Abstract Traditional affine term structure models capture much of the variation in bond yields but struggle to generate risk premia that are consistently linked to macroeconomic conditions. We propose an arbitrage-free model that incorporates regime shifts in both the dynamics of risk factors and the stochastic discount factor, explicitly pricing regime-switching risk and accounting for the nonlinearities it creates. The model features four regimes generated by two independent Markov chains, which the data associate with monetary-tightening episodes and recessionary periods, respectively. Risk premia are strongly state-dependent: long-maturity premia are positive in calm periods but turn low or negative in high-volatility, low-slope regimes, reversing the usual maturity ordering. Expected excess returns are countercyclical, declining when inflation and industrial production rise, with effects that grow stronger at longer maturities. A macroaugmented specification uncovers a structural shift around 2000 after which inflation and bond risk premia move in opposite directions, consistent with Treasuries acquiring hedging value against macroeconomic downturns.
In this study, we propose a novel approach for forecasting realized covariance (RCOV) matrices of asset returns using a Riemannian-geometry-aware deep learning framework. In this way, we account for the geometric properties of covariance matrices, naturally handling the symmetric positive-definite constraint while considering non-linear dynamics, and efficiently managing high-dimensionality. Moreover, by using the Fr & eacute;chet sample mean of RCOV matrices, we extend the heterogeneous autoregressive model to a geometric approach in the matrix-variate framework. We apply our approach to forecast daily RCOV matrices for fifty among the most capitalized companies in the S&P 500 index and show that the methods presented in this study outperform traditional approaches in terms of predictive accuracy.
Abstract We address the inefficiency of particle filters in nonlinear and non-Gaussian state space models with superposed latent states, such as multi-component stochastic volatility and intensity models. Exploiting the superposition structure, we propose a simple two-step method: apply a particle filter to the superposed state, then recover individual components via a Kalman filter. This double filtering approach reduces the effective state dimension that can lead to particle degeneracy, and allows for likelihood-based inference. Simulation results show substantial gains in efficiency over particle filters that use the optimal proposal density but do not exploit the superposition. We apply the method to study different components in the common idiosyncratic volatility of stock returns and explore their pricing implications.
Abstract We study utility-based portfolio choice when optimal risky asset weights are state-dependent functions and only a subset of assets has nonzero weight functions. We propose a penalized nonparametric estimating-equation framework that augments kernel-weighted estimating equations derived from conditional Euler equations with a functional group penalty, so that asset-specific weight functions are selected directly. Under regularity conditions, we establish the existence of the proposed estimator and prove its oracle property. Simulations under quadratic and constant relative risk-aversion utilities are consistent with the theoretical results. Empirical applications to four monthly return datasets show that the proposed portfolios achieve competitive risk-adjusted performance and favorable out-of-sample quadratic utility while using substantially fewer active positions than benchmark strategies.
We re-examine U.S. Treasury bond return predictability using real-time macroeconomic vintages and news-topic attention measures. We propose a Weighted Group Neural Network (WGNN) that imposes pre-specified macroeconomic groupings to mitigate collinearity across predictors. The loss function adopts volatility weighting, making forecast errors comparable across maturities. For non-overlapping excess bond returns, out-of-sample predictability evidence is weak; for overlapping returns, the WGNN has the highest overall out-of-sample R2, exceeding 7%, among all models, and is the only model performing well across all maturities. Text-based predictors improve the WGNN forecasts. The performance decline after removing nonlinear activation is consistent with interaction effects or state dependencies that linear models do not capture. Economic evaluation shows that portfolio gains from better statistical forecasts shrink sharply under realistic leverage constraints and financing costs, especially for short-maturity strategies. Our results point to the roles of real-time measurement, regularization, and trading frictions in assessing bond return predictability.
