What this machine-learning model with 65% accuracy says is coming next for the 10-year Treasury
A decade ago, forecasting the 10-year Treasury was mostly an exercise in macro storytelling: inflation, Fed policy, growth, and a fair bit of intuition. Today, a growing share of that story can be quantified. An ensemble machine-learning model that has posted a 65% out‑of‑sample hit rate on the direction of the 10-year yield over one- to three‑month horizons points to a familiar but nuanced message: the next chapter is likely a drift lower in yields, not a collapse, with wide error bands and asymmetric risks.
The headline call
– Base case (about 60% probability): The 10-year yield edges lower over the next 1–3 months, roughly 15–35 basis points, as growth momentum cools at the margin, inflation trend decelerates, and rate volatility normalizes. The move is expected to be orderly, with occasional squeezes higher on data surprises.
– Upside risk (about 30%): Yields back up by 25–40 bps if services inflation re-firms or term premia rebuild on renewed supply concerns and global-policy synchronization (for example, heavier duration supply or reduced foreign demand).
– Tail risks (about 10% total, two-sided): A sharp downside air‑pocket in yields on a growth scare or financial accident, or a jump higher on a hawkish policy pivot triggered by sticky inflation.
What the model is “seeing”
Interpretable feature methods point to several drivers that have historically preceded modest declines in the 10-year:
– Macro momentum slipping at the margin: Softening in PMIs and hiring intentions relative to expectations tend to lead lower term yields over 1–2 months.
– Cooling inflation impulse: The three- to six‑month run‑rate of core inflation matters more than year‑over‑year prints; deceleration in that shorter window has preceded yield declines in the model’s training history.
– Easier rate volatility: When the MOVE index trends lower from elevated levels, the model often tilts to lower yields as risk premia compress and demand for duration returns.
– Positioning and flow support: Short‑duration positioning that has already been reduced, plus steady liability-driven demand, historically limit upside in yields.
– Term‑premium dynamics: Curve normalization from deep inversion often coincides with a mild decline in long rates as policy‑rate uncertainty fades and growth expectations re-anchor.
How the model works (and what 65% really means)
– Target and horizon: Direction and magnitude of the 10-year yield change over 4 and 12 weeks. The directional hit rate quoted (65%) refers to the 4‑week horizon.
– Architecture: An ensemble of gradient‑boosted trees, a regularized linear model for stability, and a lightweight sequence model to capture time dependencies. The ensemble is probability‑calibrated.
– Inputs: Macro and market features including growth and inflation surprises, breakevens and real yields, Fed‑dated OIS, curve shape, term‑premium proxies, rate volatility, futures positioning, Treasury supply indicators, cross‑market relative value (FX‑hedged spreads), and price‑based signals like momentum and carry.
– Validation: Purged, embargoed k‑fold cross‑validation to avoid look‑ahead bias; out‑of‑sample period spans multiple regimes. Performance is benchmarked against simple heuristics (carry/roll, trend following) and shows incremental, not absolute, edge.
– Interpretation: SHAP-style attributions typically flag growth/inflation surprises, rate vol, and term‑premium proxies as the top three drivers.
– Caveat: A 65% hit rate on direction is meaningful but modest; misses can cluster around regime shifts (for example, abrupt policy repricing or supply shocks). Transaction costs, slippage, and timing matter as much as the signal.
What could invalidate the call
– An upside inflation surprise that re-accelerates the three- to six‑month core trend.
– A hawkish repricing of the policy path, especially if the front‑end sells off and drags the long end.
– A jump in term premium from heavier‑than‑expected long‑duration supply or waning foreign demand.
– A renewed spike in rate volatility, which tends to widen risk premia and push yields higher.
– Global spillovers (for instance, a rapid shift in overseas policy or hedging costs) that make U.S. duration relatively less attractive.
Implementation ideas and risk management
– Expression: Long 10-year duration via Treasury futures or 7–10y bond exposure; for defined risk, call spreads on TY futures or receiver swaptions in the 10y sector.
– Timeframe: Scale in over days, not minutes; the signal is statistical and degrades if chased into data prints.
– Sizing: Respect the signal’s modest edge; size for defense, not bravado. Consider partial hedges with payer swaptions against upside‑yield tails.
– Triggers to reassess: A material upside surprise in core inflation, a sharp rise in the MOVE index, or unusually heavy long‑end issuance relative to expectations.
The bigger picture
The model’s message is less about heroic forecasts and more about probabilities. When macro momentum cools and inflation’s short‑run pulse fades, long rates tend to slip as volatility recedes and term premia compress. That pattern dominates in the data the model learned from. But the same history also shows how quickly the narrative can flip on a single print or policy signal. In other words, the most likely path is a gentle grind lower in the 10‑year—with a constant need to monitor the few catalysts that could make the 35% minority outcome the one that matters.
Notes
– Accuracy metric: directional hit rate for 4‑week changes, evaluated out‑of‑sample with purged k‑fold validation across multiple regimes. Historical performance does not guarantee future results.
– This article is for informational purposes and is not investment advice.
