Markov Regime Switching Risk-On/Risk-Off Model for Treasuries
Data from FRED
as of 30 Sep 2025.
In this
post we explain the use of Markov Regimes Switching Risk-On/Risk-Off Model (MRS)
for Treasuries. The characteristics of Debt
securities as an asset class are very different from Equities- the issuers and
holders of Debt securities, the way they are traded, the role of liquidity, the
importance of persistence and duration in yields- all make risk-on/risk-of
models useful. Markov-type models with their wide scope for modifications e.g.
Hidden Markov, Semi Hidden Markov; and enhancements are very well suited to
this task. In this post we will show the current state of risk for different
tenors of Treasuries. We also have a
supplementary paragraph to show why Money Market Funds are low -risk.
Methodology
We model regime shifts in Treasuries using price-based,
weekly slope signals rather than raw yields. Using prices (and their derived yield
differences) ensures direct linkage to investor gains and losses, while
avoiding the unit and scaling inconsistencies that can arise when mixing yield
levels across tenors and sampling frequencies. We aggregate to weekly (Fri)
observations for stability and to minimize microstructure artifacts, holiday
gaps, and false flips that daily data can induce in regime identification.
Instead of the level of a single tenor’s yield, we use the slope (e.g.,
30Y–10Y, 10Y–2Y, 6M–3M) because slopes encapsulate the market’s relative
expectations across horizons—term premium, policy path, and growth/inflation
narratives—yielding a more robust, stationary-like signal that is less
regime-unstable than raw levels. This is particularly important when comparing
money-market vs. long-duration segments, where absolute yield moves have very different
volatilities and are not directly comparable.
Think of the Treasury market as a curve rather than isolated points.
For example, when the front end (2-year) moves one way and the long end (30-year) moves the other, the shape of the curve is telling us something about growth-inflation expectations that a single-tenor model can miss. By fitting each tenor separately, we ignore that cross-tenor “slope” signal, so the model may latch on to one stable state and hardly budge. Bringing slope data into the model gives it richer information. Therefore, we chart the 30y-10y, 10y-2y and 6m-3m risk probabilities.
We estimate regime probabilities via a robust two-state
hidden Markov model, which we refer to as a t-like Sticky HMM (t-SHMM), with
some resemblance to Student t distributions that can accommodate heavy tails and
skewness . Practically, we achieve t-style robustness with light Winsorization
of Z-scored weekly slopes and enforce persistence (“stickiness”) through high
self-transition priors; this guards against outliers and spurious whipsaws
while preserving responsiveness to genuine regime shifts. The lower-variance
state is interpreted as Risk-On for slopes, and we report the smoothed
probability of that state as our p_mean. Thresholds (e.g., 55–65%) and short
persistence filters ( 3 weeks) define actionable windows while keeping the
signal intuitive and visually consistent across tenors. Note: The charts use a
3-week persistence overlay (shaded) above a 65% threshold to highlight
sustained risk-on periods, and mitigate flip-flops.
To initiate the reader on the topic, we first show a chart
of the yields of long, medium and short-term Treasuries (T) below for you to visualize
how the 2y,10y and 30y are behaving.
The next few self-explanatory charts show the current risk level of different tenors of T.
Comments
Post a Comment