Heretic Research puts the next prediction market trade away from the front end
Heretic Research, an independent research firm focused on the intersection of crypto and AI, has published the full version of its third report arguing that the next wave of industry growth in prediction markets may come from backend infrastructure rather than trading interfaces.
The report covers 282,191 settled prediction markets through July 15, 2026 and asks a direct question: with platform competition already concentrated around traffic, licenses and liquidity, where does the next increment of industry value appear?
Its answer is the “results layer,” defined as the set of functions that compare rules, verify evidence, confirm outcomes and trigger payouts. Once those capabilities are standardized and reused across platforms, the report says, they form a second profit pool for the sector.
Heretic Research says prediction market front ends are getting crowded, while each new market still needs the same settlement work behind the scenes. Rules have to be interpreted. Evidence has to be checked. Outcomes have to be confirmed. Payouts have to be executed. Those repeated tasks, in the firm’s view, are now close to being standardized into reusable infrastructure.
The report says uncertainty in rules and settlement is already creating alpha. Its regression model found that 0.487% of disputed markets carried 8.64% of volume, while similar contracts across platforms continued to show persistent price gaps. Even so, shallow books and execution costs cap how much capital can actually be deployed into those trades. The more scalable opportunity, the report argues, is to package rule comparison, evidence verification, result confirmation and payout triggers into products that can be called repeatedly.
Heretic Research estimates observable annual revenue for the results layer at roughly $15 million to $37 million. If the same monetization covered all existing market activity, that figure would rise to about $64 million to $161 million. Under a more mature monetization level, the upper bound reaches roughly $456 million a year.
On investable exposure, the report says platform equity already reflects high front-end growth expectations, and that HYPE and ICE still offer limited direct capture of results-layer value. The more important signal to watch, it says, is which early-stage projects start winning cross-platform usage and convert that usage into attributable, renewable revenue.
A contract carries two probabilities, not one
The report starts from a simple claim: the price of a Yes contract is usually read as the probability that an event will occur, but that skips a variable that becomes decisive at settlement. Once the fact has occurred, the question is whether the platform’s rules recognize it. That recognition determines whether the contract pays.
Under that framing, every contract contains two probabilities:
- the probability that the event occurs
- the probability that the event is recognized as Yes under the rules
When rules are clear and evidence is consistent, the two sit close together. When deadlines, information sources, evidence standards or final review procedures diverge, they split. Heretic Research says that spread is not a technical glitch. It changes cash flows and can itself be treated as an asset.
The report builds its framework around three terms. Interpretive authority decides whether a fact can be recognized by the rules. “Rule spread” is the trading alpha created when that recognition bias enters prices. The results layer is the commercial form of that authority once rule comparison, evidence verification, result confirmation and payout triggers are turned into cross-platform interfaces. Heretic Research describes them as a chain: interpretive authority is the cause, rule spread is the market outcome, and the results layer is the business model built on top of it.
0.487% of markets entered dispute, but they carried 8.64% of volume
Measured by market count, disputes look like an infrequent operational issue. Heretic Research says that lens systematically understates their economic value. What matters is not how many disputes occur, but how much money is exposed when they do.
As of July 15, 2026, the firm says it drew 282,191 samples for dispute-rate and regression analysis from markets that had closed, posted at least $1,000 in final notional volume and retained UMA status-tracking data. UMA is described in the report as the onchain dispute arbitration mechanism used by Polymarket. Among that sample, markets that entered dispute accounted for just 0.487% of cases but 8.64% of final notional volume.
The conclusion was blunt: rule spread is concentrated in the most expensive markets.
Bigger markets showed higher dispute risk
When the sample is bucketed by final notional volume, dispute rates rise monotonically from 0.26% in the lowest tier to 30.14% in the highest. Heretic Research then runs a regression and says the positive relationship still holds after controlling for market year, category and duration. A 10x increase in single-market volume was associated with a roughly 2.27x increase in relative dispute propensity.
The report is careful about what that means. Because trading volume can include activity that happened after a dispute began, the result establishes correlation, not direction. High volume may attract dispute, or dispute may attract volume. Either way, the report says, it has become a recurring cost around large capital.
High-value trading pushes result certainty into pre-trade requirements
When a single event settles only a few thousand dollars, ambiguous rules can be handled ad hoc by operations staff. When the same event decides tens of millions of dollars, ambiguity changes payouts, market-making inventory and platform credibility at the same time.
That is why, the report says, four things need to be visible before an order is placed: whether rules are comparable, whether evidence can be verified, whether outcomes can be confirmed and whether payouts can be triggered. If any of those are missing, market makers charge for the uncertainty through wider spreads, smaller size or higher capital usage. Those costs are then pushed back to traders in the form of worse execution, less size and weaker capital efficiency.
That is the point of the 0.487% figure, in Heretic Research’s telling. The dispute rate is small by count, but it suppresses a far larger share of capital than that count suggests.
Interpretive authority is concentrated
The report says prediction market terms can never cover every fact pattern. Where the written terms stop, settlement becomes a question of interpretive authority. Whoever has final review can rewrite cash flows. That is the first source of rule spread inside a platform.
The Strategy case: an event happened, but the market did not recognize it
One example in the report involves a Polymarket contract tied to whether Strategy would sell bitcoin before 11:59 p.m. Eastern Time on May 31, 2026. Strategy later disclosed that it sold 32 BTC between May 26 and May 31.
The market still settled No.
Heretic Research says the dispute did not center on whether a sale occurred, but on whether the sale needed to happen before the deadline or needed to be publicly confirmed before the deadline. The original rule text was written around event timing, while supplementary guidance issued after June 1 brought public confirmation timing into the decision. According to the report, Galaxy Research’s review of the rule text and later guidance showed that traders were effectively dealing with two separate recognition standards.
The report describes that episode as a visible rule-shock event. Disclosure of the sale pushed Yes higher. A supplementary explanation then pushed it back down. The fact itself did not change. What changed was the probability that the fact would be recognized under the rules. In the report’s framing, the supplement and the vote that followed rewrote the cash flow.
Open participation did not remove concentration
Polymarket disputes go to UMA voting, and in theory any token holder can participate. In practice, the report says, voting power is allocated by token holdings. Citing a Bloomberg investigation into onchain voting, Heretic Research says that among more than 6,400 addresses that participated in dispute voting over three years, just nine wallets accounted for about half of all voting power.
The report also cites a Wall Street Journal investigation, as relayed by media reports, which found that in more than 300 disputes at least one voter also held a position in the related market.
Heretic Research says it does not need to prove manipulation to identify risk. The conflict itself is enough. A small set of wallets may both hold positions and participate in deciding the final settlement outcome. Once trading interest and adjudication power sit in the same hands, markets have to price that possibility.
A $750 bond can be enough to move a case into the dispute process
The report says Polymarket’s dispute documentation shows that a challenger generally only needs to match the proposer’s bond, commonly $750. For a trader with a large position, that cost can be trivial next to the amount of cash flow at stake.
Heretic Research gives a simple example: assume a trader holds a $100,000 Yes position and the market is proposed to settle No, which would leave that position close to worthless. Even if the trader believes the chance of winning a dispute is only 10%, spending $750 to bring the case into the dispute process may still be rational.
Ignoring research, gas and other added costs, a successful dispute would recover the $100,000 position and half of the losing side’s bond, or $375. A failed dispute would lose the trader’s own $750 bond. The report calculates expected value at about $9,362.5 in that example.
It writes the expected value formula as:
Expected dispute value = q × (V + 0.5B) − (1 − q) × B − C
where V is the recovered cash flow if the dispute succeeds, B is the dispute bond, C is research, gas and opportunity cost, and q is the estimated probability that the favorable interpretation is accepted.
The break-even condition becomes:
q > (B + C) ÷ (V + 1.5B)
As position value rises, the minimum success probability needed to justify a dispute falls. The report’s point is not that $750 decides a vote. It is that $750 can be enough to insert a favorable interpretation into the procedure that decides cash flow.
Rule spread already shows up in prices
The second source of rule spread, the report says, is differences in interpretive standards across platforms. The same “second probability” does not just become hard to predict. It becomes unequal from one venue to another.
That has turned rule spread into a tradable price distortion. Traders are not only asking whether an event will happen. They first need to know whether two contracts that look similar actually share the same deadline, source and settlement rule, and whether they would both pay under the same scenario. Only if settlement conditions are fully aligned can the gap be treated as arbitrage. If the rules differ, the trade is really a bet on whether the market has mispriced those differences.
Semantically equivalent contracts still trade apart
Heretic Research cites a 2026 cross-platform study covering ten major venues and more than 100,000 events. After manually checking natural-language descriptions, settlement semantics and time windows, the study found that some events were listed on multiple platforms and semantically equivalent markets still showed average persistent, executable price gaps of 2% to 4%. The study labeled the phenomenon “semantic non-substitutability.” Without a unified event identifier and outcome standard, prices do not fully converge.
The report then points to a more recent market observation in the same direction. Between July 1 and July 7, 2026, on year-end BTC $120,000 contracts that traded on both sides, Kalshi’s Yes price was about 2.08 percentage points higher than Polymarket’s throughout the week. Heretic Research says that Polymarket-Kalshi matched observation makes visible what would otherwise remain a theory: rules and platform structure enter the price.
Same settlement terms mean spread capture. Different terms mean judgment capture.
The report walks through a standard cross-platform setup. Assume platform A lists “Will bitcoin exceed $120,000 by year-end?” with Yes at $0.54, while platform B prices No on what appears to be the same event at $0.43. Buying both sides costs $0.97.
If the two contracts share exactly the same deadline, price source and settlement standard, one side will pay $1 regardless of outcome. As long as fees, slippage and funding leave total cost below $1, the trader can lock in the spread.
But if platform A settles on an index price at 23:59 UTC on Dec. 31 and platform B settles on a U.S. Eastern closing price, the contracts can reach different outcomes. Buying both for $0.97 is no longer riskless. The trader is judging whether the market has underestimated the chance that the two rule sets produce different results.
The report breaks cross-platform trading into two types:
- If settlement terms are identical, the edge comes from price dislocation.
- If settlement terms differ, the edge comes from understanding the rules better than the market does.
For the second type, Heretic Research says, the core variable is not whether the event happens but the probability that the contract is ultimately recognized as Yes under the rules.
It gives one example. If a Yes contract trades at $0.54 and total fees, slippage and funding per contract are $0.015, the trader needs to believe the final Yes-settlement probability is at least 55.5% for the trade to have positive expected value. The break-even recognition probability is simply market price plus all-in per-contract cost.
If the trader estimates the true settlement probability at 60%, expected profit is about $0.045 per contract. If the estimate is only 53%, the trade does not work, even if the underlying event itself still looks likely.
Arbitrage is real, but size is tiny
The report then shifts from theory to capacity. A strategy may identify real spreads, but can those spreads be monetized, and how much money can be put to work each time?
Heretic Research cites onchain research covering settled markets from April 2024 through April 2025 showing that in-market rebalancing and cross-market combination trades generated about $40 million in realized arbitrage profits. Mispricing in prediction markets is not merely theoretical, the report says. It can be found systematically and converted into actual gains.
Yet size remains constrained. A study spanning 173 NBA games reconstructed more than 75 million order-book snapshots and identified 290 combination-arbitrage segments. Median gross return per opportunity was only about 1.01%. Among those opportunities, 76.9% were limited by order-book depth, and the constrained opportunities could execute only 14.8 contracts on average. Another class of single-market anomalies lasted a median of just 3.6 seconds.
That matters. A 1% gross return sounds attractive, but if the book can only absorb around 15 contracts at close to $1 each, the trade only deploys about $15 and earns perhaps $0.15 before other considerations. The percentage return exists. The capital base does not.
Heretic Research’s conclusion is that trading rule spread is a business built on research and execution skill, not a limitless capacity trade. Contract analysis, semantic matching, actual book reading and multi-leg execution determine whether an edge can be captured. Book depth and opportunity duration determine how much capital can be scaled into it. The bigger opportunity, the report says, may be to charge repeatedly for reducing this uncertainty rather than trading each isolated instance of it.
The larger opportunity is to productize settlement
Heretic Research argues that one reason prediction market books remain shallow is settlement uncertainty itself. Market makers cannot know in advance how a platform will recognize a fact, so they protect themselves with wider spreads and smaller size.
That creates two different businesses from the same friction. Traders profit because the friction exists. The results layer profits by reducing it.
The report breaks result confirmation into three broad buckets. Facts such as prices, timestamps and onchain state can be read directly by machines. Cases where the rules are relatively clear but the evidence is scattered can be assisted by AI systems that retrieve and organize material. Cases with ambiguous rules, large stakes or conflicting evidence still need human final judgment. Productization begins when each task is standardized into a service that can be purchased and called repeatedly.
Deterministic outcomes are already being standardized
The easiest place to start is with highly verifiable events: the price of BTC at a given time, whether an address completed a transfer or whether a certain state existed before a deadline. If the source, read time and decision condition are defined in advance, a system can read the data, determine Yes or No and trigger payout without ad hoc interpretation.
Hyperliquid’s HIP-4 is presented as one attempt to turn that process into a reusable product. According to the report, HIP-4 brings binary outcome contracts into a unified trading system so rule definition, trading, result confirmation and payout can be handled inside one stack and then reused for the next batch of events.
Heretic Research says the demand side has already validated standardized outcome products. Chainlink-powered short-duration crypto markets on Polymarket have generated more than $3.4 billion in cumulative volume. Regulated venues are using a similar model. Some Kalshi gold, crude oil and agricultural contracts use Pyth-supplied external price data as the basis for settlement, showing that result confirmation can move from a platform-internal workflow to a repeatable service supplied by a specialist data provider.
AI’s role is to compress evidence cost
For many other markets, the issue is not that the rule is unknown but that the evidence is scattered across announcements, news coverage, regulatory documents or conflicting sources. In those cases, the expensive part is often not the final judgment but finding the relevant material, filtering valid evidence and presenting it clearly.
That is where AI fits best, the report says. It can accelerate retrieval, summarize sources, identify conflicts and produce an evidence package for final review. But Heretic Research says current testing does not support giving AI the last word.
A study cited in the report found that AI reached 89.58% consistency in post-dispute review of UMA cases, but only 33.88% recall in identifying cases before a dispute arose. The first figure suggests AI can often reproduce the final judgment after the record is complete. The second suggests it still misses most high-risk markets ahead of time.
The practical division of labor, the report argues, is straightforward: machines read deterministic data, AI retrieves and organizes evidence, and humans handle ambiguous rules, larger stakes and conflicting proof. Once that workflow is standardized, rule comparison, evidence verification, result confirmation and payout triggers become stand-alone interfaces for multiple platforms, market makers, institutions and agents.
How the results layer could make money
When those functions become reusable interfaces, revenue no longer comes only from final settlement. It comes from repeated calls across the trading chain. At that point, interpretive authority stops being only a governance question and starts producing a second profit pool.
The report says two conditions are needed for that revenue to stand on its own: the same capability must be used repeatedly across different markets, and platforms, institutions or developers must be willing to keep paying for those calls.
HIP-4 shows the product loop
Heretic Research says HIP-4 demonstrates the first condition. Result confirmation can be automated and then reused across different market types instead of rebuilt from scratch each time. The process runs like this:
event is written into rules → rules generate an outcome asset → the market trades → a machine confirms the outcome → funds are paid automatically → the template is reused for the next event
The report says HIP-4 posted about $277 million in cumulative volume in less than three months, including about $182 million over the last 30 days. That, in its view, shows users are willing to trade outcome assets confirmed by machines and settled automatically.
Its volume mix also matters. Sports accounted for about 82% of volume, but once large event peaks faded, activity shifted back toward crypto asset markets. Heretic Research reads that as a split between event-driven peaks from sports and steadier, higher-frequency baseline demand from crypto.
Three revenue paths are emerging
The report lays out three charging models for the results layer: result calls and payout triggers, institutional data and audit workflows, and result-data licensing and distribution.
It then points to four vendors already visible in the market:
- Chainlink: evidence that result interfaces can attract high-frequency usage. Chainlink-powered short-duration crypto markets on Polymarket have exceeded $3.4 billion in cumulative volume.
- Azuro: evidence that infrastructure usage can become real revenue. Azuro generated about $4.5 million in cumulative protocol revenue from roughly $530 million in cumulative volume, while supplying market creation, data and settlement infrastructure.
- Pyth: evidence that regulated venues will buy external result data. Kalshi uses Pyth data in some gold, crude oil and agricultural contracts.
- ICE: evidence that prediction market data can enter institutional information products. ICE has integrated Polymarket probability and sentiment data into real-time feeds and historical databases for institutional clients.
Taken together with HIP-4, the report says, those cases show a sequence that is already forming: outcome products are traded, interfaces are called repeatedly, platforms purchase external capabilities and result data enters institutional workflows.
Heretic Research also looks at market structure. In June 2026, it says, Kalshi, Polymarket US and Polymarket International produced a combined $44.8 billion in volume, implying a simple annualized run rate of $537.6 billion. Regulated markets accounted for 77.1% of that total and onchain markets 22.9%.
The revenue model differs by segment. Onchain markets split settlement across oracles, dispute systems and smart-contract payouts, making usage-based fees, settlement fees or protocol splits more natural. Regulated platforms keep final settlement responsibility, but can still buy external data, audit tooling and institutional distribution, which points more toward annual contracts, API subscriptions and data licensing.
From tens of millions to a possible $456 million
Because no one discloses “results-layer revenue” as a stand-alone line item, the report uses a three-step estimate.
First, the observable base. As of July 28, 2026, DefiLlama tracked 33 onchain prediction market protocols that generated about $146 million in revenue over the prior year and about $30.26 million over the last 30 days, according to the report. But not all of that belongs to the results layer. To isolate a supportable share, Heretic Research uses Azuro’s actual revenue split, where 10% of pool profit goes to data providers.
Applying that ratio yields roughly $14.61 million in results-layer revenue based on trailing annual revenue, or roughly $36.82 million using a simple annualization of the last 30 days. Rounded, the report places currently observable results-layer revenue at about $15 million to $37 million a year.
Second, expansion across the current market. Using the report’s earlier June volume figures, onchain markets imply annualized volume of about $123.1 billion. Dividing observable results-layer revenue by that volume gives a monetization rate of 1.19 to 2.99 basis points. In practical terms, the report says, every $10,000 of prediction market volume currently supports about $1.19 to $2.99 of results-layer revenue. If that rate applied across the full market, annual industry revenue would be about $64 million to $161 million.
Third, a mature upper bound. Azuro’s cumulative protocol revenue equals about 84.9 basis points of cumulative volume, and around 10% goes to data providers. That implies a results-layer monetization rate of about 8.49 basis points, or about $8.49 of results-layer revenue per $10,000 traded. If the whole prediction market industry reached that level, annual results-layer revenue would be about $456 million.
The report also says valuation multiples of 10x to 18x are drawn from data and financial infrastructure transactions involving IHS Markit, Black Knight and Adenza. Once full-year industry revenue for the results layer clears $100 million, Heretic Research says, that would imply enterprise value of roughly $1 billion to $1.8 billion and move the segment from early validation into an institutionally allocable stage.
What investors can actually own
The report then asks the unavoidable question: even if the results layer has industry value, can current assets capture it?
Platform equity is not the purest expression
The most obvious route is equity in prediction market platforms, but Heretic Research says that exposure includes far more than the results layer. Platform valuation also reflects traffic, licenses, liquidity, user distribution, brand and broad prediction market growth expectations.
It cites two valuation anchors. Kalshi completed a new round in May 2026 at a $22 billion valuation. Polymarket was valued at about $8 billion pre-money when ICE invested in October 2025, and ICE later made a direct $600 million investment in March 2026.
The report’s conclusion is explicit: platform equity offers exposure to prediction market growth, but not to the results layer in pure form. Its stated position is that investors should not chase platform equity at current levels.
HYPE offers platform growth, but HIP-4 is still too small
HYPE is one of the easier related assets to trade, but the report says it represents overall Hyperliquid platform growth, not stand-alone results-layer revenue. The key metric is not HIP-4 volume by itself, but HIP-4 fees as a share of Hyperliquid’s total fees.
The formula given is simple:
HIP-4 fee contribution = HIP-4 fees ÷ total Hyperliquid fees
Using $182 million in 30-day volume and a fee rate of 4 to 7 basis points, the report estimates HIP-4 monthly fees at about $72,800 to $127,400. Over the same period, Hyperliquid’s total fees were about $57.5 million. That leaves HIP-4’s contribution at just 0.13% to 0.22%.
Heretic Research says that is enough to validate product demand, but not enough to create a separate valuation increment for HYPE. Below 1%, it looks more like a product option. Above 1%, it begins to matter as a visible line of business. Above 5%, it could materially influence how HYPE is valued.
ICE is transparent, but the exposure is small
ICE is presented as the clearest public-market observation point. It owns Polymarket equity and has already integrated Polymarket data into institutional feeds and historical databases, which means both equity value and data-distribution value can pass into public financial statements.
As of March 31, 2026, ICE held Polymarket Series D/E preferred shares with a carrying value of about $2 billion, representing about 23% of issued shares and about 14% on a fully diluted basis, according to the report. In the first quarter of 2026, ICE recognized about $389 million in non-cash fair-value gains tied to observable price changes.
Still, Heretic Research says the prediction-market exposure is too small relative to ICE’s overall business to drive a major change in earnings or valuation. That makes ICE more useful as a window into whether results-layer commercialization is entering institutional budgets and public reporting than as a high-beta investment vehicle.
Early infrastructure projects are closer to the theme
The assets closest to the results layer itself, the report says, are oracles, dispute arbitration systems, distribution interfaces and specialized builder tools. Pyth and Chainlink provide deterministic data and machine confirmation. UMA and Kleros handle subjective outcomes and disputes. DFlow and similar projects link platforms, agents and application workflows.
Those projects are more focused and more likely to capture cross-platform usage directly. But they still face the same problem: product usage is not the same as revenue that clearly accrues to token holders or equity owners. Number of integrations, number of supported markets and call volume can prove demand, but not investability on their own.
The report says two signals matter most from here: whether a project can be reused across multiple markets rather than tied to one platform, and whether that usage turns into ARR, traffic fees or data-sharing revenue that clearly belongs to the relevant asset.
For now, Heretic Research says the results-layer logic is forming, but the ideal investable vehicle has not yet appeared.
The next stage may reward the infrastructure that decides payment
The report closes by saying prediction market growth has so far been concentrated on the front end, with more platforms, more categories and more competition for traffic, licenses and liquidity. As market size grows, though, the cost of rule interpretation, evidence verification and result confirmation rises with it.
Every platform building its own stack means duplicated spending and incompatible standards. Because different venues increasingly need to read the same facts, process similar evidence, interpret similar rules and bear the same settlement-error risk, Heretic Research says those functions are starting to look like shared industry infrastructure.
That is why the results layer matters in the report’s framework. It does not need to win every trader directly. It grows by serving trading demand that already exists as market count, contract variety and settlement frequency expand. Front-end competition decides where order flow goes. The results layer is the layer every platform may still need to call on the way to final payment.
There is no clear category leader yet, the report says, and revenue still does not accrue cleanly to one asset class. But the most valuable phase of an industry opportunity often comes before the investable names are fully formed. Heretic Research says the next thing to watch is which projects begin serving multiple platforms at once, which ones turn rules, evidence and result confirmation into standard products, and which ones convert that usage into durable revenue.
Its final point is that the first stage of prediction markets belonged to the expansion of trading entry points. The next stage, it argues, may send more value to the infrastructure that does not own the user relationship but decides how markets finish the final payout.