Building on the approach of Duffie and Kan (1996) who use finite maturity yields as the state vector, we propose a new representation of affine models in which the state vector is composed of infinitesimal maturity yields and their quadratic covariations. Because these variables possess unambiguous economic interpretations, they generate a representation that is globally identifiable. Further, this representation is more flexible than the maximal model of Dai and Singleton (2000) in that there are more identifiable parameters. We implement this new representation for two different three-factor models. The fact that our state vector can be estimated model-independently from yield curve data presents advantages for the estimation and interpretation of multi-factor models.

We propose a canonical representation for affine term structure models where the state vector is comprised of the first few Taylor-series components of the yield curve and their quadratic (co-)variations. With this representation: (i) the state variables have simple physical interpretations such as level, slope and curvature, (ii) their dynamics remain affine and tractable, (iii) the model is by construction 'maximal' (i.e., it is the most general model that is econometrically identifiable), and (iv) model-insensitive estimates of the state vector process implied from the term structure are readily available. (Furthermore, this representation may be useful for identifying the state variables in a squared-Gaussian framework where typically there is no one-to-one mapping between observable yields and latent state variables). We find that the 'unrestricted' A1(3) model of Dai and Singleton (2000) estimated by 'inverting' the yield curve for the state variables generates volatility estimates that are negatively correlated with the time series of volatility estimated using a standard GARCH approach. This occurs because the 'unrestricted' A1(3) model imposes the restriction that the volatility state variable is simultaneously a linear combination of yields (i.e., it impacts the cross-section of yields), and the quadratic variation of the spot rate process (i.e., it impacts the time-series of yields). We then investigate the A1(3) model which exhibits 'unspanned stochastic volatility' (USV). This model predicts that the cross section of bond prices is independent of the volatility state variable, and hence breaks the tension between the time-series and cross-sectional features of the term structure inherent in the unrestricted model. We find that explicitly imposing the USV constraint on affine models significantly improves the volatility estimates, while maintaining a good fit cross-sectionally.

In this paper, we explore the features of affine term structure models that are empirically important for explaining the joint distribution of yields on short and long-term interest rate swaps. We begin by showing that the family of N-factor affine models can be classified into N+1 non-nested sub-families of models. For each sub-family, we derive a maximal model with the property that every admissible member of this family is equivalent to or a nested special case of our maximal model. Second, using our classification scheme and maximal models, we show that many of the three-factor models in the literature impose potentially strong over-identifying restrictions on the joint distribution of short- and long-term rates. Third, we compute simulated method-of-moments estimates for several members of one of the four branches of three-factor models, and test the over-identifying restrictions implied by these models. We conclude that many of the extant affine models in the literature fail to describe important features of the distribution of long- and short- term rates. The source of the model misspecification is shown to be overly strong restrictions on the correlations among the state variables. Relaxing these restrictions leads to a model that passes several goodness-of-fit tests over our sample period.

This paper develops new results for identification and estimation of Gaussian affine term structure models. We establish that three popular canonical representations are unidentified, and demonstrate how unidentified regions can complicate numerical optimization. A separate contribution of the paper is the proposal of minimum-chi-square estimation as an alternative to MLE. We show that, although it is asymptotically equivalent to MLE, it can be much easier to compute. In some cases, MCSE allows researchers to recognize with certainty whether a given estimate represents a global maximum of the likelihood function and makes feasible the computation of small-sample standard errors.

We develop a Gaussian discrete time essentially affine term structure model with long memory state variables. This feature reconciles the strong persistence observed in nominal yields and inflation with the theoretical implications of affine models, especially for long maturities. We characterise in closed-form the dynamic and cross-sectional implications of long memory for our model. We explain how long memory can naturally arise within the term structure of interest rates, providing a theoretical underpinning for our model. Despite the infinite-dimensional structure that long memory implies, we show how to cast the model in state space and estimate it by maximum likelihood. An empirical application of our model is presented.