Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
71,901 characters · 19 sections · 8 citation commands
The Shape of Macroeconomic Beliefs
Expectations matter for monetary policy because decisions depend not only on the most likely future outcome, but also on the risks around that outcome. The distinction is especially important for inflation. A central bank faces a different communication problem when the expected CPI release is stable but the probability of a high-inflation outcome has risen. Yet many empirical measures of macroeconomic expectations are incomplete along this dimension. Surveys are direct, but are often low frequency or summarized by point forecasts. Survey densities are valuable but usually refer to lower-frequency horizons rather than to the probability distribution of a specific statistical release. Asset prices are high frequency, but extracting beliefs from them requires assumptions about risk premia, liquidity, and the mapping from asset payoffs to macroeconomic states. The object that is often most useful for monetary monitoring is a release-level subjective distribution; the econometrician usually observes only a mean or a model-dependent signal. The analysis uses prediction markets to study the distribution of short-run macroeconomic beliefs. A panel of market-implied distributions is constructed from macroeconomic event contracts traded on Kalshi. The key empirical feature is that many contracts are written on adjacent thresholds. For example, separate contracts may pay depending on whether monthly CPI inflation exceeds 0.2, 0.3, or 0.4 percent. Observing prices across these thresholds before the release makes it possible to recover a discrete distribution over the upcoming macroeconomic outcome. The resulting data reveal the market-implied mean, dispersion, skewness, entropy, and tail probabilities at standardized horizons from 30 days to one hour before resolution. The central question is whether recent macroeconomic news affects the location of prediction-market beliefs or their shape. A mean-based analysis asks whether a positive inflation surprise shifts the next expected inflation release upward. A distributional analysis asks a broader question: does the market reallocate probability mass toward high-inflation states, or become more uncertain, even when the mean remains close to consensus? This distinction matters for monetary economics. Prediction markets may be useful not because they always dominate professional forecasts in predicting the modal outcome, but because they reveal how markets price the tails of the next release.
The analysis focuses on CPI and core CPI month-over-month releases, which provide the cleanest repeated setting. The baseline distributional sample contains 282 CPI and 240 core CPI snapshots across 46 and 39 releases, respectively. These distributions are merged with Reuters Poll consensus forecasts. This allows the empirical analysis to separate three objects: the realized release, the professional consensus forecast, and the contemporaneous prediction-market distribution. The Reuters consensus is not treated as an exogenous object. It is a conditioning benchmark for the professional point forecast. The reduced-form question is whether prediction-market distributions contain additional information about residual inflation risk, conditional on that benchmark and on professional-forecaster disagreement. The first result is that Kalshi-implied means contain meaningful forecast information. For headline CPI, the Kalshi mean has an overall RMSE of 0.169 percentage points, compared with 0.331 for the previous realization, 0.280 for a three-month moving average, and 0.274 for a six-month moving average. For core CPI, Kalshi is broadly competitive with simple moving-average benchmarks, although the improvement is smaller. This forecast content is important because it shows that the recovered distributions are not noise. Forecast accuracy, however, is not the main contribution. The second result is that recent inflation news is not primarily reflected in disagreement about the center of the next release distribution. With variable-by-horizon fixed effects, the coefficient of the mean gap \(\mu^K-C\) on the lagged Reuters surprise is \(-0.023\), with a clustered standard error of 0.107. Thus, past Reuters surprises do not lead Kalshi means to deviate systematically from the current Reuters consensus. This null result is informative: the market's distinctive signal is not a persistent shift in the conditional mean relative to professional forecasters. The third and main result is distributional. Large previous surprises are associated with greater market-implied uncertainty. Positive previous surprises raise the probability assigned to fixed high-inflation outcomes. With variable-by-horizon fixed effects and the current Reuters consensus controlled for, the coefficient on the lagged Reuters surprise is 0.466 for \(\Pr^K(y>0.3)\), 0.422 for \(\Pr^K(y>0.4)\), and 0.281 for \(\Pr^K(y>0.5)\). These magnitudes imply that a 0.1 percentage point positive lagged surprise raises the probability of inflation above 0.3 percent by about 4.7 percentage points. The response is weaker when tails are defined relative to the contemporaneous consensus forecast. The interpretation is therefore not that Kalshi mechanically disagrees with Reuters after a surprise. Instead, recent inflation news changes the absolute inflation-risk environment: the mean remains close to consensus, while the market assigns more mass to high-inflation states.
Several validation and robustness exercises support this interpretation. Release-level tests show that Kalshi upper-tail probabilities predict the realization of high-inflation states, including episodes in which the market-implied mean remains close to the Reuters consensus. Staleness filters from 24 hours to one hour do not attenuate the uncertainty and fixed-tail coefficients. Direct controls for snapshot staleness, open interest, and volume leave the central estimates positive. Alternative tail-support rules and uniform-within-bin interpolation deliver the same qualitative fixed-tail result. Wild cluster bootstrap inference by release is more conservative, but the main uncertainty and fixed-tail estimates remain statistically meaningful. Split-sample regressions for CPI and core CPI preserve the signs, although precision declines. Leave-one-release-out diagnostics show that the uncertainty and fixed-tail estimates are not driven by a single inflation release.
The paper contributes to four literatures. First, it contributes to work on expectation formation and information frictions. Rational-expectations and noisy-information models emphasize the role of information sets, attention, and dispersed signals in shaping beliefs Muth1961,Lucas1972,MankiwReis2002,Sims2003,Woodford2003,MorrisShin2002,Carroll2003,CoibionGorodnichenko2012,CoibionGorodnichenko2015. The contribution is financially incentivized, high-frequency evidence on the full distribution of short-run inflation beliefs. Second, it relates to behavioral models of extrapolation, diagnostic expectations, anchoring, and salience BarberisShleiferVishny1998,DanielHirshleiferSubrahmanyam1998,HongStein1999,BordaloGennaioliShleifer2012,BordaloGennaioliShleifer2018,BordaloGennaioliLaPortaShleifer2019. The evidence is consistent with recent inflation news affecting uncertainty and tail probabilities more than one-for-one mean beliefs. Third, the paper contributes to the measurement of macroeconomic expectations, including work using survey densities Croushore1993,EngelbergManskiWilliams2009,AndradeCrumpEusepiMoench2016. Prediction markets occupy an intermediate position between surveys and standard asset-price measures: their payoffs are directly tied to macroeconomic releases, but their prices are high frequency and financially backed. Fourth, the paper adds to the prediction-market literature, which studies information aggregation, calibration, risk preferences, and market microstructure ForsytheNelsonNeumannWright1992,BergNelsonRietz2008,WolfersZitzewitz2004,Manski2006,WolfersZitzewitz2006,ArrowEtAl2008,SnowbergWolfersZitzewitz2013. Recent platform-specific work studies Kalshi and Polymarket in macroeconomic, political, and microstructure applications DiercksKatzWright2026,SwansonWangWu2025,BurgiDengWhelan2026,EichengreenViswanathNatrajWangWang2025,TsangYang2026,TsangYang2026PoliticalShocks,Dubach2026,AkeyGregoireHarvieMartineau2026,GomezCramGuoJensenKung2026Accuracy,GomezCramGuoJensenKung2025Earnings,SaguilloGhafouriKifferSuarezTangil2025,DudleyMagdaleno2026,ClintonHuang2025. The analysis is complementary to that literature because it uses prediction-market prices as a measurement device for distributional inflation beliefs and studies how recent macroeconomic news affects the shape of those beliefs. The online appendix gives a detailed comparison with recent platform-specific papers.
The rest of the paper is organized as follows. Section (ref) describes the data and the construction of market-implied distributions. Section (ref) presents the empirical framework. Section (ref) reports the main results on forecast content, consensus alignment, uncertainty, and tail risk. Section (ref) reports robustness and diagnostic checks. Section (ref) discusses interpretation, limitations, and policy relevance. Section (ref) concludes.
This section describes the construction of the Kalshi distributional panel and the external forecast data used in the empirical analysis. The goal is to convert a cross-section of event-contract prices into a repeated panel of market-implied distributions for macroeconomic releases. The main analysis uses CPI and core CPI releases because these markets have the most regular contract structure and the closest mapping to Reuters Poll consensus forecasts.
Kalshi is a real-money prediction-market exchange on which traders buy and sell contracts linked to the realization of well-defined events. In the macroeconomic markets studied here, the event is tied to a future statistical release, such as monthly CPI inflation, core CPI inflation, nonfarm payrolls, the unemployment rate, PCE inflation, Treasury rates, or recession-related outcomes. Contracts have binary payoffs. A contract that pays one dollar if an event occurs can be interpreted as a market price for that event, subject to the usual caveats about fees, risk preferences, liquidity, bid-ask spreads, market power, and heterogeneous beliefs Manski2006,WolfersZitzewitz2004,WolfersZitzewitz2006,BurgiDengWhelan2026. The raw data contain contract identifiers, event identifiers, contract titles and subtitles, parsed threshold values, timestamps, prices, volume, and open interest when available. Prices are aggregated to an hourly frequency. Hourly observations preserve the high-frequency nature of the market while reducing the influence of isolated quote-level noise. For each event and hour, the cross-section of active contracts is retained and the implied distribution is recovered when there are enough contracts to identify at least three probability bins. Table (ref) summarizes the market coverage. Panel A reports the broad macroeconomic sample. The raw panel contains 476 macroeconomic events and more than 5,000 contracts, corresponding to more than 760,000 hourly contract-level observations. CPI and core CPI are the largest and cleanest repeated settings, with 100 events each. Panel B reports the baseline month-over-month inflation sample used in the main analysis. After imposing the distributional filters and standardized snapshot horizons, the sample contains 282 CPI and 240 core CPI distributional snapshots across 46 and 39 events. The median implied mean is about 0.25 percentage points month over month, and the median implied standard deviation is about 0.10 percentage points.
Let \(y_{g,r}\) denote the outcome for variable group \(g\) and release \(r\). For inflation, \(y_{g,r}\) is the month-over-month CPI or core CPI inflation rate. At horizon \(h\), the Kalshi price cross-section implies a distribution \[ F^K_{g,r,h}(y)=\Pr^K_{g,r,h}(y_{g,r}\leq y), \] where the superscript \(K\) denotes Kalshi. Many contracts in the sample are threshold contracts. If a contract pays one dollar when \(y_{g,r}>c_j\), its price is interpreted as an implied survival probability, \[ p_{g,r,h}(c_j)\approx \Pr^K_{g,r,h}(y_{g,r}>c_j). \] For ordered thresholds \(c_1<\cdots<c_J\), adjacent prices identify probability mass over bins. For interior bins, \[ \pi^K_{g,r,h,j}=\Pr^K_{g,r,h}(c_j<y_{g,r}\leq c_{j+1})=p_{g,r,h}(c_j)-p_{g,r,h}(c_{j+1}). \] The lower and upper open-ended bins are \[ \pi^K_{g,r,h,0}=1-p_{g,r,h}(c_1),\qquad \pi^K_{g,r,h,J}=p_{g,r,h}(c_J). \] When markets are quoted as bracket contracts rather than cumulative threshold contracts, the bracket prices are used directly as bin probabilities. In either case, small no-arbitrage violations can arise from stale quotes, bid-ask effects, or asynchronous trading. The baseline procedure truncates negative probability masses at zero and renormalizes the distribution to sum to one. The empirical analysis records the raw mass and the number of bins as diagnostics. Each bin receives a representative midpoint \(m_j\). Closed bins use the midpoint of the interval. Open-ended bins use a conservative midpoint based on the adjacent grid width. This choice affects the mean and standard deviation when substantial mass lies in the open tails. The online appendix therefore reports robustness exercises that vary tail supports and interpolate survival probabilities within bins. Given normalized bin probabilities \(\pi^K_{g,r,h,j}\), the market-implied mean is \[ \mu^K_{g,r,h}=\sum_j \pi^K_{g,r,h,j}m_j, \] and the standard deviation is \[ \sigma^K_{g,r,h}=\left[\sum_j \pi^K_{g,r,h,j}(m_j-\mu^K_{g,r,h})^2\right]^{1/2}. \] The recovered distribution also yields skewness, entropy, the maximum bin probability, and economically relevant tail probabilities, \[ Tail^K_{g,r,h}(a)=\Pr^K_{g,r,h}(y_{g,r}>a). \] For inflation, the main tail thresholds are fixed monthly rates, such as 0.3, 0.4, and 0.5 percent, as well as thresholds relative to the Reuters consensus forecast. For each release, snapshots are constructed at seven standardized horizons before market resolution: \[ h\in\{30\text{ days},14\text{ days},7\text{ days},3\text{ days},1\text{ day},6\text{ hours},1\text{ hour}\}. \] At each horizon, the most recent available distribution before the target time is selected and its staleness is recorded. This creates a release-by-horizon panel of distributional objects. The baseline CPI analysis uses snapshots with enough active bins to recover a meaningful distribution. Figure (ref) illustrates the recovered object for one CPI release.
The main empirical benchmark is the Reuters Poll consensus forecast. For each CPI and core CPI release, the Reuters data report the median forecast, low forecast, high forecast, realized release, and the surprise defined as actual minus the median forecast. Let \(C_{g,r}\) denote the Reuters median forecast. The surprise is \[ s_{g,r}=y_{g,r}-C_{g,r}. \] The Reuters high-low range is used as a measure of professional-forecaster disagreement. CPI and core CPI realized values are also constructed from FRED indexes as \(100\times(I_t/I_{t-1}-1)\). The FRED-based actuals closely match the Reuters actuals in the merged sample; the mean difference is 0.001 percentage points for CPI and 0.004 percentage points for core CPI. This diagnostic is important because the analysis requires aligning the Kalshi event, the reference month, the Reuters poll, and the realized release. Figure (ref) summarizes the recovered distributions across releases. The market-implied mean is economically plausible and relatively stable across horizons. Implied uncertainty is higher far from release and tends to decline as resolution approaches. This pattern is consistent with information arriving over time and being incorporated into the distribution of market beliefs.
The empirical tests are organized around a simple reduced-form framework in which the current professional consensus summarizes the center of the next release distribution. The framework is not a structural model of prediction-market trading. Its purpose is to clarify why recent inflation news may change the shape of market-implied beliefs even when the market-implied mean remains close to the professional consensus.
Consider release \(r\) for variable \(g\), and let \(C_{g,r}\) denote the Reuters median consensus forecast. Write the release outcome as
where \(x_{g,r}\) is the residual outcome relative to the contemporaneous professional point forecast. The prediction market prices a distribution over \(x_{g,r}\). Conditional on the market information set at horizon \(h\), suppose that this residual distribution can be represented as a mixture
where \(F^{0}\) is a normal-risk residual distribution, \(F^{H}\) is a high-inflation-risk residual distribution, \(q_{g,r,h}\in[0,1]\) is the market-implied probability of the high-risk state, and \(v_{g,r,h}\) governs residual uncertainty. The high-risk state need not be interpreted as a separate structural regime. It is a reduced-form way to capture the possibility that, after recent inflation news, market participants assign more probability to states in which the next release lies in the upper part of the inflation distribution. The key assumption is that the professional consensus absorbs much of the information about the center of the next release distribution. The prediction-market mean relative to the consensus is
If the consensus forecast already incorporates the central implications of public information, or if the high-risk state mainly changes skewness and tail thickness rather than the center of the residual distribution, then changes in \(q_{g,r,h}\) need not generate a large movement in \(\mu^{K}_{g,r,h}-C_{g,r}\). In that case, the market-implied mean may remain close to the professional consensus even though the distribution around that mean changes. The same change in \(q_{g,r,h}\), however, has direct implications for upper-tail probabilities. For a fixed inflation threshold \(a\), the market-implied probability of a high-inflation outcome is
Holding \(C_{g,r}\) fixed, the effect of the high-risk weight on this probability is
For high absolute inflation thresholds, such as 0.3, 0.4, or 0.5 percent month over month, the high-risk distribution assigns more mass above the threshold than the normal-risk distribution. Thus \(F^{H}_{g,h}(a-C_{g,r};v_{g,r,h})<F^{0}_{g,h}(a-C_{g,r};v_{g,r,h})\), and an increase in \(q_{g,r,h}\) raises the fixed upper-tail probability. The variance of the market-implied distribution also increases when either the high-risk state is more dispersed or the residual uncertainty parameter \(v_{g,r,h}\) rises. Recent macroeconomic news affects the market-implied distribution through these two objects. Let the previous Reuters surprise be \(s_{g,r-1}=y_{g,r-1}-C_{g,r-1}\). The high-inflation-risk weight is allowed to satisfy
where \(\Lambda(\cdot)\) maps the index into \([0,1]\), and \(Z_{g,r}\) contains contemporaneous controls such as the current consensus forecast. A positive lagged inflation surprise therefore raises the probability attached to the high-inflation-risk state. Residual uncertainty may instead respond to the size of recent news:
where \(Range^{R}_{g,r}\) is the Reuters high-low forecast range. This captures the idea that large recent surprises make the next release more uncertain, even after controlling for contemporaneous disagreement among professional forecasters. The framework delivers four empirical implications. First, the response of the market-implied mean relative to the Reuters consensus, \(\mu^{K}_{g,r,h}-C_{g,r}\), may be small even when recent news matters. Second, large lagged surprises should be associated with higher market-implied uncertainty. Third, positive lagged surprises should raise fixed upper-tail probabilities such as \(\Pr^{K}_{g,r,h}(y_{g,r}>0.3)\), \(\Pr^{K}_{g,r,h}(y_{g,r}>0.4)\), and \(\Pr^{K}_{g,r,h}(y_{g,r}>0.5)\), conditional on the current consensus. Fourth, the response of tails defined relative to the current consensus, \(\Pr^{K}_{g,r,h}(y_{g,r}>C_{g,r}+q)\), may be weaker because these thresholds move with the professional forecast and therefore remove part of the absolute inflation-risk component. The interpretation is deliberately reduced form. The Reuters consensus is a conditioning benchmark, not an exogenous treatment. The regressions below should therefore be read as predictive conditional relationships, not as causal estimates of belief updating. Their purpose is to test whether recent inflation news is reflected mainly in the location of market-implied beliefs or instead in uncertainty and upper-tail probabilities.
The first test asks whether recent inflation surprises shift prediction-market means away from the contemporaneous professional consensus. The estimating equation is
where \(\alpha_{g,h}\) denotes variable-by-horizon fixed effects. A positive \(\beta\) would indicate extrapolation in the central forecast: after an above-consensus release, the prediction-market mean for the next release is higher than the current Reuters consensus. A coefficient close to zero does not imply that the market ignores recent news. It means only that recent news does not create a systematic gap between the Kalshi mean and the professional consensus.
The second test studies whether large surprises widen the market-implied distribution. The specification is
where \(Range^R_{g,r}\) is the Reuters high-low forecast range for the current release. The range control is useful because professional-forecaster disagreement may itself be high when the current release is difficult to forecast. The coefficient \(\eta\) therefore asks whether the market-implied distribution is wider after large previous surprises, conditional on contemporaneous survey disagreement.
The third test examines upper-tail probabilities. For fixed inflation thresholds \(a=0.3,0.4,0.5\), the specification is
The current consensus control matters because a higher expected current release mechanically raises the probability of exceeding a fixed high-inflation threshold. A positive \(\delta\) in equation (ref) means that recent inflation news predicts additional probability mass in high-inflation states, beyond what is explained by the current consensus forecast. Additional tail specifications let the threshold move with the current consensus,
for \(q\in\{0,0.1\}\). These relative-tail regressions distinguish a general increase in absolute inflation risk from a systematic disagreement with the current consensus. If lagged surprises raise fixed high-inflation tails but not consensus-relative tails, the interpretation is that the inflation-risk environment has changed, not that prediction markets are simply biased above the current professional forecast. All baseline standard errors are clustered by release event, which accounts for dependence across horizons within the same statistical release. The main specifications pool CPI and core CPI and include variable-by-horizon fixed effects; the tables also report simpler horizon and variable fixed-effect specifications where useful. The online appendix reports moving-average benchmark exercises and labor-market external-validity tests.
This section reports the main empirical results. The central finding is that the Kalshi-implied mean is informative, but it is not the main margin on which recent inflation news appears. Lagged Reuters surprises do not predict systematic movements of Kalshi means away from the current Reuters consensus. Instead, recent surprises are reflected in distributional shape: uncertainty and fixed high-inflation tail probabilities.
The first exercise evaluates whether the recovered market-implied means contain meaningful information about the upcoming release. For each release \(r\) and horizon \(h\), the Kalshi-implied mean \(\mu^K_{r,h}\) is compared with realized month-over-month inflation \(y_r\). The benchmark forecasts are the previous realization, a three-month moving average, and a six-month moving average, all constructed using only past realizations. Table (ref) reports the results.
For headline CPI, the Kalshi mean substantially improves on simple time-series benchmarks. Its RMSE is 0.169 percentage points, compared with 0.331 for the previous realization, 0.280 for the three-month moving average, and 0.274 for the six-month moving average. For core CPI, the improvement is more modest: Kalshi is broadly competitive with rolling-average benchmarks but does not dominate them. Figure (ref) plots the one-hour-ahead Kalshi-implied mean against the realized release. The figure shows substantial information in the market mean, while also making clear that individual events can contain large errors.
The Kalshi panel is next merged with Reuters Poll consensus forecasts. For each CPI and core CPI release, the Reuters data report the median forecast, low forecast, high forecast, actual release, and surprise. The merged estimation sample contains 484 release-horizon observations with lagged Reuters surprises: 262 for CPI and 222 for core CPI. These observations correspond to 79 release clusters. Standard errors are therefore clustered by release throughout the baseline analysis, and release-level validation exercises below provide a check that the results are not generated by treating horizons as independent events.
The Reuters actuals closely match the FRED-based actuals used in the Kalshi panel. The mean difference between the FRED-based actual and the Reuters actual is 0.001 percentage points for CPI and 0.004 percentage points for core CPI, supporting the event-date alignment. Table (ref) summarizes the main regressions. Panel A begins with the mean gap \(\mu^K-C\). With variable-by-horizon fixed effects, the coefficient on the lagged Reuters surprise is \(-0.023\), with a clustered standard error of 0.107. The estimate is economically small and statistically indistinguishable from zero. Past inflation surprises therefore do not lead prediction-market means to deviate systematically from current Reuters consensus forecasts. The second row of Panel A studies uncertainty. The dependent variable is the Kalshi-implied standard deviation, and the regressor is the absolute lagged Reuters surprise. The specification controls for the Reuters high-low forecast range. The coefficient is 0.089, with a clustered standard error of 0.053. This estimate is less precise than the tail estimates below, but it is positive and economically consistent with the view that large recent surprises widen the market-implied distribution. The positive role of the Reuters range in the underlying horse-race regressions also validates the distributional measure: Kalshi-implied uncertainty comoves with professional-forecaster disagreement.
The strongest evidence appears in upper-tail probabilities. Panel B of Table (ref) reports regressions for \(\Pr^K(y>0.3)\), \(\Pr^K(y>0.4)\), and \(\Pr^K(y>0.5)\), controlling for the current Reuters consensus forecast and variable-by-horizon fixed effects. The coefficients on the lagged Reuters surprise are 0.466, 0.422, and 0.281, respectively. The magnitudes are large. A 0.1 percentage point positive surprise in the previous release raises the probability assigned to inflation above 0.3 percent by about 4.7 percentage points, even after conditioning on the current consensus forecast. Panel C shows that the result is weaker when tails are defined relative to the current consensus forecast. The coefficient on \(\Pr^K(y>C)\) is 0.061 and statistically insignificant; the coefficient on \(\Pr^K(y>C+0.1)\) is 0.173 and also statistically insignificant. This difference is informative. Lagged surprises do not appear to make Kalshi systematically disagree with Reuters about whether the current release will exceed consensus. Instead, positive previous surprises raise the probability of fixed high-inflation outcomes. Recent inflation news therefore changes the perceived absolute inflation-risk environment. Figure (ref) summarizes the same evidence graphically. The coefficient for the mean gap is close to zero. The uncertainty and fixed-tail coefficients are positive. The consensus-relative tail coefficients are small and imprecise. This is the central empirical message: point forecasts miss the main signal in prediction-market data.
The previous results show that recent inflation news is reflected in the shape of the market-implied distribution rather than in a systematic gap between the Kalshi mean and the Reuters consensus. A natural question is whether these upper-tail probabilities are empirically meaningful as measures of inflation risk. This subsection therefore asks whether Kalshi-implied tail probabilities forecast the realization of high-inflation states. For each release-horizon observation, define realized tail indicators \[ \mathbf{1}\{y_{g,r}>a\}, \qquad a\in\{0.3,0.4,0.5\}, \] and compare them with the corresponding Kalshi-implied probabilities \(\Pr^K_{g,r,h}(y_{g,r}>a)\).
The exercise is deliberately simple. It does not ask whether prediction markets dominate every survey-based benchmark in every scoring rule. Instead, it asks whether the market-implied upper tail has useful early-warning content for adverse inflation outcomes, especially in states where the mean forecast may remain close to consensus. Table (ref) reports two validation exercises. Panel A sorts release-horizon observations into terciles of the Kalshi-implied probability of inflation above 0.4 percent and compares the mean predicted probability with the realized frequency of the event. The realized frequency rises sharply across the tail-risk distribution. At the one-hour horizon, the event \(y>0.4\) occurs in 7.7 percent of low-tail observations, 12.0 percent of middle-tail observations, and 57.7 percent of high-tail observations. The corresponding mean Kalshi probabilities are 0.0, 8.0, and 60.5 percent. Panel B reports linear probability regressions of realized high-inflation indicators on the corresponding Kalshi-implied tail probability. The regressions include Reuters controls: the current Reuters median consensus and the Reuters high-low forecast range. At the one-hour horizon, the coefficient on \(\Pr^K(y>0.3)\) is 0.744, with a standard error of 0.159. For the more extreme event \(y>0.5\), the coefficient is 0.722, with a standard error of 0.272. The discrimination statistics are also strong: the AUC is 0.838 for \(y>0.3\), 0.815 for \(y>0.4\), and 0.926 for \(y>0.5\). These results provide a direct validation of the inflation-at-risk interpretation. Kalshi tail probabilities are not merely transformations of noisy contract prices. They identify states in which high-inflation outcomes are more likely to occur.
The results should not be read as showing that Kalshi probabilities uniformly dominate all survey-based distributional benchmarks. Rather, they show that prediction-market tails contain economically meaningful information about high-inflation states, including after conditioning on Reuters consensus information. This validation is also useful for interpretation. The main regressions show that lagged inflation surprises raise fixed upper-tail probabilities but do not systematically move the Kalshi mean away from Reuters consensus. The early-warning evidence shows why that margin matters: a stable central forecast can coexist with a market-implied distribution that assigns substantial probability to adverse inflation outcomes. In the one-hour release-level sample, there are 30 episodes in which the Kalshi mean is within 0.05 percentage points of the Reuters consensus but \(\Pr^K(y>0.4)\) lies in the upper tail-risk tercile. In these anchored-mean/high-tail episodes, the event \(y>0.4\) occurs in 50 percent of cases. This is the sense in which prediction-market distributions can provide a real-time measure of inflation-at-risk that is not visible in point forecasts alone.
The previous subsection shows that Kalshi upper-tail probabilities contain early-warning information about high-inflation states. This subsection asks a more operational question: would a policymaker or market participant who monitors the upper tail of the Kalshi distribution identify high-inflation releases that are missed by point forecasts? I focus on headline CPI at the one-hour horizon and define a high-inflation release as \[ 1\{y_r>0.4\}, \] where \(y_r\) is monthly CPI inflation in percentage points. I compare three simple monitoring rules. The first signals high inflation when the Reuters median consensus exceeds \(0.4\) percent. The second signals high inflation when the Kalshi-implied mean exceeds \(0.4\) percent. The third uses the distributional object directly and signals high inflation when \[ \Pr^K_{r,-1h}(y_r>0.4) \geq 0.50 . \] The threshold of 50 percent is intentionally simple: it asks whether the market-implied distribution assigns the high-inflation state at least even odds. The exercise is not designed to optimize a classification rule. Its purpose is to illustrate the information lost when the monitoring problem is reduced to a point forecast.
Table (ref) reports the results. Panel A first sorts releases into terciles of the Kalshi-implied probability of inflation above \(0.4\) percent. The realized frequency of high-inflation outcomes rises from 7.1 percent in the low-tail-risk tercile to 76.9 percent in the high-tail-risk tercile. The corresponding mean Kalshi probabilities are 0.4 and 74.9 percent, respectively. Thus, in this release-level monitoring sample, the upper tail of the Kalshi distribution ranks high-inflation risk in a way that is both monotone and economically large. Panel B compares the three monitoring rules. The Reuters consensus and the Kalshi-implied mean each identify 8 of the 13 high-inflation releases, with no false alarms. The Kalshi tail rule identifies 10 of the 13 high-inflation releases, while generating only one false alarm. Its hit rate is therefore 76.9 percent, compared with 61.5 percent for either point-forecast rule, and its precision remains high at 90.9 percent. The overall accuracy of the tail rule is 90.0 percent. The exercise highlights the practical value of observing a distribution rather than only a mean. Point forecasts are conservative signals: when they cross a high-inflation threshold, the signal is very precise, but they miss several high-inflation outcomes. The Kalshi upper tail provides a complementary monitoring statistic. It identifies additional high-inflation releases by using information about the shape of the market-implied distribution, not only its center. This is the sense in which prediction-market distributions can reduce the false comfort created by apparently stable point forecasts.
Section (ref) shows that this pattern is robust to quote staleness, liquidity controls, alternative tail-support assumptions, wild cluster bootstrap inference, CPI/core CPI splits, and leave-one-release-out diagnostics. These checks are important because the empirical object is a distribution recovered from a new and sometimes thinly traded market. They also sharpen the interpretation: the stable result is distributional risk, not a persistent shift in the conditional mean relative to Reuters consensus. Taken together, the results show that prediction-market beliefs are informative, but their distinctive contribution is distributional. Kalshi means contain information about realized inflation and remain aligned with professional consensus forecasts after past surprises. At the same time, recent inflation surprises reshape the market-implied distribution. Large surprises are associated with higher uncertainty, and positive surprises increase the probability of high-inflation states. This distributional margin is difficult to detect with point forecasts alone.
The main results rely on a new market and on distributions recovered from discrete event-contract grids. This section summarizes robustness exercises designed to address the most direct threats to interpretation: stale quotes, liquidity, tail construction, finite-cluster inference, heterogeneity between CPI and core CPI, and influential releases. The tests support a narrow interpretation of the results. The robust finding is not a systematic drift of the market-implied mean away from professional consensus. It is that recent inflation news is reflected in distributional risk: implied uncertainty and fixed high-inflation tail probabilities.
Table (ref) summarizes the main diagnostics. Staleness is not driving the results. The baseline panel has a median staleness of about 6.4 hours, and all observations are within 24 hours of the target snapshot. When the sample is restricted to snapshots no more than one hour stale, the implied-uncertainty coefficient remains positive and statistically significant, and the fixed-tail coefficients become larger. For example, the coefficient for \(\Pr^K(y>0.3)\) rises from 0.466 in the baseline to 1.198 in the one-hour-staleness subsample. The mean-gap coefficient remains imprecise. The results are also robust to controlling directly for market activity. Adding continuous controls for snapshot staleness, \(\log(1+\mathrm{open\ interest})\), and \(\log(1+\mathrm{volume})\) leaves the implied-standard-deviation coefficient positive. The coefficient for \(\Pr^K(y>0.4)\) remains significant at the five percent level, while the coefficients for \(\Pr^K(y>0.3)\) and \(\Pr^K(y>0.5)\) remain positive and marginally significant. Liquidity-filter results in the online appendix show a similar pattern when liquidity is measured by open interest. Volume filters are less stable because many hourly snapshots have zero measured volume even when open interest is substantial. Finite-cluster inference is somewhat more conservative but does not overturn the main conclusion. With a wild cluster bootstrap by release, the \(p\)-values are 0.018 for implied standard deviation, 0.033 for \(\Pr^K(y>0.4)\), 0.057 for \(\Pr^K(y>0.3)\), and 0.054 for \(\Pr^K(y>0.5)\). Thus the evidence is strongest for uncertainty and the middle fixed upper-tail threshold, with the other fixed tails remaining statistically meaningful under conservative inference. Consensus-relative tail probabilities remain weak.
The online appendix reports three further checks. First, alternative tail-support rules and a uniform-within-bin interpolation deliver positive fixed-tail coefficients across constructions, indicating that the tail results are not an artifact of assigning bin probabilities to a particular support point. Second, split-sample regressions for CPI and core CPI show the same qualitative pattern but less precision, as expected given the smaller number of release clusters. The uncertainty response is particularly strong for core CPI. Third, leave-one-release-out diagnostics show that the uncertainty and fixed-tail coefficients do not depend on a single release. The mean-gap coefficient is the least stable object, which is consistent with the paper's interpretation that the distinctive information in prediction-market distributions is not the conditional mean alone.
Prediction-market prices are attractive because their payoffs are directly linked to realized macroeconomic outcomes. This feature makes them closer to beliefs about specific statistical releases than many asset-price-based measures. At the same time, the recovered objects should not be interpreted mechanically as frictionless subjective probabilities. Under strong assumptions---risk neutrality, common priors, competitive trading, no fees, and sufficient liquidity---a binary contract price equals the probability of the payoff-relevant event. In practice, prediction-market prices can embed risk premia, heterogeneous beliefs, trading fees, bid-ask spreads, depth constraints, and stale quotes. Evidence from recent studies of Kalshi and Polymarket microstructure reinforces this point: event-contract prices are informative, but they can also reflect favorite-longshot bias, liquidity provision, platform design, and heterogeneous trader sophistication BurgiDengWhelan2026,Dubach2026,AkeyGregoireHarvieMartineau2026. The objects in this paper are therefore market-implied belief distributions: they combine information, beliefs, risk preferences, and market microstructure.
This interpretation is analogous to other market-based expectation measures. Inflation swaps, TIPS breakevens, options, and futures provide valuable information, but they also embed risk and liquidity premia. Prediction markets have the advantage that the payoff is more directly tied to the macroeconomic outcome. Their disadvantage is that markets may be thinner and contract grids may be discrete or incomplete. The empirical design addresses these concerns by aggregating prices to hourly frequency, using standardized pre-release snapshots, clustering standard errors by release, and focusing on repeated patterns across events rather than on individual prices. The robustness exercises further show that the main uncertainty and fixed-tail results are not explained by stale quotes, open interest, volume, tail-support choices, or a single influential release. Survey density forecasts are an important benchmark but not a one-for-one substitute for the object studied here. The Survey of Professional Forecasters and related density data are designed for lower-frequency horizons and are well suited to studying disagreement, uncertainty, and macroeconomic risk over quarters or years Croushore1993,EngelbergManskiWilliams2009,AndradeCrumpEusepiMoench2016. Kalshi contracts instead price the probability distribution of a specific release at intramonth horizons. The Reuters poll is therefore used as the main conditioning benchmark because it provides a release-level professional point forecast and a high-low range for the same CPI and core CPI announcements. The empirical claim is not that Kalshi replaces survey densities. It is that release-contingent prediction markets provide a complementary high-frequency distribution that is unavailable from standard point forecasts. The main economic interpretation is that recent inflation news affects perceived inflation risk more than disagreement between prediction markets and professional forecasters. The Kalshi mean does not systematically move above the Reuters consensus after a positive lagged surprise. Nor does the probability of exceeding the current consensus forecast respond strongly. Instead, fixed high-inflation tail probabilities rise. This pattern is consistent with a change in the perceived inflation-risk environment: the consensus forecast may remain a good summary of the center of the distribution, while the market assigns more probability to states that would be especially salient for monetary policy and financial markets.
Kalshi prediction-market contracts make it possible to recover high-frequency distributions of short-run macroeconomic beliefs. The empirical setting is CPI and core CPI releases, where adjacent threshold contracts imply probability mass over inflation outcomes at standardized horizons before release. The resulting data reveal not only market-implied means, but also uncertainty and upper-tail probabilities. The main result is distributional. Kalshi-implied means contain useful information about realized headline CPI and remain broadly aligned with Reuters Poll consensus forecasts. Lagged Reuters surprises do not predict systematic deviations of the Kalshi mean from the current Reuters consensus. By contrast, recent inflation news predicts changes in the shape of the distribution. Large previous surprises are associated with higher implied uncertainty, and positive previous surprises raise the probability assigned to fixed high-inflation outcomes even after controlling for the current consensus forecast. The robustness evidence supports this interpretation. The uncertainty and fixed-tail results survive staleness filters, continuous controls for staleness and market activity, alternative tail-support and interpolation rules, wild cluster bootstrap inference, and leave-one-release-out diagnostics. The strongest and most stable finding is not that prediction markets persistently move their mean above professional consensus after a surprise. It is that recent inflation news is reflected in market-implied inflation risk. These findings suggest that prediction markets can add to macroeconomic monitoring by measuring short-run inflation risk in real time. Point forecasts are useful, but they can miss changes in the distribution that matter for monetary policy and financial markets. A stable expected release can coexist with a higher probability of a high-inflation state. Prediction-market distributions make this distinction observable. Future work should extend the sample as these markets mature, compare prediction-market distributions with survey density forecasts and options-implied distributions, and use full pre-release distributions to construct distributional macroeconomic news measures for structural analysis.
\phantomsection \addcontentsline{toc}{section}{Online Appendix}