A government agency publishes a number. A prediction market publishes a price. They encode fundamentally different kinds of knowledge — and the gap between them reveals how we think about uncertainty.
The Cleveland Federal Reserve publishes a number every business day at 10:00 AM Eastern. As of last week, that number was 0.24%. It's their estimate of what the next Consumer Price Index report will show for month-over-month inflation. They call it a nowcast — a forecast of the present, filling in the number before the Bureau of Labor Statistics makes it official.
On the same day, on a prediction market called Kalshi, the contract "Will CPI exceed 0.2% month-over-month?" was trading at 53 cents.
The Cleveland Fed says 0.24%. The market says 53%.
These are not contradictory. They are two fundamentally different representations of the same underlying uncertainty. And understanding the gap between them — the gap between a point and a distribution — reveals something worth knowing about how we encode what we believe.
What a Point Estimate Knows
A point estimate is the center of your uncertainty. The Cleveland Fed's 0.24% is their best single guess at what inflation will be. It's produced by a model that ingests ten data series: monthly CPI components, PCE inflation readings, weekly gasoline prices, and daily crude oil prices. It updates every business day as new data arrives.
The model is good. Historically, in the recent era from 2024 through early 2026, its standard deviation of error has been about 0.106 percentage points. That means roughly two-thirds of the time, the actual CPI prints within 0.106 points of where the nowcast said it would. That's precise.
But here's what the point estimate doesn't tell you: the shape of the uncertainty around it. Is the error symmetric? How fat are the tails? What happens when the nowcast is wrong — does it tend to miss high or low? A point estimate is the answer to one question: what's the single most likely value? It says nothing about the probability of being above or below any particular threshold.
And thresholds are exactly what prediction markets trade.
What a Price Knows
A prediction market contract at 53 cents is a statement about probability: the collective market estimate is that there's a 53% chance the event occurs. But Kalshi doesn't trade a single contract for CPI. It trades a series of contracts at different thresholds: will CPI exceed 0.0%? 0.1%? 0.2%? 0.3%? 0.4%?
Each threshold has its own price. Together, these prices encode an implied probability distribution over outcomes. When the CPI > 0.1% contract trades at 90 cents and CPI > 0.2% trades at 53 cents and CPI > 0.3% trades at 12 cents, the market is telling you it believes there's roughly a 37% chance CPI lands between 0.1% and 0.2%, a 41% chance between 0.2% and 0.3%, and a 12% chance above 0.3%.
This is richer than a point estimate. It's a full description of where the market thinks the outcome might land, with probabilities attached to each region. It encodes the center and the spread and the asymmetry.
The prediction market doesn't know more facts than the Cleveland Fed. It knows the same facts, processed through a different representation — one that explicitly accounts for the shape of uncertainty rather than collapsing it to a single number.
The Conversion
The interesting exercise is converting between these representations to see what each one implies about the other.
The Cleveland Fed gives you the center: 0.24%. If you wrap a distribution around it — say, a normal distribution with a standard deviation equal to the nowcast's historical error (0.106 percentage points in the recent era) — you can compute the probability of exceeding any threshold.
P(CPI > 0.2%) with a normal distribution centered on 0.24% and σ = 0.106%: about 64%.
The market says 53%.
That's an eleven-point gap. Your first instinct might be to call this an opportunity — the market is wrong, you can buy at 53 cents what's really worth 64 cents. But stop. The market participants can read the Cleveland Fed's website too. They've seen the same 0.24%. They've done the same arithmetic, or something like it. Twenty thousand contracts have traded on this, with tight one-cent spreads.
The gap is not evidence of market inefficiency. It's evidence that the market has a different view of the distribution than a naive Gaussian conversion produces. And the market may well be right.
Maybe the market is pricing in model risk — the possibility that the Cleveland Fed's nowcast is less reliable in the current regime than its recent track record suggests. Maybe it's accounting for the fat left tails in the historical error distribution (excess kurtosis of +3.7 over the full sample), which a Gaussian ignores. Maybe it reflects information in other economic indicators that the nowcast hasn't incorporated yet.
The exercise isn't interesting because it reveals a trading edge. It's interesting because it forces you to articulate your assumptions. Which era do you calibrate to? The full sample from 2013 to 2026 gives σ = 0.149% — wider, because it includes the COVID period when forecast errors were enormous. The recent era gives 0.106% — tighter, because inflation has been stable. The full sample is more data. The recent era is more relevant. You have to choose. Do you trust the Gaussian assumption? The historical errors have negative skew (−1.0 over the full sample) and fat tails. In the recent era, the skewness reverses: mildly positive at +0.5.
Each of these choices — calibration era, distributional assumption, tail treatment — shifts the implied probability. The conversion doesn't give you the answer. It gives you a language for comparing your beliefs to the market's beliefs, and it forces you to notice where you're making assumptions you might not have articulated otherwise.
What the Market Knows That You Don't
It's tempting to look at a gap between your model and a market price and conclude the market is wrong. The history of quantitative finance is littered with this mistake.
A prediction market with active volume and tight spreads is not a crowd of uninformed guessers. It's a mechanism that aggregates the views of everyone willing to put money behind their beliefs. Some of those participants have built exactly the distribution conversion described above. Some have better models. Some have information sources — private economic data, real-time spending metrics, alternative inflation indicators — that the Cleveland Fed's ten-input model doesn't include.
When the market diverges from a public data source, the default assumption should be that the market has already incorporated that data source and has reasons for its price. The divergence represents the market's opinion about the limitations of the public estimate, not evidence that the market hasn't seen it.
This is a useful corrective to a natural human bias: we tend to overweight information we've personally analyzed. The Cleveland Fed number feels solid because you can trace its inputs and calculate its historical accuracy. But the market price also encodes analysis — it just encodes everyone's analysis, weighted by conviction (measured in dollars).
Real informational advantage requires something the market hasn't priced: proprietary data, a genuinely superior model, structural market inefficiency (thin liquidity in a particular contract), or a timing advantage. Comparing a public nowcast to a public market price and calling the difference an edge is just assuming the other side is uninformed — which, in a liquid market, is the least likely explanation.
The Pattern Generalizes
CPI is not the only place where point estimates and distributions coexist.
The Atlanta Federal Reserve publishes GDPNow — a real-time estimate of quarter-over-quarter GDP growth, updated as new economic data arrives. Like the Cleveland Fed's inflation nowcast, it's a point estimate. And like CPI, Kalshi trades GDP threshold contracts. The same conversion applies: GDPNow says 3.1%, so what's the probability GDP exceeds 3.5%? The answer depends on the historical error distribution, which varies dramatically with time horizon — wider when the quarter is young, tighter as data accumulates.
Federal funds rate predictions follow the same structure. Weather forecasts, election models, earnings estimates — anywhere a point prediction meets a binary contract, the conversion problem appears.
The exercise is always the same: take the point estimate seriously as a center, wrap a calibrated distribution around it, and see what it implies about threshold probabilities. Not to find an edge — the market has almost certainly done this already — but to understand the shape of your own uncertainty. The conversion is a tool for self-knowledge, not market-beating.
The Deeper Point
I find myself thinking about what it means that two representations of the same reality — a point estimate and a distribution — can coexist, and that the space between them is where the interesting questions live.
A point estimate is a statement of confidence: this is what will happen. A distribution is a statement of humility: these are the things that could happen, and here is how likely each one is. The point is inside the distribution. It's the peak, the mode, the center of mass. But the distribution contains infinitely more information, because it also describes everything the point doesn't know about itself.
This distinction isn't limited to prediction markets. Every time someone gives you a number without uncertainty bounds, they're giving you a point without its distribution. A startup's revenue forecast. A project's completion date. A medical test result. The number feels precise. The precision is an artifact of the representation, not of the underlying reality.
The discipline of wrapping a distribution around a point estimate — asking not just what's your best guess? but what does the range of possible outcomes look like? — is useful far beyond any particular market. It's a habit of mind. It's the habit of not trusting a single number to carry more information than it actually contains.
The Cleveland Fed publishes 0.24%. That's what they know. The shape of the distribution around 0.24% — that's what they don't know they know. And the gap between what your distribution says and what someone else's distribution says — that's not an edge. That's a conversation. One that's worth having with enough precision to notice where you disagree, and enough humility to wonder if you're the one who's wrong.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)