← Blog/blog/yield-curve-conditional-not-forecast

The yield curve that barely predicts—and still matters

A yield curve can be a weak weekly forecaster and still be a useful stress input. Qadi, Schröder, and Medda connect a three-factor Hull–White model to a vector autoregression with exogenous inputs (VARX): first generate coherent interest-rate paths, then ask what macro-financial distribution would accompany a specified path.

The important result is not a forecasting breakthrough. Once each variable's own lags are controlled for, the realised future curve adds almost no information about weekly changes in equities, activity, inflation, or bank risk. The engine's value is conditional: it answers "what outcomes go with this curve path?" rather than "what will the curve or economy do next?"

01

One model, two probability measures

A zero-coupon bond pays one unit at maturity and nothing before. Its price encodes the yield curve. A three-factor Hull–White model writes the short rate as a deterministic curve-fitting shift plus three latent Gaussian factors. The curve's dominant shapes are roughly level, slope, and curvature, while the factors are ordered by their mean-reversion speeds. Each factor is an Ornstein–Uhlenbeck process, a mean-reverting diffusion.

The distinction that matters is between the risk-neutral measure Q, used to price bonds consistently with today's discount curve, and the physical measure P, used to describe how rates actually evolve. Their drift difference is the market price of risk λ. In this specification, θᵢ = σᵢλᵢ/aᵢ, so λᵢ = aᵢθᵢ/σᵢ.

import math

def ou_step(x, a, theta, sigma, dt, z):
    decay = math.exp(-a * dt)
    innovation_scale = sigma * math.sqrt((1.0 - math.exp(-2.0 * a * dt)) / (2.0 * a))
    return theta + decay * (x - theta) + innovation_scale * z

This is the exact discrete-time OU transition, not an Euler approximation. The z input is a standard-normal draw; the live TypeScript chart uses the same formula in the core module, with every draw routed through the repository's seeded RNG.

FactoraθσImplied λ
fast1.557-0.002100.0498-0.066
medium0.4600.013420.07100.087
slow0.140-0.005990.0392-0.021
Paper-reported December 2024 calibration for a, θ, and σ; implied λ is recomputed by the core function aθ/σ. Values are rounded for display.

The paper keeps λ constant. That makes the risk premium inspectable, but it also rules out time-varying compensation for bond risk. The authors explicitly identify a state-dependent, essentially affine price of risk as the natural extension.

02

Three factors move at three speeds

The reported mean-reversion speeds are ordered fast, medium, and slow. The chart is one seeded synthetic path using those paper-reported parameters, a zero initial factor state, and independent innovations. The paper estimates a correlated innovation matrix; its numerical matrix is not tabulated, so independence here is an explicit teaching simplification, not a reproduction of the fitted curve paths.

fast factormedium factorslow factor
Illustrative one-year physical-measure factor path. Seed 20261122; parameters from the paper's Table 2; independent standard-normal innovations in this toy implementation.
03

A stress is a path, not a level

The VARX receives weekly changes in the curve's three principal components, not the curve level itself. That makes scenario design surprisingly easy to get wrong. If the starting curve is simply moved up by 200 basis points and then held there, the transmission layer sees no continuing factor changes; the paper verifies that level-only scenarios produce distributions indistinguishable from baseline up to simulation noise.

The authors instead ramp a parallel +200 bp shift over 26 weeks. That is 200/26 = 7.69 bp per week, roughly one historical weekly standard deviation. The earlier one-week 200 bp jump was far outside the variation used to estimate the weekly transmission layer.

26-week rampone-week jumplevel-only shift
Weekly curve movement seen by the transmission layer. The ramp is core-computed from 200 bp over 26 weeks; the one-week jump and level-only series are contrast cases, not paper forecasts.
04

The curve barely improves weekly forecasts

The cleanest diagnostic compares two otherwise identical VARX fits. Mode B sets the exogenous curve factors to zero. Mode A feeds in the realised future curve changes, isolating the maximum predictive information the curve could contribute. Lower RMSE is better. The 15 values below are transcribed from the paper's Table 12 and compared by the pure core function.

Paper-reported Mode A versus Mode B comparison. Core code counts 5 wins, 0 ties, and 10 losses for the curve-conditioned model.
EvaluationPaper resultWhat it says
G3++ curve RMSE13/20 wins; 1 tieNo Diebold–Mariano rejection vs random walk at 5%
Mode A vs Mode B5/15 wins; 3/9 continuous seriesRealised curve adds little weekly point accuracy
Joint pipeline CRPS2/9 best; Mode B 6/9Pipeline and Mode B within 2% in 8/9 cells
80% interval coverage7/9 near nominalWEI 4w and MES 26w are conservatively over-covered
All entries are paper-reported. CRPS is the continuous ranked probability score, which evaluates a full predictive distribution rather than only its median.
05

What the headline still does not establish

The study is one U.S. system, one currency, weekly first differences, and a post-2010 sample. The authors warn against carrying the no-signal result over to quarterly horizons, yield levels, or other monetary regimes. Two variables—the funds rate and forward-filled CPI inflation—barely move week to week, which also makes a no-change benchmark unusually hard to beat in some evaluation cells.

The structural policy-shock magnitudes need separate caution. The identified shock is normalised to a 25 bp rate move after a relevance filter and magnitude caps, and it is orthogonal to the same week's curve movement by construction. The authors therefore interpret the sign pattern and relative responses, not the raw impact magnitudes, as causal effect sizes.

06

What to probe next

Re-run the Mode A versus Mode B information test at the frequency that matters to the institution: monthly or quarterly data may tell a different story. Then compare constant and state-dependent market prices of risk, restore the empirical factor-correlation matrix, and test whether the same conditional paths survive outside U.S. rates. For the transmission layer, the site's linear regression page is a useful bridge: a VARX is a multivariate linear model with lagged outcomes and exogenous curve factors.

References

  1. Qadi, A., Schröder, T., & Medda, F. (2026). A Three-Factor Hull–White Scenario Generator for Macro-Financial Stress Forecasting. arXiv:2611.26200 [q-fin.GN]
  2. Hull, J., & White, A. (1990). Pricing Interest-Rate-Derivative Securities. The Review of Financial Studies, 3(4), 573–592
  3. Litterman, R., & Scheinkman, J. (1991). Common Factors Affecting Bond Returns. The Journal of Fixed Income, 1(1), 54–61