Is This How America Ends?

THE HILL REPORT

Recessions Are Not Forecast in Advance

Connor Hill · InsightfulWord · August 30, 2026

The Congressional Budget Office publishes an annual assessment of how accurate its own economic forecasts have been. It has done so for decades, it compares itself against the Administration, the Blue Chip consensus and the Survey of Professional Forecasters, and the document is free.

That is a genuinely unusual institutional practice. Most organizations issuing forecasts do not publish a scorecard of their previous ones, and the absence of such a scorecard is the single most informative fact about most forecasting.

The record itself is instructive. Across forecasts going back to 1976, the mean absolute error of two-year-ahead projections of real output growth is about 0.9 percentage points. For consumer price inflation it is about 0.7 points, and for the three-month Treasury bill rate about 1.0 points.

Those are respectable numbers for a difficult task and they are large relative to the quantities being forecast. An error of 0.9 points on a growth rate that typically runs between one and three percent is not a small error.

The more consequential finding concerns turning points. The report states plainly that forecasts made just before a recession tend to be overly optimistic, because downturns cannot be accurately predicted from the information available, and that accuracy across all four sources is lower in periods overlapping a recession.

That is the professional consensus stated by the institution with the best-documented record: the specific event everyone most wants forecast is the one that is not forecastable from available data. It has been the pattern across every downturn for which records exist, and it applies to official forecasters, private forecasters and market-implied measures alike.

What follows is what the error record shows in detail, why turning points are structurally the hard part, the difference between a scenario and a prediction, what a falsifiable warning would contain, and which series move before anyone announces anything.

What the Error Record Shows

The value of a published scorecard is that it converts a general impression into measurable quantities, and several of the measurements are counterintuitive.

Accuracy does not degrade much between the two-year and the five-year horizon. That sounds like good news and is not: it indicates that the two-year forecast is not capturing much beyond the long-run averages that anchor the five-year one.

Forecasts are on average slightly too optimistic, by small amounts. The bias is modest and it is persistent, which is itself informative about the institutional pressures on forecasting.

The comparison across sources is close. The official forecaster is generally about as accurate as the private consensus, occasionally better, occasionally worse. There is no source in this comparison with a demonstrably superior record, which is what one would expect if the residual error reflects genuine uncertainty rather than differences in skill.

And errors cluster. Most of the total error across the whole record is concentrated in a small number of periods — the ones containing turning points — with long stretches of comparatively accurate forecasting in between.

That clustering explains why forecasting looks better than it is. A forecaster reviewing their own record over a decade without a downturn sees mostly small errors and concludes the method works. The method works in the conditions where the outcome was never in doubt, and fails in the conditions that determine whether the forecast was worth having.

Why Turning Points Are the Hard Part

There is a structural reason, and it is not a failure of effort or of modeling.

Economic forecasts are built substantially on the persistence of current conditions. Most of the time that works, because the economy is in fact persistent — growth this quarter is the best single predictor of growth next quarter.

A turning point is by definition a break in persistence. Forecasting one requires identifying, in advance, that the relationships holding until now are about to stop holding — which the data cannot show, because the data describe the period in which they held.

Compounding this, downturns are frequently triggered by discrete events: a policy shock, a financial failure, a geopolitical rupture, a pandemic. Those are not smoothly evolving quantities that a model can extrapolate. They are events, and their timing is not encoded in prior observations.

The consequence is that the honest form of a downturn forecast is probabilistic and unimpressive: a statement about elevated risk over a period, not a date. Forecasters who issue such statements are ignored, and forecasters who issue confident dates are quoted, which shapes the incentives of everyone in the business.

The Difference Between a Scenario and a Prediction

The distinction is the most useful analytical tool in this area and it is systematically blurred.

A scenario describes a coherent chain of events and what would follow if it occurred. It is a valuable exercise, used by central banks in stress testing and by companies in planning, and it makes no claim about likelihood.

A prediction assigns probability. It says this will happen, or is likely to, within a period.

The two are frequently presented in the same document with the same voice, and a reader cannot tell them apart unless the writer distinguishes them. A scenario written vividly reads exactly like a prediction, and its author can later claim to have described events accurately without ever having said they would occur.

The test is simple. A prediction can be wrong. If no outcome would falsify a statement, it is a scenario, and scenarios are useful for preparation and useless for positioning.

There is a further device worth recognizing because it appears constantly. A claim framed as a permanent change to conditions — a reset, a new era, a transformation — cannot be falsified by any single observation, since any period can be described as an early stage of it. Statements of that shape are unusually durable in the marketplace for exactly that reason, and their durability is a property of the framing rather than evidence of their accuracy.

What a Falsifiable Warning Contains

Four elements convert an alarming statement into one that can be evaluated afterwards.

A specific claim about an observable quantity — an index level, an unemployment rate, an inflation reading — rather than a description of conditions.

A date or a window, stated in advance.

A stated probability, or at least a stated confidence, so that being wrong once is distinguishable from being wrong systematically.

And a record: what the same source predicted previously, with dates, and what happened.

Almost nothing in the promotional forecasting genre contains all four, and the absence is not accidental. A statement with all four can be checked, and being checked is the risk the format is designed to avoid.

Where a source does publish such a record, it is worth reading carefully rather than dismissively. Several established forecasters do, and their records generally show what the official scorecard shows: reasonable accuracy in ordinary conditions, misses at turning points, and no persistent edge.

Which Series Move Before Anything Is Announced

If forecasts are weak, the alternative is to watch the data directly, and a handful of series are published frequently, revised transparently, and lead the aggregates.

Initial claims for unemployment insurance, published weekly, are among the most timely labor market indicators available and are not subject to the long revision cycles that affect other series.

The yield curve — the spread between long and short government rates — has an unusually good historical record ahead of downturns, and the reason it works is debated. Its lead time has been long and variable, which limits its usefulness for timing.

Job openings and quits from the labor turnover survey describe the demand side of the labor market with more sensitivity than the headline employment figure.

Credit conditions, from the senior loan officer survey, capture bank willingness to lend, which is one of the more direct transmission channels into activity.

None of these predicts a date. Together they describe the current state with less lag than the headline figures, which is a more modest and considerably more attainable objective than forecasting, and every one of them is free.