. Or the within-account comparison coming out flat, which would mean the post-loss state changes the timing and size of orders without changing their quality. Either result would remove the basis for the delay intervention in Table 2.
Claim. For traders in the top quintile of marketable-order share, the spread accounts for a measurable and material fraction of gross return, and a meaningful share of that amount is recoverable, in the sense that a resting order at the touch would have filled within their own holding horizon.
Confirms. Direct measurement. Sum the distance from mid at submission across fills, express it as a share of gross return, and then replay each marketable order as a hypothetical resting order against the recorded book to estimate the fill probability within the trader’s own median holding period. This is the cheapest of the three propositions to test, since it needs no outcome data at all.
Refutes. A low replayed fill rate. If resting orders at the touch would rarely have filled on the venue’s actual books, then crossing was the price of participating, and the intervention returns nothing. Thin event markets make this a live possibility, and it is the outcome that would most change the design.
Claim. Per-account calibration error predicts subsequent net return after controlling for trade frequency, and it does so through size: the miscalibrated trader stakes more when they are more wrong. Implied probabilities are recovered by the fractional-Kelly inversion of Box 1 and scored with the standard Brier decomposition into calibration and resolution terms (Murphy). The sizing rule is an assumption and therefore part of the hypothesis: this proposition tests calibration and the adequacy of proportional sizing jointly, and a result should report the fit of the sizing model alongside the calibration term rather than only the latter.
Confirms. Calibration estimated on one period predicting return in the next, with the effect surviving a control for turnover, and with an interaction showing the effect concentrated in the accounts whose size varies most with implied confidence. A demonstration that recommended-size adjustments derived from measured calibration improve realized risk-adjusted return against a holdout would be the strongest form.
Refutes. Calibration error being uncorrelated with subsequent return once frequency is controlled, which would collapse the overconfidence dimension into the simpler overtrading finding and make the sizing recommendation redundant. Alternatively, calibration failing to persist across periods at the account level, which would mean the parameter is not a stable trait and cannot be personalized on at all.
Claim. Delivering the persona-fitted intervention improves the treated trader’s realized outcome relative to a randomized holdout, on a pre-registered endpoint, within a defined horizon.
This is the only proposition that requires an experiment in place of an observational panel, and it is the one that actually matters commercially. The design is standard: randomize at the account level, hold out a control that receives the structural facts of §5 and no behavioural recommendation, pre-register the endpoint, and run long enough to cover several resolution cycles. The failure mode to guard against is measuring the wrong thing: an intervention that reduces trading will improve return per trade almost mechanically while possibly reducing total return, so the endpoint has to be stated in currency over a fixed window. A ratio will move for the wrong reason.
Every pattern in this paper is a human one. Chasing, adding to losers, longshot preference, and overconfidence are descriptions of a person under stress, and a market-making bot exhibits none of them while producing order flow that can superficially resemble several. On venues with meaningful automated participation, mixing the two populations will bias every estimate in the paper, most severely the spread-crossing and cancel-rate measures, where automated flow dominates the counts.
Classification of order flow by participant type is an established exercise. Kirilenko, Kyle, Samadi, and Tuzun classify every account in an audit-trail dataset into categories including high-frequency traders, intermediaries, and fundamental traders, using trading volume and end-of-day inventory patterns. Hasbrouck and Saar identify algorithmic activity from strategic runs of linked messages in raw order data, and Hendershott, Jones, and Menkveld use normalized message traffic as an algorithmic-trading proxy. The features these approaches rely on (message-to-trade ratios, inter-order timing regularity, inventory mean reversion, activity around the clock) transfer directly to an event venue. The honest statement is that this classification step has to run first and has to be validated, and any persona work published without it should be discounted heavily. It is also the same machinery the decompositional source of §1 requires, so a venue that builds it once gets both.
A useful feature of this venue is that much of the required evidence is already public. Polymarket settles on a public chain, and its data interface serves per-wallet positions and activity to anyone who asks, with no credentials and no relationship with the venue. Each filled trade carries the wallet, the timestamp, the market, the side, the size in dollars, the price paid, and the settling transaction. Closed positions carry the average entry price and the realized result. Resolution outcomes are public by construction.
That is enough to estimate most of the persona and to run two of the four propositions. Order size against the trader’s own distribution, the interval between a realized loss and the next order, holding period against the contract’s remaining life, exposure concentrated in one event cluster, and the gap between implied and realized accuracy are all recoverable from public fills. Propositions one and three are therefore open to any researcher, on the venue’s full history, at the cost of the compute to assemble the panel.
One test comes before those two and is worth running first, because the rest of the paper rests on it. A persona is only meaningful if these features are stable within an account across time and separable across accounts. Estimate the feature vector on one period, re-estimate it on the next, and measure both the within-account correlation and the across-account spread. High stability and wide separation establish that a persona is a trait worth personalizing on. Low stability collapses the paper, since a parameter that does not persist cannot be used to fit anything to anybody, and it collapses it cheaply and early.
Two things stay out of reach of public data. Cancels and the orders that never filled are matched off-chain and leave no public record, and the prevailing quote at the instant of each submission requires an orderbook history that has to be captured as it happens, since a book that was not recorded cannot be reconstructed afterward. Proposition two needs both, which makes it a prospective study that begins when continuous capture begins. Proposition four needs a randomized holdout inside a live product and belongs to whoever operates one.
The honest summary is a split one. The behavioural and calibration claims are testable now, by an outside researcher, on public data. The execution-cost claim and the intervention claim need capture and a product respectively. Until those run, the epistemic status of the paper’s central claims is: mechanism well established in adjacent markets, magnitude unknown in this one, transfer untested.
Survivorship. Any persona estimated on accounts with enough history is estimated on the traders who lasted, and the ones who chased hardest are disproportionately the ones who stopped. This biases the measured value of every pattern downward, and it is the same problem Brown, Goetzmann, Ibbotson, and Ross document for performance studies generally. The mitigation is to build the panel from account opening and follow every account including the ones that go quiet, which is expensive and is the only version worth trusting.
Personas that never stabilize. A large share of accounts on any consumer venue place a handful of orders and stop. For them the persona is the population prior forever, and the honest product response is the abstention path in §7: structural facts only, no behavioural claims. A system that manufactures a persona for these accounts will produce confident nonsense at exactly the scale where nonsense is most visible.
Self-fulfilling recommendations. Once a trader is told they chase, their behaviour changes, and the parameter is thereafter measuring a mixture of the trait and the response to being told about it. This is Goodhart’s law in its behavioural form, where a measure that becomes a target stops being a good measure, and it also breaks the estimator, since the intervention contaminates the data the intervention is estimated from. Randomized holdouts are the only clean answer, and they are the reason §8 insists on them.
Correlation and causation. Every proposition in §8 except the intervention experiment is observational. A trader who chases and underperforms may be underperforming for a reason that also causes chasing, with inexperience the obvious candidate. Within-account comparisons and the standard panel controls help, and they do not close the gap; only the randomized intervention does.
The trades that were never placed. A persona built from executed orders sees only the decisions that survived to become orders. The trader who considered a position and closed the app, the one whose limit never filled, and the one who traded the same view on another venue are all invisible, and they are not missing at random: the unplaced trades are systematically the hesitant ones. This is sample selection on an unobserved variable, and it means the persona describes the trader’s executed behaviour, with their intentions left outside the model. Client-side signals (ticket abandonment, time on a market page, cancelled compositions) would narrow the gap and carry their own surveillance cost, which is a decision a venue should make deliberately and disclose.
The consumption case, again. Some of what this paper labels error is a purchase. A trader who buys a longshot on their team knows the price and is buying the afternoon. The persona can detect the pattern and cannot, from order flow alone, distinguish a costly mistake from a cheap pleasure. The design response is to make every intervention informative and refusable: show the trader their own record, and let the order through.
Conclusion. The argument is short. A prediction market settles zero-sum, and it trades continuously for weeks before it settles, which makes the tradeable object an option with a deadline whose price path is driven by a mixture of populations. Three sources of variation follow, and none of them requires a trader to be right at resolution: the structural, the informational, and the decompositional. All three are reached through timing, sizing, and order type, which are facts about how a person trades. Those facts are visible in the trader’s own order flow at the account level, and a persona is a modest, operational model of them: a few parameters with honest uncertainty, estimated by pooling toward the population until the trader’s own record earns the right to move them. It buys two things. It selects which of the small set of decisive facts to put in front of a specific trader, and it lets a recommender fit its advice to the person receiving it, provided the recommender is scored on the trader’s outcomes against a holdout.
The mechanisms are well established in equities and racetracks, the magnitudes in event markets are unmeasured, and the transfer is the open question. The three families are not equally supported, and a product should ship them in the order the evidence runs: execution first, where the effect is large and directly measurable, then sizing, then the state-dependent patterns whose transfer is still a hypothesis. Testing the first two needs a panel assembled from public per-wallet history and automated flow separated from human flow before anything else. Testing the interventions needs a randomized holdout inside a live product, which only a venue can run. The propositions in §8 are written so that a negative result is as publishable as a positive one, which is the condition under which the exercise is worth doing at all.