Every recession looks obvious in hindsight. The housing bubble, the dot-com mania, the oil shocks: after the fact, the warning signs seem to have been flashing everywhere. Yet in real time, forecasters have a notoriously weak record of calling turning points. Professional forecasters have rarely predicted a recession even a few months before it began.

The problem is not a shortage of data. It is the opposite. Thousands of economic series rise and fall every month, and most of them fluctuate for reasons unrelated to the business cycle. The real challenge is separating the variables that genuinely move ahead of the economy from those that simply move.

This puzzle has occupied economists since Wesley Mitchell and Arthur Burns began cataloguing cyclical indicators at the National Bureau of Economic Research in the 1930s. Their question still stands: which signals carry real information about the future, and why? The answer depends less on any single indicator than on understanding the economic mechanism behind each one, and on recognising when that mechanism stops working.

The logic of leading indicators rests on a simple observation: economic decisions take time to execute. A firm that orders new machinery today will produce with it months from now. A developer who obtains a building permit will break ground later. These commitments are recorded before the activity they represent, and that timing is the source of their predictive value.

New orders for nondefense capital goods, the ISM manufacturing new orders index, and residential building permits have long been the workhorses of this approach. Housing has a particularly strong record. Residential investment peaked in early 2006, nearly two years before the Great Recession began, and permits fell sharply well ahead of the broader downturn. Economist Edward Leamer's well-known argument that “housing is the business cycle” reflects this pattern across most postwar US recessions.

Yet the classic indicators have real weaknesses. Their lead times vary widely, from a few months to well over a year, which makes timing difficult. They are revised, sometimes substantially, after initial release. And the structure of the economy has shifted. Manufacturing now accounts for a much smaller share of output and employment than when many of these indicators were first selected, so factory orders capture a narrower slice of activity than they once did.

There is also the persistent problem of false alarms. Manufacturing indicators contracted in 2015 and 2016 amid an energy-sector slump and a strong dollar, yet the broader economy kept growing. A slowdown in one sector does not necessarily spread. Whether it does depends on how weakness transmits through employment, household income, and credit conditions. The indicator itself cannot reveal whether that transmission is happening. It only shows that one link in the chain has weakened.

Takeaway

A leading indicator is only as reliable as the causal chain linking it to the broader economy. When the economy's structure changes, that chain can weaken, and so does the signal.

Financial markets offer something real-economy data cannot: forward-looking prices, updated continuously and never revised. Asset prices aggregate the expectations of millions of participants with money at stake. In principle, this should make them powerful forecasting tools. In practice, their record varies enormously by market.

Stock prices are the weakest of the group. Paul Samuelson's 1966 quip that the stock market had predicted nine of the last five recessions remains apt. Equities respond to discount rates, risk appetite, and sentiment as well as to expected earnings. The 1987 crash, when the Dow fell more than 22 percent in a single day, was followed by years of continued expansion.

The yield curve has a far better record. The spread between ten-year Treasury yields and three-month bill rates turned negative ahead of every US recession since the late 1960s, typically six to twenty-four months in advance, a relationship documented by Arturo Estrella and Frederic Mishkin. The mechanism is intuitive: tight monetary policy lifts short rates while markets expect future cuts as growth slows. However, the inversion that began in 2022 persisted unusually long without a recession arriving on schedule, a reminder that even strong historical relationships can bend.

Credit spreads, the gap between corporate bond yields and Treasury yields, measure lenders' perception of default risk and their willingness to bear it. Research by Simon Gilchrist and Egon Zakrajšek found that the “excess bond premium,” the portion of spreads not explained by expected defaults, is particularly informative. When it rises, credit supply is tightening, which directly reduces investment and hiring. Spreads often widen later than the yield curve inverts, making them useful confirmation.

Takeaway

Financial markets predict recessions best when a signal reflects a concrete mechanism, such as tight monetary policy or tight credit, rather than shifts in mood alone.

If every individual indicator suffers from false alarms and shifting lead times, the natural response is to combine them. The Conference Board's Leading Economic Index blends ten components spanning manufacturing orders, building permits, jobless claims, stock prices, credit conditions, the interest rate spread, and consumer expectations. The idea is that noise specific to each series partially cancels out, leaving the common cyclical signal more visible.

Composite indices are not immune to error. The LEI declined persistently through 2022 and 2023, a pattern that historically signalled recession, yet output and employment kept growing. Many components were tied to manufacturing and interest rates, the areas hit hardest by monetary tightening, while a services-heavy economy supported by household savings and strong labour demand proved resilient. Combining correlated indicators provides less diversification than it appears to.

Statistical models attempt to go further. Probit models, such as the one published by the Federal Reserve Bank of New York, translate the yield curve and other variables into an explicit recession probability. Dynamic factor models, developed by James Stock and Mark Watson, extract common movements from hundreds of series at once. Machine learning methods add the ability to detect nonlinear patterns. Each approach replaces binary warnings with probabilities that can be weighed and updated.

A fundamental limit remains. The United States has experienced only about a dozen recessions since the Second World War, so every model is estimated on very few observations. Each downturn also had a different trigger: oil shocks, monetary tightening, financial crises, a pandemic. Milton Friedman's warning about the “long and variable lags” of monetary policy applies here too. Even when a forecaster identifies the right mechanism, its timing remains uncertain.

Takeaway

Combining indicators reduces noise but cannot create information about shocks that have never occurred. The most useful forecast is a well-calibrated probability, not a confident prediction.

The search for reliable recession predictors reveals a consistent lesson. Indicators work when they capture a genuine transmission mechanism. Orders and permits reflect committed spending. The yield curve reflects monetary policy and expectations about it. Credit spreads reflect the availability of financing.

Signals weaken when the economy's structure changes or when shocks arrive from outside the patterns the indicators were built on. Recent years, with prolonged inversions and falling composite indices alongside continued growth, have tested many long-standing relationships.

For anyone tracking the cycle, the practical approach is to ask not only what an indicator says, but why it should matter, and whether that reason still holds today.