1. Introduction
1.1 Prediction Markets
The pricing of macroeconomic events has long been the domain of professional financial markets. Institutional traders in interest rate futures have spent decades translating Federal Reserve policy expectations into probability estimates through futures pricing. Tools such as the CME FedWatch Tool have made this process more accessible to the broader public. However, over the past several years, a parallel system of market-based forecasting has emerged in the form of prediction markets. Platforms like Polymarket allow participants to trade binary contracts on real-world outcomes, producing their own implied probability estimates derived not from futures pricing but from bets of thousands of individual traders. When applied to the same event, such as the probability of a Federal Reserve rate cut at a specific FOMC meeting, both markets produce a probability-like estimate intended to answer the same question. The fact that these numbers frequently disagree is the starting point of this article.
Prediction markets operate by allowing participants to trade contracts whose payoff depends on whether a given outcome materialises, effectively pooling dispersed information into a single price signal. The market-clearing price, expressed as a probability between zero and one, reflects the collective expectation of traders willing to put capital behind their views. Prediction markets are well-established and are effective at aggregating dispersed private information, often outperforming expert forecasts and surveys because participants have a direct financial incentive to be accurate. Research in this area has also shown that prediction market prices incorporate new information quickly and are largely resistant to manipulation. On these grounds, a Polymarket contract on a Fed rate cut is not merely a speculative instrument; it is a meaningful probability estimate produced by a market with real money at stake.
1.2 CME-Implied Probabilities
CME Fed Funds futures operate on a different logic but produce a similar output. The CME FedWatch Tool derives implied probabilities from 30-Day Fed Funds futures contracts, whose prices reflect the market's expectation of the average Effective Federal Funds Rate over a given month. By comparing the implied rate before and after a scheduled FOMC meeting, one can extract the probability assessed by the market of a rate change at that meeting. Unlike Polymarket, this is a market dominated by professional fixed-income traders, where deep liquidity, narrow bid-ask spreads, and continuous price updating driven by macroeconomic events and Federal Reserve communication are the norm. The CME-implied probability can, in this sense, be interpreted as the institutional market’s estimate of what the Fed will do.
What makes the comparison between the two venues interesting is that, despite pricing the same event, they differ in nearly every other respect. Polymarket attracts a primarily retail, event-driven participant base. Its contracts are binary and subject to liquidity conditions that can be limited and variable depending on the event in question. The CME, by contrast, is a deeply liquid exchange populated by professional traders for whom rate expectations are crucial to their positioning. These differences create room for the two markets to arrive at different probability estimates and, crucially, to correct those differences at different speeds.
1.3 The Puzzle: Divergence Between Two Markets
The pricing of macroeconomic events has long been the domain of professional financial markets. Institutional traders in interest rate futures have spent decades translating Federal Reserve policy expectations into probability estimates through futures pricing. Tools such as the CME FedWatch Tool have made this process more accessible to the broader public. However, over the past several years, a parallel system of market-based forecasting has emerged in the form of prediction markets. Platforms like Polymarket allow participants to trade binary contracts on real-world outcomes, producing their own implied probability estimates derived not from futures pricing but from bets of thousands of individual traders. When applied to the same event, such as the probability of a Federal Reserve rate cut at a specific FOMC meeting, both markets produce a probability-like estimate intended to answer the same question. The fact that these numbers frequently disagree is the starting point of this article.
Prediction markets operate by allowing participants to trade contracts whose payoff depends on whether a given outcome materialises, effectively pooling dispersed information into a single price signal. The market-clearing price, expressed as a probability between zero and one, reflects the collective expectation of traders willing to put capital behind their views. Prediction markets are well-established and are effective at aggregating dispersed private information, often outperforming expert forecasts and surveys because participants have a direct financial incentive to be accurate. Research in this area has also shown that prediction market prices incorporate new information quickly and are largely resistant to manipulation. On these grounds, a Polymarket contract on a Fed rate cut is not merely a speculative instrument; it is a meaningful probability estimate produced by a market with real money at stake.
1.2 CME-Implied Probabilities
CME Fed Funds futures operate on a different logic but produce a similar output. The CME FedWatch Tool derives implied probabilities from 30-Day Fed Funds futures contracts, whose prices reflect the market's expectation of the average Effective Federal Funds Rate over a given month. By comparing the implied rate before and after a scheduled FOMC meeting, one can extract the probability assessed by the market of a rate change at that meeting. Unlike Polymarket, this is a market dominated by professional fixed-income traders, where deep liquidity, narrow bid-ask spreads, and continuous price updating driven by macroeconomic events and Federal Reserve communication are the norm. The CME-implied probability can, in this sense, be interpreted as the institutional market’s estimate of what the Fed will do.
What makes the comparison between the two venues interesting is that, despite pricing the same event, they differ in nearly every other respect. Polymarket attracts a primarily retail, event-driven participant base. Its contracts are binary and subject to liquidity conditions that can be limited and variable depending on the event in question. The CME, by contrast, is a deeply liquid exchange populated by professional traders for whom rate expectations are crucial to their positioning. These differences create room for the two markets to arrive at different probability estimates and, crucially, to correct those differences at different speeds.
1.3 The Puzzle: Divergence Between Two Markets
The December 2025 (350–375 vs decrease_25) pair is a clear illustration of the dynamic explored in this article. The two series diverge through October and November, with CME running persistently above Polymarket, before a sharp joint collapse in mid-November temporarily realigns both markets. What follows is a rapid convergence towards certainty as decision day approaches, with both series closing above 90%. This demonstrates both the persistence of mid-cycle gaps and their eventual closure. By contrast, the June 2026 (350–375 vs no_change) comparison is instructive because it remains unresolved. CME has consistently priced a higher probability of no change than Polymarket across the sample window, with the gap widening further from mid-February onward, suggesting institutional traders repriced faster and more decisively than prediction market participants, consistent with the lag hypothesis this article sets out to test.
To interpret such divergence, it is important to keep in mind that prices in prediction markets should not be read as exact probabilities because frictions, trader composition, and differences in market design all create a wedge between prices and true probabilities. Even when individual participants are systematically biased, the market as a whole can still aggregate information effectively, suggesting that Polymarket's retail composition does not disqualify it as a forecasting tool. Moreover, local mispricings can coexist with broader market efficiency as distortions produced by irrational traders are not necessarily permanent.
1.4 Hypothesis and Article Structure
Taken together, the above suggests a specific and testable hypothesis: that large gaps between CME-implied and Polymarket-implied probabilities are temporary, and that the two series tend to converge as the event approaches and information becomes more symmetric across participant groups. Whether that convergence is driven primarily by the prediction market adjusting towards the institutional benchmark or by some other dynamic is not predetermined; it is an empirical question with direct implications for how practitioners should interpret divergences between the two markets in real time.
This article tests that hypothesis using data from the 2025–2026 FOMC meeting cycle. We begin by situating the comparison within the literature on prediction market efficiency and the conditions under which temporary mispricings arise. We then examine the structural characteristics of the two markets and what a persistent gap between them can mean. From there, we lay out the data and methodology before presenting the empirical results, their robustness, and what they suggest about information transmission between retail and institutional markets.
To interpret such divergence, it is important to keep in mind that prices in prediction markets should not be read as exact probabilities because frictions, trader composition, and differences in market design all create a wedge between prices and true probabilities. Even when individual participants are systematically biased, the market as a whole can still aggregate information effectively, suggesting that Polymarket's retail composition does not disqualify it as a forecasting tool. Moreover, local mispricings can coexist with broader market efficiency as distortions produced by irrational traders are not necessarily permanent.
1.4 Hypothesis and Article Structure
Taken together, the above suggests a specific and testable hypothesis: that large gaps between CME-implied and Polymarket-implied probabilities are temporary, and that the two series tend to converge as the event approaches and information becomes more symmetric across participant groups. Whether that convergence is driven primarily by the prediction market adjusting towards the institutional benchmark or by some other dynamic is not predetermined; it is an empirical question with direct implications for how practitioners should interpret divergences between the two markets in real time.
This article tests that hypothesis using data from the 2025–2026 FOMC meeting cycle. We begin by situating the comparison within the literature on prediction market efficiency and the conditions under which temporary mispricings arise. We then examine the structural characteristics of the two markets and what a persistent gap between them can mean. From there, we lay out the data and methodology before presenting the empirical results, their robustness, and what they suggest about information transmission between retail and institutional markets.
2. Literature Review
The emergence of prediction markets across an increasingly broad range of macroeconomic and political outcomes has meaningfully expanded their potential utility in causal inference, particularly for institutional investors and academic researchers seeking real-time probability estimates exogenous to survey-based methods. Wolfers and Zitzewitz (2004) document encouraging evidence of informational efficiency, cross-exchange price consistency, internal coherence across related contract families, and the absence of exploitable time-series patterns, while carefully stopping short of claiming full efficiency, simultaneously documenting longshot bias, partisan trading behaviour, and cases where markets faithfully aggregated systematically misleading public information. This nuanced empirical picture becomes more consequential as prediction markets scale: broader event coverage means that prediction markets’ use as a gauge for market-based expectations is growing in importance. In the context of comparing CME-listed contracts with retail-dominated venues such as Polymarket, this literature directly motivates examining whether informational efficiency and cross-market consistency hold across structurally different trading environments.
Information aggregation, incentives for truthful revelation, and incentives for information discovery (Snowberg et al. 2012) form the three pillars that guarantee the functioning of prediction markets, justifying their informational value as much as their role as betting platforms. However, prediction markets are inherently imperfect. Their prices may be distorted by behavioural biases, by events for which dispersed information is limited, by concentrated insider information, and by ambiguity in defining the winning conditions of a market; an issue that becomes more pronounced as the breadth of tradable events expands. When comparing prediction market platforms, another key differentiating factor is market design, specifically the mechanism through which contracts are structured, matched, and settled. Prediction markets therefore reflect expectations in a broadly effective way, but prices should not be interpreted as exact probabilities. These frictions are likely to differ systematically between regulated, institutionally intermediated venues such as CME and more flexible, retail-oriented platforms like Polymarket, providing a natural setting to test how market design influences price quality.
Hayek’s hypothesis, formulated in 1945, holds that markets aggregate dispersed private information through the price mechanism despite individuals possessing incomplete knowledge. Forsythe, Rietz and Ross (1999) provide empirical support for this in prediction markets, even when most traders exhibit biases, a result explained by the marginal trader hypothesis: prices are set by a relatively small subset of rational traders actively posting competitive limit orders. Oliven and Rietz (1995) identify these as market makers, who commit significantly fewer errors than price-taking participants. This mechanism is particularly relevant when contrasting CME, where professional liquidity provision is central, with retail-heavy platforms such as Polymarket, where a thinner layer of sophisticated market makers may weaken the strength of Hayekian information aggregation.
Oliven and Rietz (2004) further show that even in markets that are efficient in aggregate, individual traders frequently violate no-arbitrage conditions and the law of one price, with such violations occurring in a substantial share of transactions. Importantly, these distortions tend not to persist, as rational market makers face strong incentives to correct mispricing to avoid adverse selection. This dynamic, where local inefficiencies are endogenously corrected by better-informed traders, creates conditions under which temporary deviations from equilibrium can emerge and subsequently converge. In the context of CME versus Polymarket, this implies that structural differences in participant composition and liquidity provision may generate short-lived cross-market mispricings, which can be arbitraged away. This directly motivates an empirical investigation of price convergence and the feasibility of systematic trading strategies across these venues.
Information aggregation, incentives for truthful revelation, and incentives for information discovery (Snowberg et al. 2012) form the three pillars that guarantee the functioning of prediction markets, justifying their informational value as much as their role as betting platforms. However, prediction markets are inherently imperfect. Their prices may be distorted by behavioural biases, by events for which dispersed information is limited, by concentrated insider information, and by ambiguity in defining the winning conditions of a market; an issue that becomes more pronounced as the breadth of tradable events expands. When comparing prediction market platforms, another key differentiating factor is market design, specifically the mechanism through which contracts are structured, matched, and settled. Prediction markets therefore reflect expectations in a broadly effective way, but prices should not be interpreted as exact probabilities. These frictions are likely to differ systematically between regulated, institutionally intermediated venues such as CME and more flexible, retail-oriented platforms like Polymarket, providing a natural setting to test how market design influences price quality.
Hayek’s hypothesis, formulated in 1945, holds that markets aggregate dispersed private information through the price mechanism despite individuals possessing incomplete knowledge. Forsythe, Rietz and Ross (1999) provide empirical support for this in prediction markets, even when most traders exhibit biases, a result explained by the marginal trader hypothesis: prices are set by a relatively small subset of rational traders actively posting competitive limit orders. Oliven and Rietz (1995) identify these as market makers, who commit significantly fewer errors than price-taking participants. This mechanism is particularly relevant when contrasting CME, where professional liquidity provision is central, with retail-heavy platforms such as Polymarket, where a thinner layer of sophisticated market makers may weaken the strength of Hayekian information aggregation.
Oliven and Rietz (2004) further show that even in markets that are efficient in aggregate, individual traders frequently violate no-arbitrage conditions and the law of one price, with such violations occurring in a substantial share of transactions. Importantly, these distortions tend not to persist, as rational market makers face strong incentives to correct mispricing to avoid adverse selection. This dynamic, where local inefficiencies are endogenously corrected by better-informed traders, creates conditions under which temporary deviations from equilibrium can emerge and subsequently converge. In the context of CME versus Polymarket, this implies that structural differences in participant composition and liquidity provision may generate short-lived cross-market mispricings, which can be arbitraged away. This directly motivates an empirical investigation of price convergence and the feasibility of systematic trading strategies across these venues.
3. Comparing CME and Polymarket
3.1 The Object of Comparison
The central question of this paper is whether the probabilities implied by Polymarket contracts on Fed Funds rate moves differ systematically from those implied by Fed Funds futures, as indicated by CME’s FedWatch Tool.
3.2 CME FedWatch and Institutional Price Formation
CME has been running its well-known FedWatch Tool for over 13 years now, and has since become the go-to source for inferring the likelihood of Federal target rate movements. Practically, it derives meeting-by-meeting probabilities for future target-rate outcomes from the prices of 30-Day Fed Funds futures contracts. In other words, FedWatch does not collect opinions directly; rather, it transforms futures prices into a distribution of possible policy outcomes.
The logic is as follows. The tool compares the implied rate from the futures contract with the current Fed Funds target range, assuming that any change in rate will come in the standard 25-basis-point (0.25%) increment used by the Fed. Then, from the outstanding Fed Funds futures prices, it calculates the probability of each potential outcome: hold, one cut, two cuts, one hike, and so on. For meetings further in the future, the tool extends this framework through a probability tree, linking the possible outcomes of one meeting to those of the next. Consider this example: suppose the current target range is 3.50–3.75%, and the Fed Funds futures contract for the next FOMC meeting month implies a rate of 3.60%. That implied rate is closer to the current range than to a 25 bps cut, which would bring the range to 3.25–3.50%. In this case, the tool would assign a high probability to a hold and a lower probability to a cut.
The market for 30-day Fed Funds futures is populated mainly by banks, asset managers, and other institutional investors that have the liquidity to trade these contracts, which are worth roughly $400,000 each. The implications this has for information quality are non-trivial.
Trades in 30-day Fed Funds futures by large institutional players are often the result of research efforts by highly skilled teams with access to the best infrastructure and highest-quality information. The large ticket size also leaves less room for pure speculation. Let us also not forget that many of these institutions are highly regulated and face tight restrictions on directional speculation. There is also little to no room for “manipulation” through the purchase of large positions.
3.3 Polymarket and Retail Price Formation
Polymarket probabilities result from a different process. Instead of inferring them from an underlying interest-rate futures contract, it reports the market price of a binary event contract, where the contract payoff depends directly on whether a specific policy outcome occurs. The price of a “Yes” contract can be interpreted as the market-implied probability of that event, although subject to frictions such as liquidity and, even if to a lesser extent, transaction costs. Thus, the probabilities we see in Fed-rate-decision markets are the result of thousands of transactions between market players with differing views. Some of them may hold very hawkish views, others may be willing to take the opposite side out of conviction, others may trade despite not having enough information, and others may constantly be on the hunt for mispriced events. Polymarket could even be used for cross-platform hedging purposes; i.e., one could buy a 5-cent “Yes” on a rate cut to hedge a rate-hike bet.
By contrast, as mentioned already, on Polymarket the maximum payout on a single yes/no contract is $1, so the cost of one contract cannot exceed $0.99. It is becoming increasingly common for hedge funds to deploy prediction-market strategies, but their activity in the space is nowhere near what would be required to cut through the noise produced by the retail crowd. At the same time, since the size of these prediction markets is far smaller, it is, first, cheaper to alter significantly the “priced” probability of an event occurring and, second, far harder to extract clear information from prices when the market also reflects $5 bets.
3.4 Why Cross-Market Comparison Is Informative
This fundamental difference is a striking example of why comparing CME and Polymarket is informative. Although both markets produce probability-like measures about future Fed decisions, they do so through fundamentally different mechanisms. Comparing CME and Polymarket, therefore, allows us to test whether a decentralised prediction market delivers signals consistent with those extracted from a traditional benchmark derivatives market, or whether it embeds systematically different expectations. 30-day Fed Funds futures prices, and by extension the CME FedWatch Tool, act as a sort of “involuntary” yet necessary check on the amount of speculation and potential misinformation that can be embedded in prediction-market probabilities. The gap between CME and Polymarket discussed in the next section reflects these structural differences.
The central question of this paper is whether the probabilities implied by Polymarket contracts on Fed Funds rate moves differ systematically from those implied by Fed Funds futures, as indicated by CME’s FedWatch Tool.
3.2 CME FedWatch and Institutional Price Formation
CME has been running its well-known FedWatch Tool for over 13 years now, and has since become the go-to source for inferring the likelihood of Federal target rate movements. Practically, it derives meeting-by-meeting probabilities for future target-rate outcomes from the prices of 30-Day Fed Funds futures contracts. In other words, FedWatch does not collect opinions directly; rather, it transforms futures prices into a distribution of possible policy outcomes.
The logic is as follows. The tool compares the implied rate from the futures contract with the current Fed Funds target range, assuming that any change in rate will come in the standard 25-basis-point (0.25%) increment used by the Fed. Then, from the outstanding Fed Funds futures prices, it calculates the probability of each potential outcome: hold, one cut, two cuts, one hike, and so on. For meetings further in the future, the tool extends this framework through a probability tree, linking the possible outcomes of one meeting to those of the next. Consider this example: suppose the current target range is 3.50–3.75%, and the Fed Funds futures contract for the next FOMC meeting month implies a rate of 3.60%. That implied rate is closer to the current range than to a 25 bps cut, which would bring the range to 3.25–3.50%. In this case, the tool would assign a high probability to a hold and a lower probability to a cut.
The market for 30-day Fed Funds futures is populated mainly by banks, asset managers, and other institutional investors that have the liquidity to trade these contracts, which are worth roughly $400,000 each. The implications this has for information quality are non-trivial.
Trades in 30-day Fed Funds futures by large institutional players are often the result of research efforts by highly skilled teams with access to the best infrastructure and highest-quality information. The large ticket size also leaves less room for pure speculation. Let us also not forget that many of these institutions are highly regulated and face tight restrictions on directional speculation. There is also little to no room for “manipulation” through the purchase of large positions.
3.3 Polymarket and Retail Price Formation
Polymarket probabilities result from a different process. Instead of inferring them from an underlying interest-rate futures contract, it reports the market price of a binary event contract, where the contract payoff depends directly on whether a specific policy outcome occurs. The price of a “Yes” contract can be interpreted as the market-implied probability of that event, although subject to frictions such as liquidity and, even if to a lesser extent, transaction costs. Thus, the probabilities we see in Fed-rate-decision markets are the result of thousands of transactions between market players with differing views. Some of them may hold very hawkish views, others may be willing to take the opposite side out of conviction, others may trade despite not having enough information, and others may constantly be on the hunt for mispriced events. Polymarket could even be used for cross-platform hedging purposes; i.e., one could buy a 5-cent “Yes” on a rate cut to hedge a rate-hike bet.
By contrast, as mentioned already, on Polymarket the maximum payout on a single yes/no contract is $1, so the cost of one contract cannot exceed $0.99. It is becoming increasingly common for hedge funds to deploy prediction-market strategies, but their activity in the space is nowhere near what would be required to cut through the noise produced by the retail crowd. At the same time, since the size of these prediction markets is far smaller, it is, first, cheaper to alter significantly the “priced” probability of an event occurring and, second, far harder to extract clear information from prices when the market also reflects $5 bets.
3.4 Why Cross-Market Comparison Is Informative
This fundamental difference is a striking example of why comparing CME and Polymarket is informative. Although both markets produce probability-like measures about future Fed decisions, they do so through fundamentally different mechanisms. Comparing CME and Polymarket, therefore, allows us to test whether a decentralised prediction market delivers signals consistent with those extracted from a traditional benchmark derivatives market, or whether it embeds systematically different expectations. 30-day Fed Funds futures prices, and by extension the CME FedWatch Tool, act as a sort of “involuntary” yet necessary check on the amount of speculation and potential misinformation that can be embedded in prediction-market probabilities. The gap between CME and Polymarket discussed in the next section reflects these structural differences.
4. Data
4.1 Data Sources
The analysis relies on two daily time series for each FOMC meeting under consideration: implied probabilities derived from CME 30-day Fed Funds futures, and market prices from Polymarket prediction markets. Both series are observed at end-of-day frequency, and each dataset covers the period from the opening of the corresponding Polymarket contract to the day of the FOMC decision. In total, five meetings are examined: December 2025, January 2026 and March 2026, which have resolved, and April 2026 and June 2026, which were still unresolved at the time of writing.
4.2 CME-Implied Probabilities
The first series is constructed following the CME FedWatch methodology, which derives the market-implied probability distribution over possible FOMC policy outcomes from the settlement prices of 30-day Fed Funds futures contracts. The underlying intuition is that the futures price reflects market expectations of the average EFFR over the delivery month and, by comparing this to the midpoints of possible target rate ranges, one can back out the implied probability assigned to each outcome.
For each meeting and each day, the result is a discrete probability distribution over a set of target rate buckets. In all cases, the tail buckets represent cumulative probability, while the interior buckets represent point probabilities for specific 25bps target ranges. The full distribution sums to 100% by construction.
4.3 Polymarket-Implied Probabilities
The second series is sourced from Polymarket via its API. For each FOMC meeting, Polymarket creates a dedicated market, with the possibility to trade different outcomes corresponding to possible rate decisions. As in the CME distribution, the tail outcomes represent cumulative probabilities of having a cut or hike equal or higher than stated. Across all markets, decrease_25 and no_change represent point outcomes, making them less noisy and therefore the primary focus of the analysis.
Polymarket defines the price of each outcome as the midpoint between the best bid and ask. In our sample, quoted bid-ask spreads are very small and typically around 0.1 percentage points, so midpoint prices should provide a close approximation of market-implied probabilities. As a result, the midpoint price provides a close approximation to the true market-implied probability, and the spread is not expected to introduce any systematic bias in our statistical testing. Prices are expressed as percentages and, across all outcome tokens for a given contract, sum to 100 after normalisation.
4.4 Dataset Overview
Table 1 summarises the five datasets used in the analysis. For each meeting, we report the sample period, the number of daily observations, and the set of CME buckets and Polymarket outcomes available.
The analysis relies on two daily time series for each FOMC meeting under consideration: implied probabilities derived from CME 30-day Fed Funds futures, and market prices from Polymarket prediction markets. Both series are observed at end-of-day frequency, and each dataset covers the period from the opening of the corresponding Polymarket contract to the day of the FOMC decision. In total, five meetings are examined: December 2025, January 2026 and March 2026, which have resolved, and April 2026 and June 2026, which were still unresolved at the time of writing.
4.2 CME-Implied Probabilities
The first series is constructed following the CME FedWatch methodology, which derives the market-implied probability distribution over possible FOMC policy outcomes from the settlement prices of 30-day Fed Funds futures contracts. The underlying intuition is that the futures price reflects market expectations of the average EFFR over the delivery month and, by comparing this to the midpoints of possible target rate ranges, one can back out the implied probability assigned to each outcome.
For each meeting and each day, the result is a discrete probability distribution over a set of target rate buckets. In all cases, the tail buckets represent cumulative probability, while the interior buckets represent point probabilities for specific 25bps target ranges. The full distribution sums to 100% by construction.
4.3 Polymarket-Implied Probabilities
The second series is sourced from Polymarket via its API. For each FOMC meeting, Polymarket creates a dedicated market, with the possibility to trade different outcomes corresponding to possible rate decisions. As in the CME distribution, the tail outcomes represent cumulative probabilities of having a cut or hike equal or higher than stated. Across all markets, decrease_25 and no_change represent point outcomes, making them less noisy and therefore the primary focus of the analysis.
Polymarket defines the price of each outcome as the midpoint between the best bid and ask. In our sample, quoted bid-ask spreads are very small and typically around 0.1 percentage points, so midpoint prices should provide a close approximation of market-implied probabilities. As a result, the midpoint price provides a close approximation to the true market-implied probability, and the spread is not expected to introduce any systematic bias in our statistical testing. Prices are expressed as percentages and, across all outcome tokens for a given contract, sum to 100 after normalisation.
4.4 Dataset Overview
Table 1 summarises the five datasets used in the analysis. For each meeting, we report the sample period, the number of daily observations, and the set of CME buckets and Polymarket outcomes available.
Table 1: Dataset Summary
| Meeting | Sample Period | Obs. | CME Buckets | Polymarket Outcomes |
|---|---|---|---|---|
| Dec 2025 | Aug 6, 2025 – Dec 10, 2025 | 127 | (0–350), (350–375), (375–400), (400+) |
dec50, dec25, nc, inc25 |
| Jan 2026 | Sep 23, 2025 – Jan 28, 2026 | 128 | (0–325), (325–350), (350–375), (375+) |
dec50, dec25, nc, inc25 |
| Mar 2026 | Nov 4, 2025 – Mar 18, 2026 | 134 | (0–325), (325–350), (350–375), (375+) |
dec50, dec25, nc, inc25 |
| Apr 2026 | Nov 19, 2025 – Apr 11, 2026 | 122 | (0–325), (325–350), (350–375), (375+) |
dec50, dec25, nc, inc25 |
| Jun 2026 | Dec 16, 2025 – Apr 11, 2026 | 99 | (0–325), (325–350), (350–375), (375–400), (400+) |
dec50, dec25, nc, inc25, inc50 |
dec50 = decrease ≥50bps, dec25 = decrease 25bps, nc = no change, inc25 = increase ≥25bps (exception for June 2026 with inc25 = increase > 25bp, inc50 = increase ≥50bps)
4.5 Series Selection for Statistical Testing
As noted above, our primary focus for the lead-lag analysis will be the decrease_25 and no_change outcomes, with the additional inclusion of increase_25 for June 2026, since that market also includes increase_50 as a cumulative tail outcome. This choice is motivated by the fact that point probabilities are available on both the CME and Polymarket sides, meaning that these series represent the market-assigned probability of a specific, well-defined outcome.
The full set of outcomes will be subject to stationarity and cointegration testing for completeness. The lead-lag inference and strategy implementation will centre on these two point-probability series.
As noted above, our primary focus for the lead-lag analysis will be the decrease_25 and no_change outcomes, with the additional inclusion of increase_25 for June 2026, since that market also includes increase_50 as a cumulative tail outcome. This choice is motivated by the fact that point probabilities are available on both the CME and Polymarket sides, meaning that these series represent the market-assigned probability of a specific, well-defined outcome.
The full set of outcomes will be subject to stationarity and cointegration testing for completeness. The lead-lag inference and strategy implementation will centre on these two point-probability series.
5. Methodology
5.1 Overview
The aim of this section is to establish whether CME-implied probabilities systematically lead Polymarket prices, providing a statistical foundation for the trading strategy. The analysis proceeds in three stages. First, we assess the stationarity properties of each series. Second, we test for cointegration between each CME-Polymarket pair. Third, depending on the cointegration outcome, we apply either an Error Correction Model (ECM) or a Granger causality test to identify the direction of the lead-lag relationship.
As explained above, within each meeting we primarily focus on the pairs corresponding to decrease_25 and no_change, and additionally increase_25 for the June 2026 meeting.
5.2 Stationarity Analysis
We begin by testing each series individually for the presence of a unit root, using two complementary tests: the Augmented Dickey-Fuller (ADF) test and the Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test.
The ADF test evaluates the null hypothesis that the series contains a unit root and is therefore non-stationary. A rejection at the 5% significance level is taken as evidence of stationarity. The KPSS test takes the opposite stance, with the null hypothesis of level stationarity, so that a rejection indicates non-stationarity. Running both tests jointly allows for a more robust classification than either test alone, since each is subject to different size and power limitations.
Table 2 summarises the four possible verdicts of the two tests.
The aim of this section is to establish whether CME-implied probabilities systematically lead Polymarket prices, providing a statistical foundation for the trading strategy. The analysis proceeds in three stages. First, we assess the stationarity properties of each series. Second, we test for cointegration between each CME-Polymarket pair. Third, depending on the cointegration outcome, we apply either an Error Correction Model (ECM) or a Granger causality test to identify the direction of the lead-lag relationship.
As explained above, within each meeting we primarily focus on the pairs corresponding to decrease_25 and no_change, and additionally increase_25 for the June 2026 meeting.
5.2 Stationarity Analysis
We begin by testing each series individually for the presence of a unit root, using two complementary tests: the Augmented Dickey-Fuller (ADF) test and the Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test.
The ADF test evaluates the null hypothesis that the series contains a unit root and is therefore non-stationary. A rejection at the 5% significance level is taken as evidence of stationarity. The KPSS test takes the opposite stance, with the null hypothesis of level stationarity, so that a rejection indicates non-stationarity. Running both tests jointly allows for a more robust classification than either test alone, since each is subject to different size and power limitations.
Table 2 summarises the four possible verdicts of the two tests.
Table 2: ADF and KPSS verdicts
| ADF rejects | ADF does not reject | |
| KPSS rejects | Inconclusive | Not stationary (I(1)) |
| KPSS does not reject | Stationary (I(0)) | Inconclusive |
For the ADF test, lag length is selected automatically via the Akaike Information Criterion (AIC). For the KPSS test, the number of lags is selected using the automatic bandwidth selection procedure. Both tests are run under the constant (level) specification.
5.3 Cointegration Analysis
For pairs where both series are classified as I(1), we test for cointegration using the Engle-Granger two-step procedure. Cointegration implies that, despite each series being individually non-stationary, a linear combination of the two is stationary, capturing a long-run equilibrium relationship between CME and Polymarket implied probabilities.
The Engle-Granger test is well-suited to our bivariate setting. In the first step, we estimate the long-run OLS regression:
Polyt = α + β · CMEt + εt
where εt = residual term
We are testing whether the residuals are stationary. A stationary residual series indicates that the two markets share a long-run equilibrium. Since the test is not symmetric with respect to the choice of dependent variable, we run it in both directions and report both p-values.
To control for the inflated false positive rate arising from testing both directions, we apply a Bonferroni correction, setting the significance threshold at 0.025 rather than 0.05. A pair is declared cointegrated if the minimum of the two p-values falls below this corrected threshold, ensuring significance in at least one direction after adjusting for multiple testing.
5.4 Lead-Lag Analysis
Depending on the outcome of the cointegration test, we apply one of two methods to identify which market leads the other.
5.4.1 Error Correction Model
For cointegrated pairs, we estimate an Error Correction Model (ECM) following the Engle-Granger two-step approach. Having obtained the equilibrium error from the long-run regression in Step 1, the ECM is estimated in Step 2 as:
To control for the inflated false positive rate arising from testing both directions, we apply a Bonferroni correction, setting the significance threshold at 0.025 rather than 0.05. A pair is declared cointegrated if the minimum of the two p-values falls below this corrected threshold, ensuring significance in at least one direction after adjusting for multiple testing.
5.4 Lead-Lag Analysis
Depending on the outcome of the cointegration test, we apply one of two methods to identify which market leads the other.
5.4.1 Error Correction Model
For cointegrated pairs, we estimate an Error Correction Model (ECM) following the Engle-Granger two-step approach. Having obtained the equilibrium error from the long-run regression in Step 1, the ECM is estimated in Step 2 as:
ΔPolyt = αPoly · εt−1 +
∑
k
i=1
γiPoly ΔCMEt−i +
∑
k
i=1
δiPoly ΔPolyt−i + ut
ΔCMEt = αCME · εt−1 + ∑ k i=1 γiCME ΔCMEt−i + ∑ k i=1 δiCME ΔPolyt−i + vt
ΔCMEt = αCME · εt−1 + ∑ k i=1 γiCME ΔCMEt−i + ∑ k i=1 δiCME ΔPolyt−i + vt
The key parameters are the adjustment coefficient αpoly and αCME. A significant negative α indicates that the corresponding market error-corrects towards the long-run equilibrium, that is, it is the follower. Conversely, an insignificant α indicates that the market does not adjust, identifying it as the leader. Our hypothesis predicts that αpoly will be negative and significant while αCME will be insignificant, consistent with CME being the price discovery venue.
The number of lagged differences k is selected separately for each pair by minimising the AIC over the Polymarket equation, with a maximum of five lags. A verdict of "CME leads" is assigned when αpoly is negative and significant at the 5% level and αCME is not. A verdict of "Poly leads" is assigned in the reverse case. If both coefficients are significant and negative, the verdict is "Biderictional". If neither is significant, the result is "Inconclusive".
5.4.2 Granger Causality
For pairs that are not cointegrated, either because one series is stationary, I(0), or because the Engle-Granger test fails to reject the null of no cointegration, we test for Granger causality. Where necessary, the data are transformed to ensure stationarity prior to testing.
A variable X is said to Granger-cause Y if lagged values of X contain statistically significant predictive power for Y beyond what is already captured by lagged values of Y alone. We test both directions: whether lagged changes in CME-implied probabilities (ΔCME) help predict changes in Polymarket probabilities (ΔPoly), and whether lagged ΔPoly helps predict ΔCME. The optimal lag length is selected via AIC on the Polymarket equation with a maximum of four lags. As with the cointegration step, we apply a Bonferroni correction to account for multiple testing across directions, setting the significance threshold at 0.05/2 = 0.025. The F-test p-value at the AIC-selected lag is used for inference.
The same four-way verdict structure applies: "CME leads", "Poly leads", "Bidirectional", or "No lead".
5.5 Strategy
The backtest implements a mean-reversion pyramid strategy on the spread between CME-implied and Polymarket-implied probabilities for each FOMC outcome. For every meeting–outcome pair, it computes spread = CME − Polymarket and stays flat until the spread breaches ±2pp: a positive breach triggers a long on Polymarket (CME is pricing the outcome higher, so we expect Polymarket to converge upward), a negative breach triggers a short. While in position the strategy pyramids, adding one unit each time the spread moves a further 2pp in the entry direction relative to the last add, and exits the full stack once the spread reverts by 2pp from the first entry level (with a forced close at the end of the sample if still open). PnL is then booked as the change in the Polymarket probability between entry and exit, in percentage points, scaled by the number of units held. The resulting figure therefore measures how much of the Polymarket series move the strategy captures, rather than the dollar return one would realise actually trading Polymarket shares, which would additionally depend on the share price level at entry/exit, fees, slippage and the binary payoff mechanics.
The number of lagged differences k is selected separately for each pair by minimising the AIC over the Polymarket equation, with a maximum of five lags. A verdict of "CME leads" is assigned when αpoly is negative and significant at the 5% level and αCME is not. A verdict of "Poly leads" is assigned in the reverse case. If both coefficients are significant and negative, the verdict is "Biderictional". If neither is significant, the result is "Inconclusive".
5.4.2 Granger Causality
For pairs that are not cointegrated, either because one series is stationary, I(0), or because the Engle-Granger test fails to reject the null of no cointegration, we test for Granger causality. Where necessary, the data are transformed to ensure stationarity prior to testing.
A variable X is said to Granger-cause Y if lagged values of X contain statistically significant predictive power for Y beyond what is already captured by lagged values of Y alone. We test both directions: whether lagged changes in CME-implied probabilities (ΔCME) help predict changes in Polymarket probabilities (ΔPoly), and whether lagged ΔPoly helps predict ΔCME. The optimal lag length is selected via AIC on the Polymarket equation with a maximum of four lags. As with the cointegration step, we apply a Bonferroni correction to account for multiple testing across directions, setting the significance threshold at 0.05/2 = 0.025. The F-test p-value at the AIC-selected lag is used for inference.
The same four-way verdict structure applies: "CME leads", "Poly leads", "Bidirectional", or "No lead".
5.5 Strategy
The backtest implements a mean-reversion pyramid strategy on the spread between CME-implied and Polymarket-implied probabilities for each FOMC outcome. For every meeting–outcome pair, it computes spread = CME − Polymarket and stays flat until the spread breaches ±2pp: a positive breach triggers a long on Polymarket (CME is pricing the outcome higher, so we expect Polymarket to converge upward), a negative breach triggers a short. While in position the strategy pyramids, adding one unit each time the spread moves a further 2pp in the entry direction relative to the last add, and exits the full stack once the spread reverts by 2pp from the first entry level (with a forced close at the end of the sample if still open). PnL is then booked as the change in the Polymarket probability between entry and exit, in percentage points, scaled by the number of units held. The resulting figure therefore measures how much of the Polymarket series move the strategy captures, rather than the dollar return one would realise actually trading Polymarket shares, which would additionally depend on the share price level at entry/exit, fees, slippage and the binary payoff mechanics.
6. Results and Robustness
6.1 Statistical Results
6.1.1 Lead-Lag
The lead-lag analysis produces consistent findings across all five FOMC meetings and both methodological approaches: where a statistical relationship is detectable, CME-implied probabilities consistently lead Polymarket prices. We find no evidence of the converse. This result holds across different outcome pairs and frameworks, which lends the result considerable robustness given the size of the sample.
6.1.2 Time Series Properties (Stationarity)
The stationarity analysis indicates that the vast majority of series are integrated of order one, I(1). This is the expected result for probability series that begin away from their terminal value and drift towards it as the meeting date nears. However, the few exceptions – notably the December 2025 (0–350), January 2026 (0–325) and April 2026 (375+) CME buckets, which tested as stationary – reflect series that had already converged to near-certainty by the start of their sample windows, leaving little variation to analyse. A small number of additional series, concentrated in March 2026 and June 2026, returned inconclusive verdicts where the ADF and KPSS tests disagree; these are treated as non-stationary for the purposes of subsequent testing. For the focus pairs, decrease_25 and no_change, the I(1) classification is consistent across all meetings, providing a clean and consistent foundation for cointegration testing.
6.1.3 Cointegration and Error Correction
Out of the eleven CME-Polymarket pairs tested using the Engle-Granger procedure, the following four are found to be cointegrated: the January 2026 (350–375) vs no_change pair, both March 2026 pairs and the April 2026 (325–350) vs decrease_25 pair. For each of these, the ECM produces a clear “CME leads” verdict. The Polymarket adjustment coefficient αpoly is negative and significant in all cases – ranging from -0.142 (April 2026) to -0.243 (January 2026) – while the CME adjustment coefficient αCME is insignificant. This pattern is consistent with the theoretical expectation: Polymarket error-corrects itself towards the CME equilibrium, while CME itself does not adjust.
The consistency of this result across all four independent meeting-pair combinations, estimated with maximum five lags each, reduces the likelihood that it is driven by any single outlier.
6.1.4 Granger Causality
For the remaining seven pairs, where cointegration could not be established, Granger causality tests are applied to first differences. Five of these seven return a “CME leads” verdict, with p-values for the CME-to-Polymarket direction close to zero and p-values on the converse direction uniformly above 0.10. The two exceptions – April 2026 (350–375) vs no_change and June 2026 (375–400) vs increase_25 – return a “No lead” verdict, indicating that neither series contains significant predictive power over the other at the tested lag length. It is important to note that neither of these exceptions constitutes evidence of Polymarket leading; they reflect an absence of detectable directionality.
6.1.5 Summary of Empirical Findings
Taken together, nine of the chosen eleven pairs produce a “CME leads” verdict, with the remaining two inconclusive and none favouring Polymarket. This uniformity across meetings, outcome types and estimation methods constitutes the central empirical finding of the analysis.
6.2 Trading Strategy Results
Below, we present four of the eleven outcome pairs: three profitable cases and the single loss-making example, showing both the spread dynamics and cumulative PnL.
6.1.1 Lead-Lag
The lead-lag analysis produces consistent findings across all five FOMC meetings and both methodological approaches: where a statistical relationship is detectable, CME-implied probabilities consistently lead Polymarket prices. We find no evidence of the converse. This result holds across different outcome pairs and frameworks, which lends the result considerable robustness given the size of the sample.
6.1.2 Time Series Properties (Stationarity)
The stationarity analysis indicates that the vast majority of series are integrated of order one, I(1). This is the expected result for probability series that begin away from their terminal value and drift towards it as the meeting date nears. However, the few exceptions – notably the December 2025 (0–350), January 2026 (0–325) and April 2026 (375+) CME buckets, which tested as stationary – reflect series that had already converged to near-certainty by the start of their sample windows, leaving little variation to analyse. A small number of additional series, concentrated in March 2026 and June 2026, returned inconclusive verdicts where the ADF and KPSS tests disagree; these are treated as non-stationary for the purposes of subsequent testing. For the focus pairs, decrease_25 and no_change, the I(1) classification is consistent across all meetings, providing a clean and consistent foundation for cointegration testing.
6.1.3 Cointegration and Error Correction
Out of the eleven CME-Polymarket pairs tested using the Engle-Granger procedure, the following four are found to be cointegrated: the January 2026 (350–375) vs no_change pair, both March 2026 pairs and the April 2026 (325–350) vs decrease_25 pair. For each of these, the ECM produces a clear “CME leads” verdict. The Polymarket adjustment coefficient αpoly is negative and significant in all cases – ranging from -0.142 (April 2026) to -0.243 (January 2026) – while the CME adjustment coefficient αCME is insignificant. This pattern is consistent with the theoretical expectation: Polymarket error-corrects itself towards the CME equilibrium, while CME itself does not adjust.
The consistency of this result across all four independent meeting-pair combinations, estimated with maximum five lags each, reduces the likelihood that it is driven by any single outlier.
6.1.4 Granger Causality
For the remaining seven pairs, where cointegration could not be established, Granger causality tests are applied to first differences. Five of these seven return a “CME leads” verdict, with p-values for the CME-to-Polymarket direction close to zero and p-values on the converse direction uniformly above 0.10. The two exceptions – April 2026 (350–375) vs no_change and June 2026 (375–400) vs increase_25 – return a “No lead” verdict, indicating that neither series contains significant predictive power over the other at the tested lag length. It is important to note that neither of these exceptions constitutes evidence of Polymarket leading; they reflect an absence of detectable directionality.
6.1.5 Summary of Empirical Findings
Taken together, nine of the chosen eleven pairs produce a “CME leads” verdict, with the remaining two inconclusive and none favouring Polymarket. This uniformity across meetings, outcome types and estimation methods constitutes the central empirical finding of the analysis.
6.2 Trading Strategy Results
Below, we present four of the eleven outcome pairs: three profitable cases and the single loss-making example, showing both the spread dynamics and cumulative PnL.
The strategy is evaluated at a threshold of 2pp, meaning a trade is entered when the gap between CME-implied probabilities and Polymarket-implied probabilities exceeds 2pp, and exited when it narrows back by the same magnitude. PnL is reported in cumulative percentage points of probability – that is, the sum of probability-point gains and losses across all trades for each meeting, rather than a realised monetary return.
Across the five meetings and eleven outcome pairs, the strategy produces positive cumulative PnL in ten out of eleven cases. The strongest cumulative gains are observed for December 2025, where the decrease_25 and no_change pairs yield 275.9pp and 243.9pp respectively, with hit rates of 81.2% and 88.9% across 16 and 18 trades. Two of the June 2026 pairs also perform relatively strongly, with decrease_25 and no_change generating 175.9pp and 181.3pp, with hit rates of 94.1% and 82.6%, respectively.
Performance is more modest in the intermediate meetings. January 2026 and March 2026 both produce positive but smaller cumulative gains, ranging from 90.2pp to 133.4pp, with hit rates between 58.8% and 83.3%. The weakest results occur for April 2026 no_change, which delivers a near break-even result of 5.1pp despite a 72.2% hit rate, and June 2026 increase_25, which is the only loss-making pair at -0.7pp across just four trades, making it too small a sample to draw strong conclusions.
Overall, hit rates are consistently above 70% across most specifications, with January 2026 decrease_25 as the main exception at 58.8%. This pattern is consistent with the earlier lead-lag results: where CME systematically leads Polymarket, deviations that open in the direction implied by CME tend to close as Polymarket adjusts, generating a high proportion of profitable trades.
Since the analysis is conducted at a single threshold of 2pp, comparing performance across alternative thresholds such as 1pp and 3pp would allow for an assessment of threshold sensitivity, and remains a potential direction for future refinement.
6.3 Limitations and Robustness
While the lead-lag test findings are consistent, several limitations affect how the results and derived trading strategy should be interpreted. These fall into two broad categories: data and measurement limitations and strategy-level concerns.
The first concerns the asymmetric observation frequency between the two series. CME-implied probabilities are recorded once daily, at the end of the day, while Polymarket prices are continuously observed throughout the trading day. The strategy used exploits this directly by using a CME signal at time t to predict Polymarket at time t+1, which makes the setup both transparent and operationally implementable. However, this asymmetry would mean that any intraday Polymarket movement occurring after the CME close is not captured until the following day’s CME reading. During especially high-activity time windows such as FOMC press conferences or macroeconomic data releases, this gap could appear wider than if both series were observed on the same frequency. Therefore, the lead-lag relationships should be understood as operating at a frequency of hours to days, as opposed to mere minutes.
Another related concern is stale quoting on Polymarket. Since Polymarket is a thinner, retail-oriented venue, prices on lower-volume bets can remain unchanged for extended periods of time; not because traders agree with the prevailing price, but because no trade has occurred. Snowberg, Wolfers and Zitzewitz (2012) identify thin markets as a recognised structural weakness of such prediction markets, noting that a lack of active participation reduces active members’ incentive to discover and trade on new information. Where this applies, the Polymarket series may reflect the preceding transaction rather than a current market-clearing price, which would cause the CME-Polymarket gap to appear larger and more persistent than it truly is.
At the strategy level, a concern is that transaction costs are not accounted for in the raw backtest. Polymarket charges a fixed fee on each trade, and the effective bid-ask spread on lower-liquidity contracts can be materially significant relative to the gaps being exploited. A gap of one or two percentage points may be partially or entirely consumed by such costs, meaning that only the wider thresholds are likely to survive the frictions. Additionally, position size is a further limitation: some of Polymarket’s order books can be characterised by low liquidity, meaning that attempting to trade a meaningful size at the quoted price may move the market against the position before it is filled.
Two additional limitations affect the interpretation of the backtest specifically. First, the strategy is tested over a limited number of FOMC meetings, which limits statistical power and raises the possibility of overfitting – the chosen gap thresholds may perform well over this sample without generalising to future meetings. There is also a degree of event-definition mismatch between the two. CME-implied probabilities are derived from Fed Funds futures, which price the average Federal Funds rate across an entire calendar month rather than the outcome of a single meeting. Polymarket’s contracts, by contrast, are binary and tied directly to a specific meeting outcome. Snowberg, Wolfers and Zitzewitz (2012) note that prediction market contracts must be precisely specified, and that ambiguity in contract definition is a recognised source of noise in cross-market comparisons. Some portion of the observed gap between the two series may therefore reflect this structural difference in contract design, as opposed to exploitable mispricing.
Across the five meetings and eleven outcome pairs, the strategy produces positive cumulative PnL in ten out of eleven cases. The strongest cumulative gains are observed for December 2025, where the decrease_25 and no_change pairs yield 275.9pp and 243.9pp respectively, with hit rates of 81.2% and 88.9% across 16 and 18 trades. Two of the June 2026 pairs also perform relatively strongly, with decrease_25 and no_change generating 175.9pp and 181.3pp, with hit rates of 94.1% and 82.6%, respectively.
Performance is more modest in the intermediate meetings. January 2026 and March 2026 both produce positive but smaller cumulative gains, ranging from 90.2pp to 133.4pp, with hit rates between 58.8% and 83.3%. The weakest results occur for April 2026 no_change, which delivers a near break-even result of 5.1pp despite a 72.2% hit rate, and June 2026 increase_25, which is the only loss-making pair at -0.7pp across just four trades, making it too small a sample to draw strong conclusions.
Overall, hit rates are consistently above 70% across most specifications, with January 2026 decrease_25 as the main exception at 58.8%. This pattern is consistent with the earlier lead-lag results: where CME systematically leads Polymarket, deviations that open in the direction implied by CME tend to close as Polymarket adjusts, generating a high proportion of profitable trades.
Since the analysis is conducted at a single threshold of 2pp, comparing performance across alternative thresholds such as 1pp and 3pp would allow for an assessment of threshold sensitivity, and remains a potential direction for future refinement.
6.3 Limitations and Robustness
While the lead-lag test findings are consistent, several limitations affect how the results and derived trading strategy should be interpreted. These fall into two broad categories: data and measurement limitations and strategy-level concerns.
The first concerns the asymmetric observation frequency between the two series. CME-implied probabilities are recorded once daily, at the end of the day, while Polymarket prices are continuously observed throughout the trading day. The strategy used exploits this directly by using a CME signal at time t to predict Polymarket at time t+1, which makes the setup both transparent and operationally implementable. However, this asymmetry would mean that any intraday Polymarket movement occurring after the CME close is not captured until the following day’s CME reading. During especially high-activity time windows such as FOMC press conferences or macroeconomic data releases, this gap could appear wider than if both series were observed on the same frequency. Therefore, the lead-lag relationships should be understood as operating at a frequency of hours to days, as opposed to mere minutes.
Another related concern is stale quoting on Polymarket. Since Polymarket is a thinner, retail-oriented venue, prices on lower-volume bets can remain unchanged for extended periods of time; not because traders agree with the prevailing price, but because no trade has occurred. Snowberg, Wolfers and Zitzewitz (2012) identify thin markets as a recognised structural weakness of such prediction markets, noting that a lack of active participation reduces active members’ incentive to discover and trade on new information. Where this applies, the Polymarket series may reflect the preceding transaction rather than a current market-clearing price, which would cause the CME-Polymarket gap to appear larger and more persistent than it truly is.
At the strategy level, a concern is that transaction costs are not accounted for in the raw backtest. Polymarket charges a fixed fee on each trade, and the effective bid-ask spread on lower-liquidity contracts can be materially significant relative to the gaps being exploited. A gap of one or two percentage points may be partially or entirely consumed by such costs, meaning that only the wider thresholds are likely to survive the frictions. Additionally, position size is a further limitation: some of Polymarket’s order books can be characterised by low liquidity, meaning that attempting to trade a meaningful size at the quoted price may move the market against the position before it is filled.
Two additional limitations affect the interpretation of the backtest specifically. First, the strategy is tested over a limited number of FOMC meetings, which limits statistical power and raises the possibility of overfitting – the chosen gap thresholds may perform well over this sample without generalising to future meetings. There is also a degree of event-definition mismatch between the two. CME-implied probabilities are derived from Fed Funds futures, which price the average Federal Funds rate across an entire calendar month rather than the outcome of a single meeting. Polymarket’s contracts, by contrast, are binary and tied directly to a specific meeting outcome. Snowberg, Wolfers and Zitzewitz (2012) note that prediction market contracts must be precisely specified, and that ambiguity in contract definition is a recognised source of noise in cross-market comparisons. Some portion of the observed gap between the two series may therefore reflect this structural difference in contract design, as opposed to exploitable mispricing.
7. Conclusion
In conclusion, the analysis provides consistent evidence that CME-implied probabilities lead Polymarket prices in the context of FOMC rate decisions. Across five meetings, eleven CME-Polymarket pairs and two distinct frameworks, the direction of price discovery runs consistently from CME to Polymarket in the cases where a directional relationship is detected. Where cointegration was detected, Polymarket error-corrects itself towards the CME equilibrium; in cases where it is not, Granger causality tests confirm the same relationship. We find no evidence of the converse.
At the strategy level, a convergence trade entered at a 2pp threshold generates positive cumulative PnL in ten of the eleven pairs across the five meetings, with hit rates consistently above 70%. This provides a practical counterpart to the statistical findings – the lead-lag relationship is not only detectable but also appears tradable in raw terms. However, transaction costs and liquidity constraints could materially affect real-world returns.
The findings are consistent with what differences in participant bases and market structures would predict. CME is an institutional market, populated by professional fixed-income traders with strong incentives to process macroeconomic information quickly and accurately. On the other hand, Polymarket attracts a retail-oriented participant base for whom FOMC outcomes may be one of many event-driven trades. The lead-lag relationship identified here likely reflects this mismatch in information processing speeds, rather than a persistent arbitrage opportunity – as suggested by Oliven and Rietz (2004).
Several directions for future research follow from these findings. A larger sample set spanning more FOMC meetings would allow for more reliable threshold calculations, as well as reduce the overfitting concern raised in Section 5.2. Additionally, extending this comparison to other central bank decisions – such as ECB or Bank of England meetings, where similar prediction market contracts exist – would test whether the CME-leads finding is specific to the Federal Reserve or reflects a more general relationship between institutional and retail-oriented venues.
At the strategy level, a convergence trade entered at a 2pp threshold generates positive cumulative PnL in ten of the eleven pairs across the five meetings, with hit rates consistently above 70%. This provides a practical counterpart to the statistical findings – the lead-lag relationship is not only detectable but also appears tradable in raw terms. However, transaction costs and liquidity constraints could materially affect real-world returns.
The findings are consistent with what differences in participant bases and market structures would predict. CME is an institutional market, populated by professional fixed-income traders with strong incentives to process macroeconomic information quickly and accurately. On the other hand, Polymarket attracts a retail-oriented participant base for whom FOMC outcomes may be one of many event-driven trades. The lead-lag relationship identified here likely reflects this mismatch in information processing speeds, rather than a persistent arbitrage opportunity – as suggested by Oliven and Rietz (2004).
Several directions for future research follow from these findings. A larger sample set spanning more FOMC meetings would allow for more reliable threshold calculations, as well as reduce the overfitting concern raised in Section 5.2. Additionally, extending this comparison to other central bank decisions – such as ECB or Bank of England meetings, where similar prediction market contracts exist – would test whether the CME-leads finding is specific to the Federal Reserve or reflects a more general relationship between institutional and retail-oriented venues.
Written by: Fabio Visconti, Marc Weigand, Aaryan Kandala, Audrey Victoria Venturi, Pietro Ferraro