← back to writing

Trading a rise in Treasury term premium

Independent strategy framework. The historical discussion uses published Federal Reserve analysis of 2023. Yield decompositions, trade sizes and scenarios below are illustrative.

For a persistent increase in long-end risk compensation, my preferred candidate would be a DV01-matched 2s10s steepener: long a two-year Treasury and short a ten-year Treasury. The condition is that the long end faces greater repricing than the front end. I would want evidence from several term-premium estimates and policy expectations before treating a selloff that way.

I find this an appealing problem because the observable part is easy: a Treasury yield goes up. Explaining the increase requires a view about something much harder to observe: how much investors expect short rates to average, and how much compensation they require for committing to a longer bond.

That distinction can change which part of the curve I want to own.

The same yield increase can contain different messages

An investor comparing a long bond with a sequence of shorter investments faces uncertainty about future rates and returns. Economists use the term premium to describe the extra compensation embedded in the longer yield, beyond the expected path of short rates.

A useful schematic, using a model-consistent zero-coupon yield, is:

yt(n)≈1n∑j=0n−1Et[rt+j]+TPt(n).y_t^{(n)}\approx\frac{1}{n}\sum_{j=0}^{n-1}E_t[r_{t+j}] +TP_t^{(n)}.

Here, the expected short rates refer to equal periods across the bond's life. The expression summarizes an economic decomposition; estimating its components requires a model and consistent rate conventions.

Suppose an illustrative ten-year yield begins at 4.00%, comprising a 3.50% expected average short rate and a 0.50% term premium. It then rises to 4.50%. Two possible decompositions are:

Illustrative component Expected-rate-led move Premium-led move
Expected average short rate 4.00% 3.50%
Term premium 0.50% 1.00%
Fitted ten-year yield 4.50% 4.50%

The observed increase is 50 bp in both cases. The first interpretation points toward a higher expected rate path. The second points toward a change in the compensation investors require for holding duration. An outright short ten-year position could benefit from either increase, but a curve hedge can behave differently depending on how the two-year yield responds.

The New York Fed's introduction to the Adrian–Crump–Moench model explains this decomposition and its use of zero-coupon yield data. I would preserve that convention when reconciling its fitted yield and components. Subtracting a model term premium from an unrelated par yield can create a residual that looks meaningful without matching the model's expected-rate component.

Autumn 2023 made the distinction consequential

In his November 16, 2023 remarks, Roberto Perli reported that ten-year Treasury yields had increased by more than 100 bp from the beginning of July to the end of October. The models monitored by the New York Fed produced different estimates, but on average attributed most of that change to term premium. Survey-based decompositions supported a similar broad interpretation.

That is stronger evidence than looking at a steepening curve and assigning the move a label. It is also narrower than claiming that the entire selloff reflected one risk premium, or that the estimates were observable facts.

The information available during the move was less tidy. In an October 18 speech, Governor Christopher Waller described a roughly 90 bp increase in the ten-year yield since July, alongside a much smaller increase at shorter maturities. He discussed stronger economic data, deficits and issuance among the possible influences.

I like reading those accounts together. The October speech shows the questions policymakers faced while the episode was unfolding. The November account offers a later decomposition. A retrospective can use both, provided it keeps their timing visible.

A model estimate needs some company

The Federal Reserve's study of term-premium robustness compares ACM with the Kim–Wright model. It finds that survey inputs and estimation samples help explain material differences between their estimates. Even the sign and volatility of the premium can be less robust than a chart suggests.

That changes how I would use the output. I would compare the direction and persistence of the change across models, examine survey expectations, and look at short-rate pricing for corroboration. Futures and overnight-indexed swaps are useful cross-checks, but their prices also contain risk premiums; they do not provide a complete, assumption-free forecast of policy.

A stable current Fed target would not settle the question, either. Investors can revise expectations for future easing or the longer-run neutral rate while today's target remains unchanged. A rise in distant forward rates can reflect those revisions as well as higher term premium.

For a research tracker, I would retain each model's fitted yield, expected-rate component and premium, together with its observation and retrieval dates. A backtest would need estimates available at the time, or a recursive reconstruction. Today's revised history could give a past decision information the analyst did not have.

The premium can also be negative in a model. The New York Fed documents such estimates in its historical analysis. The word “premium” does not impose a positive floor, and a move back above zero does not, by itself, establish that long bonds are cheap.

Keep real yields and term premium on the right axes

The TIPS decomposition in this portfolio separates nominal yields into real yields and inflation compensation. The decomposition here asks a different question: expected short rates versus compensation for holding a longer bond.

Those are overlapping ways to examine the same nominal yield. Adding a nominal term premium to real yield plus breakeven inflation would double-count part of the explanation. Real yields themselves can include a real term premium, and inflation compensation can contain risk and liquidity premiums.

For 2023, the evidence that real yields drove much of the nominal selloff therefore sits comfortably beside the evidence for higher term premium. The two statements address different dimensions of the repricing. Neither, alone, identifies a change in expected inflation.

Translate the thesis into a cash-curve position

The trade candidate is a long two-year Treasury against a short ten-year, with matched dollar sensitivity to a one-basis-point yield move, or DV01. It benefits if the ten-year yield rises relative to the two-year yield. That is the exposure I would seek if persistent long-end risk compensation were the central view.

For illustration, assume a long two-year market value of USD 5 million with modified duration 1.8, and a short ten-year market value of USD 1.2 million with modified duration 7.5. Both legs have DV01 of USD 900 per bp:

DV01≈V×Dmod×0.0001.\mathrm{DV01}\approx V\times D_{\mathrm{mod}}\times0.0001.

With yield changes in basis points, first-order price P&L is:

Πprice≈900(Δy10−Δy2).\Pi_{\mathrm{price}}\approx900(\Delta y_{10}-\Delta y_2).

Assume coupon income, roll, financing and execution together contribute a net drag of USD 1,500 over a one-month horizon. This is a hypothetical allowance, not a financing quote or an estimate from the historical episode.

Hypothetical scenario 2-year move 10-year move Net P&L
Long-end-led selloff +5 bp +25 bp +USD 16,500
Front-end-led selloff +25 bp +10 bp −USD 15,000
Growth-led rally −35 bp −20 bp +USD 12,000
Parallel selloff +20 bp +20 bp −USD 1,500

The growth-rally row is instructive. The position can make money while yields fall if the two-year rallies more. It trades the slope; it does not isolate term premium. Conversely, a stronger-than-expected inflation or labor-market report could hurt if investors reprice the front end more than the ten-year.

Matching DV01 removes a small parallel shift to first order. It leaves curve, convexity, liquidity and changing-sensitivity risks. The USD 3.8 million difference between the legs' market values also needs a funding plan. Actual implementation requires security-level prices, accrued interest and repo terms, including availability of the short.

Persistence is the part I would need to get right

An entry would require evidence that long-end pressure persists while near-term policy expectations remain more stable. I would monitor supply absorption, rate volatility, model changes and the behavior of the curve around macro releases. A single high premium estimate would leave too much unresolved.

Against the entry slope, an illustrative plan could target 15 bp of steepening, trigger a risk review after 10 bp of flattening, and close after one month if the catalyst fails to develop. Those levels describe a research scenario. They need calibration, and an exit instruction cannot guarantee execution at the intended price.

The thesis weakens if models attribute more of the move to expected short rates, uncertainty recedes, or demand absorbs duration without further concession. Perli's discussion makes a related point: long-run sources of uncertainty could sustain higher premiums, while short-lived concerns could unwind. A high premium is not a timetable for its own reversal.

My conclusion is conditional: use a steepener to express expected relative pressure on the long end, with several sources supporting the diagnosis and funding included in the return. The question I want to investigate next is whether agreement across models improves the trade signal, or mainly improves the confidence of the explanation. Those are different tests, and I would want to see both results.

Related frameworks: two ways the Treasury curve can steepen, Treasury carry after the funding bill, and trading an auction concession.