Yield Spreads and Interest Rate MovementsA Bird's Eye View

Let's start with a habit we all share. When the yield curve is steep, we nod and say, "Rates are going up." That intuition comes from a classic idea called the expectations theory of the term structure. In plain English, the long-term yield should be today's short rate plus the market's best guess of future short rates, with a fixed extra bit for bearing maturity risk. So, the spread, which is the long rate minus the short rate, ought to be a window into that future path. A high spread indicates higher future short rates. Since long rates average those future short rates, long yields should move in line with that story. Campbell and Shiller looked at this intuition with a very simple, but very demanding test. If the theory is right, a one-unit change in today's spread should translate one-for-one into the correct change in the long rate over the life of the short bond. That "one-for-one" is the knife edge. They also turn the logic around by taking what the spread claims about future short-rate changes and asking whether the spread itself matches that "perfect foresight" forecast. Again, if the theory holds with a constant term premium, you should see a slope of one. To do this cleanly, you need clean yields. They used McCulloch's zero-coupon U.S. Treasury term structure, collected monthly from nineteen fifty-two through early nineteen eighty-seven. This allows for a comparison of pure discount bonds across a grid that runs from one month out to ten years. Zeroes matter because they strip away coupon complications and compounding conventions, allowing for a direct comparison across maturities. Here's what they observed when they ran those baseline regressions. Across almost any pair of maturities, a steep spread does predict that short rates will rise. That aspect hums along as the expectations theory would suggest. But the long rate? It tends to fall over the life of the short bond, which is exactly the opposite of the naive story. One concrete example involves a three-month versus six-month maturity pair. The slope linking the current spread to the subsequent change in the six-month yield is about minus one point twenty-eight. If you translate that into the earlier Campbell-Shiller-Schoenholtz benchmark, the implied number is roughly minus zero point fourteen. Same spread, two very different behaviors: the short end climbs, while the long end slips. That "wrong sign" isn't a fluke of one maturity. It shows up all over the short end and in most subsamples. The early nineteen fifties are a partial exception, but from the nineteen sixties onward, the pattern is persistent. This creates a paradox that's hard to ignore: the term structure's slope contains real information about how short rates will evolve, yet it systematically misleads you about where long yields are going in the near term. If you rely on the steepness of the curve to bet on long yields rising soon, the data indicate you'll often be leaning the wrong way. Why trust these regression lines? The authors are careful about the econometrics. When predicting an n-period change using a signal that updates every month, your measurement windows overlap. This makes standard errors tricky. Hansen and Hodrick provided one fix for this kind of overlap, while White's robust corrections help with heteroskedasticity. Even with these adjustments, finite samples can make inference too optimistic. Therefore, Campbell and Shiller go further. They build a vector autoregression, which you can think of as a compact summary of how the short rate and the spread move together over time, utilizing four monthly lags. They then use it to generate the path of short-rate changes that the spread is supposed to be indicating. From that vector autoregression, you can back out a "theoretical spread," which you would expect if the expectations theory held, based on the model's forecasts. Comparing it to the actual observed spread, if the theory is correct with a constant term premium, the two should move together. And they do—sort of. The correlation is almost always positive and often quite high. At the very long end, out around ten years, it approaches one. However, the magnitude is off. The theoretical spread is too quiet. Its standard deviation is typically about half of the actual spread's. In other words, markets move the curve's slope around a lot more than a simple average of expected short rates would justify. Take a breath here, because this mismatch matters. The positive correlation indicates that the slope isn't noise; it is genuinely tied to expected short-rate paths. The volatility gap reveals that there's an extra ingredient that causes the spread to swell and shrink beyond what the expectations story, with a fixed term premium, would allow. That's the puzzle in a single glance. Could this all be an artifact of small samples and overlapping data windows? They stress-test that idea with Monte Carlo simulations. They feed artificial shocks through the estimated vector autoregression, recreate the regressions, and ask: if the true world obeyed the expectations theory, how often would we see numbers like those in the data just by chance? There is some finite-sample bias, particularly at the short end where the overlap is worst, but it doesn't overturn the pattern. The short end keeps rejecting the simple theory; the long end is less decisive. Some long-horizon coefficients hover near the one-for-one benchmark, while others do not, but you don't get a clean exoneration by blaming finite samples. That leaves interpretation. The easiest first step is to relax the strictest part of the theory: make the spread a scaled version of what the expectations-with-constant-premium model would predict. You can imagine a "multiplier" bigger than one. The shape of the relationship remains intact—the theoretical and actual spreads move together—but the actual spread swings wider. This adjustment fits two facts at once: the high correlation and the roughly two-to-one volatility ratio. It also allows for the regression slopes to be negative in the long-rate-change test, because rescaling the signal can flip the sign once you translate between the two representations of the theory. A close cousin of that story involves time-varying term premia. If investors demand a higher premium exactly when they also expect higher future short rates, the two components of the spread—expectations and premia—will move in the same direction. Consequently, the spread rises by more than the expectations component alone. Then, when short rates climb as anticipated, the long rate doesn't need to rise in lockstep in the near term, because part of the initial steepness was due to that premium. This provides one way to achieve "short rates up, long yields down" over the life of the short bond without disregarding information. Could it just be measurement noise that's conveniently uncorrelated? Campbell and Shiller test that orthogonality idea with instrumental variables. If the error in the spread were white noise, completely uncorrelated with past information, then using lagged spreads or moving averages of the short rate as instruments should yield a slope near one when you regress the observed spread on the theoretical one. They do not observe this. The instrumental variable slopes are consistently negative and often larger in magnitude than the ordinary least squares versions. When they substitute instruments like a twelve-month moving average of the short rate or a sixty-month one, the basic message remains unchanged. This leans against a "just noise" explanation and supports the idea of a structured deviation—multipliers, premia, or some combination. There's a stylistic way to express the same idea that dates back to earlier work by Campbell and Shiller: overreaction versus underreaction. Consider the long rate as a weighted average of current and future short rates. If it underreacts to the current short rate and overweights more distant expectations, you can end up with an apparent "too steep" spread today, which unwinds as the near-term long rate drifts down and the short rate climbs. The distributed-lag weights in their estimates often sum to about one, as the theory would require, but they are in the wrong places in time. That misplacement is sufficient to flip the sign in the near term without violating the long-run arithmetic. One subtlety that's easy to overlook is that horizon matters. The predictive grip of the spread on short-rate changes generally strengthens as you look further out, echoing findings by Fama and Bliss. At very short horizons, around nine to twelve months, it wobbles, then firms up again toward multi-year horizons. Concurrently, the long-rate result is most stark in the near term, where a steep curve is followed by a drift down in the longer yield. By the time you reach tests keyed to multi-year changes in long rates, the evidence becomes more mixed; some specifications get close to the one-for-one benchmark while others still show that negative slope. The vector autoregression view ties this together: as you extend the horizon, the correlation between the actual and theoretical spreads rises, often approaching one by ten years, but the volatility gap persists. If you're trying to apply this knowledge, here's the takeaway. Spreads do a reasonable job telling you about the path of short rates. However, they are not a great signal for imminent jumps in long yields. In fact, a steep curve often correlates with downward pressure on long yields over the life of the short bond. Therefore, for a bond investor relying on the slope to time long-end moves, this serves as a cautionary sign. For a macro forecaster, it's a reminder to separate expectations about policy paths from time-varying premia that can thicken or thin the curve. Methodologically, this offers a lesson in humility. Overlapping horizons can make confident t-statistics appear stronger than they truly are. Robust corrections—Hansen and Hodrick for overlap, Newey and West for general dependence—are helpful, but simulations that mimic your exact sample can mean the difference between a fragile claim and a sturdy one. Campbell and Shiller's Monte Carlo runs serve as a template for that discipline. Where does this leave the expectations theory? It is not dead, nor triumphant, but trimmed. The core insight—that the curve's slope is a window onto future short rates—survives. The strict version, which posits a time-invariant term premium and one-for-one regression slopes, does not fit the postwar U.S. data, especially at the short end. A modest generalization—allowing the premium to vary with the cycle and permitting the spread to be a scaled-up version of its textbook self—brings the facts back into view without resorting to magic. If you want to advance this work, two paths look especially promising. One is to incorporate direct measures of risk compensation, such as swap spreads, survey expectations, and macro state variables. This would reveal how much of the "too steep" spread they absorb when the curve fattens. The other is to continue investigating dynamics: richer vector autoregressions, or models that allow today's shocks to shift not just the level of expected short rates but their uncertainty, might help explain why the correlation increases with horizon while the variance mismatch stubbornly remains. But that's for another day. For now, the simplest reading is the most useful: when the curve is steep, expect short rates to rise—and don't be surprised if long yields sag before they catch up.