Don't Worry About the Future.
- James C. McGrath
- 17 hours ago
- 14 min read
Updated: 17 minutes ago
Until You Really Understand the Present
Prediction Markets Are on Point
Prediction markets are once again part of the public conversation. Over the past year, they have attracted renewed attention in the popular press, a growing alternative to polls or expert forecasts. This attention reflects genuine developments: improved market infrastructure, broader participation, growing institutional curiosity, and, in some cases, integration with crypto rails and AI-assisted forecasting tools. Dissatisfaction with traditional opinion surveys and expert judgment has only amplified interest.
Much of this coverage, however, has focused on whether prediction markets “get it right,” whether they can outperform polls, or whether they are poised for explosive growth. Far less attention has been paid to a quieter but more consequential question: what information these markets reveal while uncertainty remains unresolved.
This article is conservative. Its purpose is not to forecast the future size of prediction markets, nor to evangelize for their adoption. Instead, it asks a more modest and, I believe, a more useful question: what do prediction markets already reveal when they function well? In particular, what can be learned from how beliefs evolve over time, before outcomes are known?
Three Ways of Answering the Same Question
Consider a single, concrete event: a Federal Open Market Committee (FOMC) policy decision. There are at least three familiar ways to aggregate information about such an event.
Polls summarize stated opinions. They capture what respondents say they believe at a moment in time, typically without requiring them to act on those beliefs or update them continuously.
Capital markets, such as interest-rate futures, embed expectations in prices shaped by risk transfer, hedging demand, and institutional constraints. These prices are highly informative, but they are not designed to be easily interpreted as beliefs about discrete outcomes.
Prediction markets occupy a distinct middle ground. Participants express beliefs by trading contracts that resolve to simple, well-defined outcomes, and they can revise those beliefs continuously as information arrives.
Prediction markets occupy a middle ground between polls and capital markets: more disciplined than opinion surveys, more interpretable than risk-transfer prices, and uniquely suited to revealing how beliefs change over time. These mechanisms are complements, not substitutes. The error is to evaluate prediction markets as if their sole purpose were to generate a final answer comparable to a poll result or a futures price.
Why “Being Right” Is the Wrong Yardstick
Ex-post correctness is an appealing metric. It is clean, binary, and intuitive. A market either converged to the realized outcome or it did not. It crystallizes regret or exaltation... 'I could have made a lot of money from that!'
But correctness is not the only source of informational value. Terminal prices collapse a rich process into a single endpoint. They tell us where beliefs landed, but not how they got there, how much disagreement existed along the way, how quickly information was incorporated, or how fragile consensus may have been. Two markets that end at the same price can reflect radically different informational environments.
Most of the informational action in prediction markets occurs before outcomes resolve. Beliefs shift as participants interpret data releases, speeches, rumors, and second-order expectations about how others will react. By the time a decision is announced, much of this work has already been done. Markets that end up “wrong” often exhibit just as much—and sometimes more—belief updating, disagreement resolution, and information flow as markets that end up “right.” This motivates a different empirical focus: not whether markets converge correctly, but how beliefs evolve while uncertainty is still live.
To test this empirically, I turn to a concrete setting where belief evolution can be observed directly: FOMC decision markets on Kalshi, a regulated, open-access prediction market platform offering contracts that resolve to clear, binary outcomes.
A Concrete Setting: 2025 FOMC Decisions on Kalshi
In 2025, the Federal Reserve shifted from holding rates steady to beginning a cautious easing cycle. After keeping the federal funds target unchanged through mid-year, the FOMC delivered three consecutive 25 bps cuts in the second half of the year, starting at the September meeting. This broadly conformed to market expectations that easing would begin in 2025 as inflation cooled and growth slowed, though the pace was more gradual than some early-year forecasts.
FOMC decisions are particularly well suited to studying belief dynamics. They are recurring events with high stakes, well-defined resolution, and intense information flow. Each meeting generates a strip of mutually exclusive contracts tied to possible target-rate outcomes, allowing beliefs to be expressed as a distribution rather than a single point estimate. All market price data used in this article are drawn from publicly available historical price series. No proprietary or non-public information is used. Meeting dates and announcement timing are aligned to the scheduled FOMC calendar as recorded in Federal Reserve Economic Data, but no macroeconomic variables from FRED enter the analysis.
Exhibit 1 FOMC Meetings and Associated Prediction Markets (2025)
Source: Kalshi (public historical data)
This table summarizes all FOMC-related prediction markets closing in 2025, grouped by meeting. Each meeting is represented by a strip of mutually exclusive contracts, rather than a single forecast.
One might expect that there would be more activity on Kalshi when something is about to happen, but the data are more nuanced than that.
What I Measure
For each contract, I collected one-minute price candlesticks in a standardized event window: two days prior to market close through 24 hours after. Prices are indexed by both calendar time and normalized event time—minutes relative to market close—placing all meetings on a common temporal axis.
From these paths, I focused on a small set of interpretable, market-internal descriptors:
Total belief movement: cumulative absolute price changes prior to resolution
Volatility: dispersion of beliefs during the pre-resolution period
Reversals: frequency of sign changes in belief direction
Pre- vs post-resolution movement: comparison of belief updating before and after outcomes are known
Participation intensity: trading volume and open interest
These measures describe how beliefs evolve, not which outcome ultimately prevails.
Equally important is what I don't try to do: I'm not forecasting FOMC decisions. I'm not evaluating trading profitability. I'm not attempting to score markets by ex-post accuracy.
This restraint is intentional: terminal accuracy collapses belief trajectories into a single binary result and obscures the informational process that unfolds beforehand.
Exhibit 2 Meeting-Level Belief Dynamics Summary
Source: Kalshi (public historical data)
This table reports meeting-level summaries of belief dynamics, including pre-resolution churn, volatility, reversals, and participation. The dispersion across meetings highlights substantial heterogeneity in belief environments that is not captured by outcome correctness.
Pre- and post-resolution belief movement are measured as cumulative absolute changes in contract prices. Volatility reflects dispersion of belief paths prior to resolution. Reversals count sign changes in belief direction during the pre-resolution period. Volume is total traded contracts across all outcomes for the meeting.
The following charts show belief dynamics visually. First, (a) variation across meetings, then (b) the relationship between churn and reversals, finally (c) detailed paths for a single meeting.
Exhibit 3 Pre-Resolution Churn by Meeting
Source: Kalshi (public historical data)
This bar chart reports cumulative belief movement prior to resolution for each FOMC meeting in 2025. Substantial variation across meetings highlights differences in the intensity of belief updating and disagreement while uncertainty remains live, independent of the realized policy outcome.
Exhibit 4 Relationship Between Churn and Reversals Across Meetings
Source: Kalshi (public historical data)
Each point represents one FOMC meeting. The x-axis shows cumulative belief movement prior to resolution; the y-axis shows the number of belief reversals during the same period. Meetings with higher pre-resolution churn tend to exhibit more frequent reversals, reflecting sustained belief updating and disagreement while uncertainty remains unresolved rather than post-resolution correction.
Exhibit 5 Belief Paths Around a Single FOMC Meeting
Source: Kalshi (public historical data)
This figure plots normalized belief paths for all target-rate contracts associated with a single FOMC meeting. Each line represents the close price of one contract, expressed in minutes relative to market close (t = 0). Substantial belief updating, reversals, and cross-contract coordination are visible well before resolution, while post-resolution adjustments are smaller and shorter-lived.
Testing for Patterns in Belief Evolution
To be clear: when I examine 'announcement moves' below, I am not attempting to predict Fed decisions or market outcomes. Rather, I'm testing whether the pattern of belief evolution itself—how much uncertainty remains to be resolved at announcement time—follows predictable structures. Large announcement moves indicate markets were still actively processing information; small moves suggest beliefs had already converged. The question is whether we can anticipate these different information-processing regimes, not where prices will end up.
The objective is not to predict policy decisions or improve forecasts, but to test whether belief evolution itself exhibits interpretable structure. Using meeting-level summaries of belief churn, volatility, reversals, and participation, I apply a few simple, interpretable models—such as decision trees—trained solely on market-internal features.
Having characterized how beliefs evolve, I now test whether these evolutionary patterns themselves follow predictable structures. This is fundamentally different from predicting outcomes - I'm asking whether we can anticipate the type of belief process we'll observe, not where it will end.
These models offer little leverage in predicting realized outcomes. They do not meaningfully separate “right” from “wrong” meetings. At the same time, they are effective at organizing belief dynamics. They recover distinctions already visible in the data: meetings characterized by orderly convergence versus those marked by sustained debate, reversals, and late-stage uncertainty. Machine learning succeeds precisely where it should. Whether or not it improves outcome prediction? That's a question for another time.
Note that 'announcement move' here is not about predicting the Fed's decision - it's about predicting how much belief updating work remains to be done when the announcement arrives. Large announcement moves indicate the market was still actively processing information; small moves suggest beliefs had already converged. The ML models test whether we can anticipate these different information processing regimes.
What the analysis does reveal is the following: in 2025, meetings where there was higher market volatility or uncertainty leading up to the announcement (higher 'Pre-Churn') tended to experience larger shifts in market belief when the announcement actually came out (larger 'Announcement Move'). Conversely, meetings preceded by more stable market beliefs saw more modest shifts post-announcement.
The following chart distills this. Below, each data point represents a single FOMC meeting in 2025 where market data was analyzed.
X-axis ('Pre-Churn'): Shows the standard deviation of normalized market belief in the two-day period leading up to the FOMC meeting. This acts as a proxy for pre-event market volatility or 'nervousness'. A higher value on this axis means market beliefs were fluctuating more significantly before the announcement.
Y-axis ('Announcement Move'): Shows the change in average normalized market belief from the day before the FOMC announcement to the day of the announcement. This quantifies the impact or shift in collective market belief directly attributable to the FOMC's decision.
There is a statistically meaningful positive correlation here, which boils down the insight that periods of greater pre-FOMC market turbulence are associated with more pronounced market reactions to the Fed's decisions.
Exhibit 6 'Pre-Churn vs. Announcement Move for Meetings'
Source: Kalshi (public historical data)
This correlation reveals something crucial: markets that exhibit rich belief dynamics before announcements continue processing information actively through resolution. The informational work doesn't suddenly begin at announcement time—it's a continuous process, with pre-announcement volatility signaling how much uncertainty remains to be resolved.
What This Means for Practitioners
For practitioners—investors, analysts, and policymakers—the implications are quite useful, if you look at them the right way.
First, disagreement and churn are signals, not noise. High belief movement often reflects genuine uncertainty and active information processing, even when markets ultimately converge incorrectly.
Second, timing matters. Early repricing and pre-resolution dynamics frequently contain more information than announcement-day reactions.
Third, interpretability beats accuracy. Prediction markets are most useful when treated as diagnostic instruments that reveal how expectations are forming, not as oracles that deliver final answers.
Drilling deeper, the modeling reveals some additional insights:
Early Warning Signals for Directional Shifts: The analysis demonstrates that pre-event market dynamics, specifically pre-event belief volatility and average trading volume, are statistically significant indicators of subsequent market belief shifts. Investment managers may be able to leverage changes in these metrics on prediction markets as an early signal to anticipate the direction of market sentiment leading up to FOMC announcements.
High Confidence in Identifying Stable vs. Volatile Belief Environments: Practitioners can distinguish meetings likely to generate substantial belief revision from those with predetermined consensus. When the model predicts a non-positive shift, it's highly likely to be correct, and it effectively captures all actual non-positive shifts. This offers a strong signal for managers looking to mitigate downside risk or position for neutral-to-negative market reactions around rate decisions.
Nuanced Insights for Positive Moves: When the model does predict a positive shift, it is highly accurate. However, it only captured half of the actual positive shifts. For investment managers, this could imply that while a positive prediction from the model is a strong indicator, the absence of such a prediction doesn't necessarily rule out a positive move. Further analysis or complementary signals might be needed for identifying all instances of positive sentiment shifts.
Quantifying Market Conviction: The pre-event volatility metric showed a high correlation to subsequent market belief shifts. This indicates that periods of higher market uncertainty or debate (as reflected in belief dispersion on Kalshi) are reliably associated with larger price movements following the actual Fed decision. Managers can use this to gauge the potential impact magnitude and adjust their positions accordingly.
The model successfully distinguishes between different types of information environments: some meetings generate sustained debate and belief revision, others see orderly convergence. I'd argue that the taxonomy of belief evolution is more valuable than knowing which outcome will prevail.
Used this way, prediction markets complement rather than replace other tools. They provide a real-time window into belief formation that polls and traditional market prices are not designed to offer.
Conclusion:
The central claim of this article is not that prediction markets are especially good at producing correct answers. It is that they are unusually good at revealing how beliefs evolve while uncertainty is still unresolved.
The data support this view directly. Across FOMC-related prediction markets on Kalshi, belief updating is concentrated well before decisions resolve, varies meaningfully across meetings, and often exhibits substantial churn regardless of whether markets ultimately converge to the realized outcome.
Markets that end up “wrong” frequently display belief dynamics that are just as rich—and sometimes richer—than markets that end up “right.”
Machine learning reinforces, rather than challenges, this interpretation. When applied to market-internal features alone, simple and interpretable models help organize belief dynamics into recognizable patterns, while offering little leverage in predicting outcomes themselves. This asymmetry underscores where the informational value of prediction markets truly lies.
There are some specific takeaways I'd emphasize:
Prediction Markets as Quantitative Tools, Not Just Sentiment Barometers:
While prediction markets are often seen as broad sentiment indicators, this analysis demonstrates their utility as platforms for generating quantifiable features and targets related to market belief. These can be rigorously analyzed using machine learning to derive actionable insights, moving beyond simple 'yes/no' outcomes.
Improving the 'Activity-Equals-Action' Hypothesis:
The initial intuition that "more activity on Kalshi when something is about to happen" is broadly supported by the role of average trading volume and pre-event belief volatility as significant features.
However, the analysis clarifies that it's not just any activity, but the nature of that activity (e.g., its volatility and sheer volume) that provides predictive power for the direction and intensity of market belief shifts. The data are indeed more nuanced than a simple direct correlation.
Objective Reflection of Evolving Expectations:
Prediction markets, as evidenced by their behavior around the 2025 Fed rate cuts, offer a dynamic and real-time reflection of aggregated market expectations.
Their value lies in providing a leading indicator of how market participants collectively interpret information and price in future events, distinct from conventional economic forecasts or polls.
Robustness of Findings:
The significant gap between the actual classification model's accuracy (0.8750) and the negative test's accuracy (0.6250, close to a random baseline) confirms that the observed patterns are statistically meaningful and not due to chance. This solidifies the conclusion that prediction markets capture genuine market dynamics that can be leveraged for informed decision-making.
What does this all mean? In a nutshell, I'd assert that prediction markets are not best understood as machines that produce correct answers. They are systems that allow us to dynamically measure uncertainty in real-time. Practitioners willing to look beyond the final verdict, actively managing positions dispassionately as conditions change, can extract genuine economic value from these increasingly ubiquitous tools.
By shifting focus from terminal accuracy to belief evolution, we can better understand not just what markets think, but how collective intelligence forms under uncertainty.
Addendum: Words of Caution
This is a small dataset. While the analytical approach (LOOCV) is appropriate for this sort of analysis, and the results show internal statistical significance, especially when compared to the negative test, it is crucial to interpret these conclusions with caution regarding their generalizability. They suggest strong patterns within the observed 2025 data, providing valuable insights and a robust framework for analysis. However, for broader predictive claims or application to different economic periods, a larger dataset would be highly desirable to confirm these findings and establish greater confidence in their robustness across diverse market conditions.
Crucially, the significant difference between the actual model's accuracy (0.8750) and the negative test's accuracy (0.6250) indicates that the model is indeed capturing real patterns within this dataset, rather than just random noise. If the accuracies were similar, it would strongly suggest the findings are not meaningful.
For more details on the analysis, please continue to the appendix.
Appendix:
Methodologies
This analysis systematically quantifies market belief shifts around Federal Open Market Committee (FOMC) meetings using Kalshi market data for the year 2025. The methodology involves:
Data Acquisition & Preparation: Raw 1-minute candlestick data and market metadata for FOMC-related contracts (KXFED series) were collected. Market metadata was used to identify distinct FOMC meetings (via event_ticker aka 'mnemonic' in the tables) and their corresponding close_time.
Daily Aggregation & Belief Vector Construction: For each meeting, 1-minute candlestick data was aggregated to a daily resolution, capturing the last price and total volume for each contract per day. Contract prices for mutually exclusive outcomes within a given meeting were then normalized daily to form 'belief vectors', representing probabilities summing to 1.0 per day.
Feature Engineering: Pre-event features were engineered from the daily belief vectors and volumes within a 2-day window prior to each market's close_time:
pre_avg_belief: Average normalized belief.
pre_avg_volume: Average daily trading volume.
pre_belief_std: Standard deviation of normalized belief (proxy for pre-event volatility/churn).
num_contracts_pre: Number of unique contracts in the pre-event window.
Target Variables Definition: Two primary target variables were defined for each meeting:
ann_move (Regression Target): The change in average normalized belief from the day before the event to the day of the event, quantifying the magnitude and direction of the announcement's impact.
regime (Classification Target): A binary variable (0 or 1) indicating whether the ann_move was positive (1) or non-positive (0), representing the directional outcome.
Meaningfulness Statistics
Correlation between Pre-Churn and Announcement Move: A positive Pearson correlation coefficient of r = 0.5742 was observed between pre_churn (pre-event belief volatility) and ann_move (change in belief around the event). This indicates a moderately strong linear relationship, suggesting that higher pre-event volatility is associated with larger belief shifts during the announcement period.
Regression Model Performance (Predicting pre_churn):
Decision Tree Regressor (LOOCV): Achieved a Mean Absolute Error (MAE) of 0.1557 and a Mean Squared Error (MSE) of 0.0341. These metrics provide a measure of the model's accuracy in predicting the magnitude of pre-event volatility.
Feature Importances (Random Forest Regressor): For predicting the targets, the Random Forest model highlighted the following features as most impactful:
pre_belief_std (Pre-event Belief Standard Deviation): 0.5490 (normalized importance).
pre_avg_volume (Pre-event Average Volume): 0.2677 (normalized importance). These large importance values indicate that these features are robust predictors.
Classification Model Performance (Predicting regime):
Decision Tree Classifier (LOOCV): Demonstrated an overall accuracy of 0.8750 in predicting the directional regime of market belief shift. The detailed classification report showed high precision and recall for the 'non-positive' class (Precision: 0.86, Recall: 1.00) and moderate performance for the 'positive' class (Precision: 1.00, Recall: 0.50).
Negative Test Results: A negative test, where the classification model attempted to predict randomly shuffled regime labels, yielded an accuracy of 0.6250. The significantly higher accuracy of the actual classification model (0.8750) compared to the random baseline (expected closer to 0.5 or the negative test result) strengthens the statistical validity of the findings, indicating that the model captures genuine patterns rather than spurious correlations.
Statistically Meaningful Conclusions
Quantifiable Link Between Pre-Event Market Dynamics and Post-Event Belief Shifts: There is a clear, quantifiable relationship where increased market activity and belief volatility prior to an FOMC meeting are strongly associated with more substantial shifts in market belief after the announcement. This suggests that pre-event market 'nervousness' or information asymmetry is a key indicator of subsequent price action.
Predictive Power of Pre-Event Volatility and Volume: Pre-event belief volatility (pre_belief_std) and average trading volume (pre_avg_volume) are demonstrably the most significant features for forecasting market belief shifts. Practitioners should consider these metrics as primary signals when assessing potential market reactions to upcoming FOMC decisions.
High Accuracy in Forecasting Directional Shifts: The ability to predict the direction (positive vs. non-positive) of market belief shifts with an accuracy of 87.5% is robust and actionable. This provides a strong quantitative basis for anticipating the general sentiment response to FOMC announcements.
Validation of Model Robustness: The substantial performance gap between the actual model and the negative control test (87.5% vs. 62.5% accuracy) confirms that the observed patterns are statistically significant and not merely artifacts of random chance. This reinforces the reliability of the identified predictive relationships.