## _wp10258

## Source details

**Canonical URL:** [_wp10258](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2010/_wp10258.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2010/_wp10258.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2010/_wp10258.pdf.json)

---

### Introduction and scope
- Objective: discuss estimation of term-structure models and estimate the United States’ term structure of Treasury bond yields.
- Dataset span: 1972 to 2007.
- Sample period (empirical application): January 1972 to December 2007 (1972:1-2007:12).
- Yield data: U.S. Treasury monthly zero-coupon bond yields (annualized, continuously compounded), 432 months; 3,888 monthly observations; 9 maturities: 3, 6, 12, 24, 36, 48, 60, 84 and 120 months.
- Macroeconomic variables: (i) inflation: annual percentage change in the monthly price deflator for personal consumption expenditures; (ii) manufacturing capacity utilization; (iii) monthly average federal funds rate.
- Software: MCM Term Structure Software (MCMTS) developed by Fund staff; capable of modeling at least nine models, generating Monte Carlo paths, computing risk indicators such as Value-at-Risk (VaR).

### Major model classes reviewed
- Nelson-Siegel Models (NSMs)
  - Functional-form approach with at least three latent factors interpreted as level, slope, and curvature.
  - Can include observable macroeconomic variables; state-space representation with VAR(1) dynamics and diagonal measurement-error covariance H in yield-only and augmented models.
  - Decay parameter λ controls exponential decay; example value that maximizes curvature at exactly 30 months is λ = 0.0609.
  - Estimation approaches:
    - Two-step: fix λ, OLS cross-sectionally to extract factors, then estimate VAR(1).
    - One-step: joint maximum likelihood in state-space via Kalman filter; example parameter counts: yield-only model has 36 parameters; yield-macro model has 81 parameters.
- Affine Term-Structure Models (ATSMs)
  - Short rate specified as linear combination of state variables; state variables follow SDEs with drift and diffusion driven by Wiener processes.
  - No-arbitrage pricing via equivalent martingale measure Q; bond prices solved from Backward Kolmogorov PDE / Feynman-Kac representation.
  - Cox-Ingersoll-Ross (CIR) model:
    - Short rate under P: dr_t = κ(θ - r_t) dt + σ sqrt(r_t) dW_t^P (equation (28)).
    - Positivity: if κ > 0 and θ > 0 short rate non-negative; if 2·κ·θ ≥ σ^2 short rate strictly positive.
    - Under Q: dr_t = [κ(θ - r_t) - λσ sqrt(r_t)] dt + σ sqrt(r_t) dW_t^Q (λ market price of risk).
    - Bond price exponential-affine: b(t,r,τ;β) = exp[A(τ;β) - C(τ;β) · r_t] (equation (31)); yields affine in r_t via y(t,τ;β) = [A(τ;β)/τ] + C(τ;β) · r_t.
  - Affine multi-factor models: exponential-affine bond price conjecture b(x_1,x_2,t,T) = exp[α(τ) + β_1(τ) x_1 + β_2(τ) x_2], reduction of PDE to ODEs (Riccati equations) with boundary conditions α(0)=0, β_1(0)=0, β_2(0)=0.

### State-space, equivalent martingale measure, and Monte-Carlo implications
- Equivalent martingale measure Q
  - Q equivalent to P: for any event E, P(E) > 0 ⇔ Q(E) > 0.
  - Under Q discounted security prices are martingales; instantaneous expected return on any security equals the short rate.
  - Construction via Girsanov: Radon-Nikodym derivative ξ as in equations (10)–(19); market price of risk process θ enters change of measure.
  - Practical note: Monte-Carlo simulations under Q appropriate for pricing (expectations under Q), but not appropriate for computing VaR (VaR requires real-world measure P).
- No-arbitrage PDE and pricing representation
  - Bond price b satisfies Backward Kolmogorov (no-arbitrage) PDE; Feynman-Kac representation (equation (27)) gives b as Q-expectation of discounted payoff.

### Empirical findings: stylized facts and NSM fit (U.S., 1972:1–2007:12)
- Stylized facts preserved verbatim:
  - The average yield curve is upward sloping and concave.
  - The yield curve assumes a variety of shapes through time, including upward sloping, downward sloping, humped, and inverted humped.
  - Yield dynamics are persistent, while spread dynamics are less persistent.
  - The short end of the yield curve is more volatile than the long end.
  - The level of the yield curve is highly persistent, but exhibits small variation relative to its mean. In fact the level is more persistent than any single yield.
  - The slope is less persistent than any single yield but highly variable relative to its mean.
  - The curvature is the least persistent of all factors and displays the largest variability relative to its mean.
- Nelson-Siegel three-factor fit (yield-only model, selected exact statistics):
  - Yields (Table 1, selected):
    - 3M: Mean 36.101 Std. Dev. 12.969 Minimum 0.860 Maximum 15.999 ρ(1) 0.981 ρ(12) 0.777 ρ(30) 0.454
    - 120 (level): Mean 7.514 Std. Dev. 2.484 Minimum 3.372 Maximum 15.065 ρ(1) 0.991 ρ(12) 0.873 ρ(30) 0.728
    - Slope: Mean 1.413 Std. Dev. 1.447 Minimum -3.223 Maximum 4.140 ρ(1) 0.941 ρ(12) 0.419 ρ(30) -0.154
    - Curvature: Mean -0.016 Std. Dev. 0.782 Minimum -2.062 Maximum 3.012 ρ(1) 0.838 ρ(12) 0.332 ρ(30) 0.094
  - Residuals (Table 2, selected):
    - 3M residuals: Mean -0.059 Std. Dev. 0.092 Minimum -0.628 Maximum 0.226 ρ(1) 0.655 ρ(12) 0.274 ρ(30) 0.108
    - 120 residuals: Mean 0.032 Std. Dev. 0.069 Minimum -0.159 Maximum 0.311 ρ(1) 0.739 ρ(12) 0.356 ρ(30) 0.032
  - Estimated factors (Table 3, selected):
    - Level: Mean 7.723 Std. Dev. 2.377 Minimum 3.934 Maximum 14.399 ρ(1) 0.992 ρ(12) 0.879 ρ(30) 0.734
    - Slope: Mean 1.693 Std. Dev. 1.903 Minimum -5.136 Maximum 5.370 ρ(1) 0.949 ρ(12) 0.456 ρ(30) -0.134
    - Curvature: Mean -0.472 Std. Dev. 2.040 Minimum -5.953 Maximum 6.007 ρ(1) 0.884 ρ(12) 0.439 ρ(30) 0.062
  - Correlations:
    - Corr(level estimates, empirical level) = 0.99
    - Corr(slope estimates, empirical slope) = 0.99
    - Corr(curvature estimates, empirical curvature) = 0.97
  - Factor persistence (VAR own-lag coefficients):
    - Estimated own-lag coefficients: level 1.00, slope 0.92, curvature 0.84 (stationarity assured by eigenvalues with modulus less than 1).

### Yield-macro model: parameters and diagnostic statistics
- Yield-macro model augmenting NSM with CU, INF, FFR (yield loadings only on yield factors).
- Estimated factors (Table 5, selected):
  - Level: Mean 7.706 Std. Dev. 2.396 Minimum 3.921 Maximum 14.503 ρ(1) 0.992 ρ(12) 0.877 ρ(30) 0.733
  - Slope: Mean 1.674 Std. Dev. 1.875 Minimum -5.023 Maximum 5.279 ρ(1) 0.946 ρ(12) 0.442 ρ(30) -0.134
  - Curvature: Mean -0.601 Std. Dev. 2.125 Minimum -6.259 Maximum 5.756 ρ(1) 0.890 ρ(12) 0.465 ρ(30) 0.050
- Measurement errors (Table 6, selected):
  - 3M: Mean -0.059 Std. Dev. 0.084
  - 120: Mean 0.038 Std. Dev. 0.072
  - Average standard deviation for maturities 6 to 60 months ≈ 6.4 basis points (statement preserved from source).
- VAR(1) parameter estimates (Table 7, selected entries preserved with standard errors and t-stats):
  - Level(-1) row highlights:
    - 0.97 (0.02476) [39.2045], -0.09 (0.04906) [-1.78343], 0.21 (0.08044) [2.64645], 0.29 (0.04287) [6.84723], 0.56 (0.04298) [12.9222], -0.04 (0.02178) [-1.78560]
  - Slope(-1) row highlights:
    - 0.00 (0.02398) [-0.17922], 1.00 (0.04753) [20.9513], -0.18 (0.07793) [-2.32893], -0.21 (0.04153) [-5.04640], -0.52 (0.04164) [-12.5521], 0.03 (0.0211) [1.22017]
  - Curvature(-1) row highlights:
    - -0.03 (0.00872) [-2.91824], -0.06 (0.01728) [-3.75630], 0.83 (0.02834) [29.4552], 0.02 (0.0151) [1.26505], 0.05 (0.01514) [3.27703], -0.02 (0.00767) [-2.02061]
  - Interpretation notes preserved:
    - Diagonal elements significant; 12 out of 30 off-diagonal elements appear insignificant.
    - Lower-left 3x3 block (macroeconomic→yield factors) has five insignificant coefficients; upper-right 3x3 block (yield factors→macroeconomic) has three insignificant coefficients.

### Impulse-response results and economic interpretation
- Four groups of impulse responses analyzed: (i) macro→macro; (ii) macro→yield; (iii) yield→macro; (iv) yield→yield.
- Macro→macro responses:
  - Macroeconomic variables show significant persistence except for FFR.
  - An increase in FFR lowers CU over following years; FFR rises with CU and INF consistent with an estimated monetary policy reaction function.
  - INF increases with CU and has a negative delayed response to FFR.
- Macro responses to yield shocks:
  - Macro variables mostly insignificant to curvature shocks.
  - Increase in slope produces an important decline in FFR (close relation between slope and monetary policy instrument).
  - Increase in level raises CU, FFR, and INF; Fed accommodated only a fraction of expected inflation by increasing nominal funds rate, reducing CU, and limiting increase in INF during sample.
- Yield responses to macro shocks:
  - Macroeconomic shocks have little influence on curvature.
  - Positive shock to FFR quickly increases the slope (steeper or less negatively sloped curve).
  - Positive shocks to CU, and to a lesser degree INF, produce inverse and delayed responses in slope, suggesting bond market anticipates monetary tightening.
  - Shocks to INF and CU generate a prolonged increase in the level factor, suggesting long-term inflation expectations may not be firmly anchored.
  - Positive shock to FFR induces a small temporary increase in the level factor.
- Yield responses to yield shocks:
  - The three yield curve factors display significant persistence.
  - Shock to level increases slope (associated with lowering of short end relative to long end and loosening of monetary policy).
  - Surprise increase in slope reduces level, shifting entire curve downwards with relatively stronger decline in the short end.

### Estimation techniques, algorithms, and computational notes
- MCMTS estimation approach:
  - One-stage Powell method used for Nelson-Siegel and CIR model estimation.
  - Nelson-Siegel: fixed decay parameters in MCMTS to stabilize state-space framework; joint optimization via Kalman filter with OLS-based cross-sectional extraction to compute H and AR/VAR to compute Q.
  - CIR: state-innovation covariance Q_jj depends on β_{j,t-1} and is dynamically updated in the Kalman filter.
- Algorithmic steps (Nelson-Siegel high-level):
  - For given λ compute loadings x_{1,τ} = 1; x_{2,τ} = (1 - exp(-λ τ)) / (λ τ); x_{3,τ} = x_{2,τ} - exp(-λ τ).
  - Cross-sectional OLS per t to obtain β_t and ε_t; compute H from ε_t series.
  - AR(1)/VAR(1) on β_t to obtain μ, Φ, Q.
  - Kalman filter to compute adjusted factors and log-likelihood L = Σ_{t=1}^T [ -1/2 ln | f_{t|t-1} | -1/2 η_{t|t-1}' f_{t|t-1}^{-1} η_{t|t-1} ].
- Convergence and computational burden (Appendix Table 1 preserved):
  - Nelson-Siegel model with one decay parameter: Number of Parameters: 1; Parameters: λ; Number of Iterations before Convergence: <10
  - Nelson-Siegel model with two decay parameters: Number of Parameters: 2; Parameters: λ_1, λ_2; Number of Iterations before Convergence: <50
  - One-factor Cox-Ingersoll-Ross model: Number of Parameters: 4; Parameters: λ, σ, θ, κ; Number of Iterations before Convergence: <200
  - Two-factor Cox-Ingersoll-Ross model: Number of Parameters: 8; Parameters: λ_1, σ_1, θ_1, κ_1 ; λ_2, σ_2, θ_2, κ_2; Number of Iterations before Convergence: <2,000

### Monte Carlo simulation and diagnostics
- Simulation framework:
  - Factor dynamics driven by AR(1) or VAR(1); user sets number of simulated paths N and number of steps S.
  - Nelson-Siegel: draw K i.i.d. standard normals ω ∼ N(0, I); obtain q = C ω with C = Cholesky(Q); use q to drive state evolution.
  - CIR: Euler discretization β_{j,t+Δ} = β_{j,t} + κ_j(θ_j - β_{j,t}) Δ + σ_j sqrt(β_{j,t} Δ) z_{j,t}, z_{j,t} ∼ N(0, Δ).
  - MCMTS computes summary statistics (mean, median, standard deviation, percentiles) and displays histograms and simulated paths.

### Software features, user workflow, and proposed extensions
- MCMTS capabilities:
  - Create datasets, select models (Nelson-Siegel family, CIR family, Macro-Factor Models), estimate parameters, run Monte Carlo simulations, estimate impulse responses and variance decomposition.
  - Implemented as an Excel add-in with algorithms developed in C#.
- User workflow and defaults:
  - Data sheets required: "Yield Curve" (zero-coupon yields by maturity), "Macro Factors".
  - Frequency options: daily, weekly, monthly.
  - Simulation horizon defaults: Daily: one year after end date; Weekly: three years after end date; Monthly: ten years after end date.
  - Nelson-Siegel alternatives: three-factor base; Bjork and Christensen four-factor; Bliss three-factor; Svensson four-factor; Adjusted Svensson four-factor.
  - Factor relationship options: Independent Factors, AR(1), VAR(1).
  - Optimization methods: No Optimization; Minimize Measurement RMSE; Maximum Likelihood with Kalman Filter.
  - CIR options: One-factor or two-factor CIR.
  - Model evaluation: cross-sectional fit, time-series fit, Chi-square test of overall fit.
- Proposed extensions preserved from source:
  - Multi-country/global yield curve analysis via hierarchical dynamic factor models with global and country-specific factors using state-space and Bayesian MCMC methods.
  - Introduction of no-arbitrage restrictions: extend Nelson-Siegel to arbitrage-free (affine arbitrage-free Nelson-Siegel) specifications to integrate absence of arbitrage while maintaining Nelson-Siegel factor structure.

*Source: _wp10258 - References .............................................................................................................*

### References .............................................................................................................

### _wp10258 - References .............................................................................................................

### Introduction
- The modeling of the term structure of interest rates helps investors, market analysts, and policymakers:
  - Value financial instruments, including credit derivatives.
  - Simulate interest rate scenarios.
  - Assess impact of interest rate movements on default probabilities of different financial instruments.
- The need for term-structure modeling increased due to the sharp interest rate movements in the context of the global financial crisis.

### Major model classes reviewed
- Nelson-Siegel Models (NSMs)
  - Postulate a particular functional form for the term structure of interest rates.
  - Do not need to depend on the existence of arbitrage possibilities.
  - Rely on at least three latent factors interpreted as level, slope, and curvature.
  - Can include observable macroeconomic variables to capture joint behavior of the yield curve and macroeconomic variables.
- Affine Term-Structure Models (ATSMs)
  - May depend on the absence of arbitrage opportunities.
  - Assume unobservable factors underlying the term structure follow stochastic processes.
  - Include traditional models such as:
    - General single-factor model.
    - Cox-Ingersoll-Ross (CIR) model.
    - Affine multi-factor models.

### Paper scope and dataset
- Primary objective: discuss estimation of term-structure models and estimate the United States’ term structure of Treasury bond yields.
- Dataset span: 1972 to 2007.
- Software:
  - Developed by Fund staff.
  - Capable of modeling the term structure using at least nine models.
  - Enables generation of different paths of the term structure.
  - Facilitates computation of risk indicators such as Value-at-Risk (VaR) for managing investment-holdings risk.

### Structure and methodological focus
- Sectional coverage (as presented in the source):
  - Reviews of simple versions of NSMs and ATSMs, with focus on an extension of the CIR model.
  - Illustration of software capabilities through estimation of the U.S. Treasury yield curve.
  - Discussion of possible extensions of the software.
- Appendices and technical content (headings and focus areas listed in source):
  - Appendix I. Term-Structure Modeling Using the MCM Term Structure Software
    - I. The Short Rate
    - II. Equivalent Martingale Measure
    - III. Existence of an Equivalent Martingale Measure
    - IV. Equivalent Martingale Measures and Monte-Carlo Simulations
    - V. Single-Factor Term Structure Models
    - VI. The Cox-Ingersoll-Ross Model
    - VII. Affine Multi-Factor Models
  - Appendix II. Estimation Techniques
    - I. Implementation of Nelson-Siegel Models
    - II. Implementation of Cox-Ingersoll-Ross Models
    - III. Estimation Analysis
    - IV. Monte Carlo Simulations
  - Appendix III. Description of the MCM Term Structure Software
  - Appendix Table 1. Estimation analysis for Nelson-Siegel and CIR models

### Empirical and analysis outputs highlighted
- Tables (selected titles and focus):
  - 1. Yield-Only Model Descriptive Statistics, Yield Curves
  - 2. Yield-Only Model Descriptive Statistics, Yield Curve Residuals
  - 3. Yield-Only Model Descriptive Statistics, Estimated Factors
  - 4. Yield-Only Model VAR Parameter Estimates
  - 5. Yield-Macro Model Descriptive Statistics, Estimated Factors
  - 6. Summary Statistics for Measurement Errors of Yields
  - 7. Yield-Macro Model Paramenter Estimates
- Figures (selected titles and focus):
  - 1. Observed Yield Curves
  - 2. Selected Fitted (Model Based) Yield Curves
  - 3. Observed and Estimated Average Yield Curve
  - 4. Estimates of the Level, Slope, and Curvature in the Yields-Only Model
  - 5. Model-Based vs. Data-Based Level, Slope, and Curvature
  - 6. Yield-only Model Level Factor and Inflation
  - 7. Yield-only Model Slope Factor and Capacity Utilization
  - 8. Impulse-Response Functions

### Key applications and implications
- Applications of the modeling framework and software:
  - Valuation of bonds and credit derivatives.
  - Interest-rate scenario simulation via Monte Carlo.
  - Computation of risk metrics such as Value-at-Risk (VaR).
  - Assessment of policy impacts on the yield curve and associated financial risks.

*Source: _wp10258 - References .............................................................................................................*

### Appendix I presents a detailed derivation of some of the main ATSM, Appendix II

### _wp10258 - Appendix I presents a detailed derivation of some of the main ATSM, Appendix II

### Background: term structure modeling aims and challenges
- The term structure (yield curve) links short-, medium-, and long-term nominal bond rates and exhibits time-varying shapes that follow systematic patterns.
- Key modeling objective: provide a parsimonious representation that (i) summarizes information for a large number of traded nominal bonds at a point in time, (ii) reproduces historical stylized facts of the average shape of the yield curve, and (iii) forecasts future interest rates.
- Most models assume a small number of unobservable (latent) factors with factor loadings relating yields of different maturities to those factors.
- Two popular factor-model classes: Nelson-Siegel Models (NSMs) and Affine Term-Structure Models (ATSMs).

### Nelson-Siegel Models (NSMs)
- Core idea: approximate the instantaneous forward rate and derive the corresponding zero-coupon yield curve as a function of a small set of parameters and loadings.
- Common specification uses three time-varying latent factors interpreted as:
  - Level (long-term factor)
  - Slope (short-term factor)
  - Curvature (medium-term factor)
- The decay parameter controls exponential decay and the maturity at which the curvature loading attains its maximum; an example value that maximizes curvature at exactly 30 months is ߣ = 0.0609.

Key empirical and practical properties
- Parsimonious: three loadings (1, a decaying function of maturity, and a concave function of maturity) give flexibility to reproduce a range of observed yield-curve shapes.
- Forward and yield curves start at an instantaneous rate that depends on level and slope and level off at the finite infinite-maturity value of the level factor.
- Time-series dynamics of the three latent factors underlie historical yield-curve dynamics.
- Limitations: difficulty fitting yield curves when data are dispersed with multiple interior minima and maxima; extensions (additional factors or decay parameters) can improve cross-sectional fit but complicate estimation and may have unclear benefits for forecasting and dynamics, especially for multi-country analysis.

State-space representation (general framework)
- NSM can be written in state-space form with:
  - Transition equation: the state vector (for three-factor NSM) 3x1 and assumed to follow a VAR(1) in deviations from the mean; transition matrix A is 3x3; transition disturbance ߟ_t has 3x3 non-diagonal covariance matrix Q.
  - Measurement equation: an A×3 measurement matrix Λ whose columns are the loadings on the three factors; measurement disturbance ߝ_t has A×A diagonal covariance matrix H.
- Assumptions for estimation: disturbances orthogonal to each other and to initial state.

Yield curve models augmented with macro factors
- Diebold, Rudebusch, and Auroba (2006) extend the state vector to include three macroeconomic variables: manufacturing capacity utilization, the federal funds rate, and annual price inflation.
- Augmented state vector dimension increases from 3 to 6; correspondingly ߤ is 6x1, A becomes 6x6, and ߟ_t is 6x1.
- Measurement matrix Λ in the yield-macro model contains six columns: the three leftmost are the yield-factor loadings; the three rightmost are zeroes (yields load only on yield factors).
- Transition disturbance covariance matrix Q becomes 6x6 and is non-diagonal; measurement disturbance covariance matrix H remains diagonal.
- After estimation, interactions between yields and macro variables are explored via impulse response functions and variance decompositions; main empirical finding summarized in this content:
  - Bidirectional causality is likely present.
  - Effects of macroeconomic factors on future yield curves appear relatively more important than effects of yield-curve factors on future macroeconomic developments.
  - Market yields still contain relevant predictive information about the stance of monetary policy.

Estimation approaches for Nelson-Siegel specifications
- Two broad estimation approaches:
  1. Two-step approach
     - Example: Diebold and Li (2006).
     - Fix the decay parameter ߣ_t at a prespecified value constant over time (ߣ_t = ߣ), then use ordinary least squares (OLS) cross-sectionally at each time t to estimate factor time series.
     - Second step: estimate transition equation (VAR(1)) typically assuming A and Q diagonal.
     - Practical simplification: fixing ߣ avoids nonlinear least squares at each t.
  2. One-step approach
     - Example: Diebold, Rudebusch, and Auroba (2006).
     - Place model in state-space form and estimate all parameters jointly via maximum likelihood using Kalman filter techniques.
     - Produces correct inference via standard theory but requires numerical optimization over many parameters (example parameter counts: yield-only model has 36 parameters; yield-macro model has 81 parameters).
- Implementation choices described in the source:
  - Decay parameter ߣ is fixed over time but not prespecified; its value is chosen to optimize estimation performance.
  - Two-step implementation: choose ߣ to minimize root mean square error (RMSE) of the measurement equation by:
    - For a candidate ߣ, compute Λ, run OLS cross-sectionally to extract latent factors and measurement errors, compute measurement disturbance covariance matrix and RMSE, and select ߣ with lowest RMSE.
    - Then estimate VAR(1) transition equation by OLS to obtain A, ߤ, and transition disturbance covariance Q.
  - One-step implementation: maximize the state-space log-likelihood over ߣ, ߤ, Λ, Q, and R using Kalman filter prediction-error decomposition.
    - Kalman filter iterations use unconditional mean and variance of the latent factor vector for initial conditions; the latent factor vector ܨ_t is estimated as part of the procedure.

Practical notes on identification and inference
- One-step (full likelihood / Kalman filter) approach yields better large-sample inference properties but increases computational burden and dimensionality of numerical optimization.
- Diebold, Rudebusch, and Auroba (2006) use an identification ordering for impulse responses when constructing orthogonal shocks from Q; their ordering (for identification) is: ߚ_level, ߚ_slope, ߚ_curvature, manufacturing capacity utilization, inflation, federal funds rate. Results reported as robust to alternative identification strategies.

### Affine Term-Structure Models (ATSMs)
- ATSMs are another factor-class linking term structure dynamics to a set of state variables; used mainly in academic and market finance contexts.
- Modeling building blocks:
  - Instantaneous short-term rate (the instantaneous interest rate) is specified as a linear combination of state variables.
  - State variables follow stochastic differential equations (SDEs): a drift component plus diffusion (variance) component driven by Wiener processes.
  - Price of a pure discount bond is modeled as a function of maturity and the state variables; bond price dynamics require stochastic calculus for derivation.
- Conceptual note: the instantaneous interest rate is a theoretical construct (limit of short-maturity yields) and has no direct empirical counterpart.

*Source: _wp10258 - Appendix I presents a detailed derivation of some of the main ATSM, Appendix II*

### Appendix I summarizes the

### _wp10258 - Appendix I summarizes the

### Model assumptions and mathematical framework
- The appendix summarizes the assumptions underlying stochastic equations and derives a one-single factor model, the CIR, and multi-factor Affine Term-Structure Models (ATSM).
- Generic notation: change in the price of a bond denoted ܲ݀, used to obtain the relative change in the value of a riskless self-financing portfolio composed of a long position in a pure discount bond with instantaneous maturity and as many short positions (each with a different maturity) as there are risk factors.
- Objective: select portfolio weights to eliminate underlying sources of risk so the portfolio earns the risk-free rate (absence of arbitrage), producing a partial differential equation (PDE) that is solvable analytically because the term structure model is affine.
- Key stochastic elements:
  - An SDE is of the form ܺ݀௧ߤൌሺܺ௧ݐ,ሻߪ൅ݐ݀ሺܺ௧ݐ,ሻܹ݀௧, where ܹ௧ is a scalar Wiener process.
  - A standard scalar Wiener process ܹ௜ሺݐሻ has normal distribution with mean zero and variance ݐ, i.e., ܹ௜ሺݐሻܰ׽ሺݐ.0ሻ.
  - If there are ݋ state variables, the model requires ݋+1 bonds to construct a riskless portfolio.
  - A function Թ is affine if ݂ሺݔሻ=ߚ൅ߙ்ݔ for all Թאݔ௡.

### Multi-factor extension of Cox-Ingersoll-Ross (CIR)
- Instantaneous short-term interest rate ݕ௧ is a linear combination of ݋ independent state variables ݖଵ(ݐ),..,ݖ௡(ݐ):
  - Equation (14): ݕ௧ൌ∑ݖ௜(ݐ)ൡ௜ୀଵ.
- Each state variable follows a square-root (CIR-type) process:
  - Equation (15): ݖ݀௜(ݐ)ߢൌ௜ߠ൫௜ݖെ௜(ݐ)ߪ൅ݐ݀൯௜ඥݖ௜(ݐ)ܹ݀௜(ݐ), for ݅ൌ1,...,݋.
  - Drift term is mean-reverting with parameters ߢ௜ and ߠ௜; diffusion term has volatility parameter ߪ௜.
  - The ܹ௜(ݐ) are independent scalar Wiener processes defined on probability space (Ω,࣠,Զ).
- Zero-coupon bond price as a function of maturity ߬ and state variables:
  - Equation (16): ܲ(߬)ܲൌሺݖ,߬ଵ,..,ݖ௡ሻ.
  - The PDE from riskless portfolio selection is solved analytically under affine specifications.
- CIR n-factor solution: closed-form expressions for the loadings of bond price on state variables given by functions ܣ௜(߬) and ܤ௜(߬):
  - Equations (19) and (20) provide these expressions (symbols preserved as in source).
  - γ୧ൌඥሺκ୧൅λ୧ሻଶ൅2σ୧ଶ, where λ୧ is the market price of risk.

### Application to the United States: data and empirical setup
- Sample period: January 1972 to December 2007 (1972:1-2007:12).
- Yield data:
  - U.S. Treasury monthly zero-coupon bond yields (annualized, continuously compounded).
  - 432 months; 3,888 monthly observations.
  - 9 maturities: 3, 6, 12, 24, 36, 48, 60, 84 and 120 months.
- Macroeconomic variables (monthly, same period):
  - (i) inflation: annual percentage change in the monthly price deflator for personal consumption expenditures;
  - (ii) real economic activity relative to potential: manufacturing capacity utilization;
  - (iii) monetary policy instrument: monthly average federal funds rate.
- Modeling approach:
  - Assume three factors—level, slope and curvature—suffice to summarize the term structure dynamics.
  - Level defined as the 4-year yield; slope as one-month minus 4-year yield; curvature as twice the one-year yield minus the sum of the one-month and four-year yields.

### Stylized facts of the U.S. yield curve (1972:1–2007:12)
- Identified stylized facts (verbatim where numeric or hedged language appears in source):
  - The average yield curve is upward sloping and concave.
  - The yield curve assumes a variety of shapes through time, including upward sloping, downward sloping, humped, and inverted humped.
  - Yield dynamics are persistent, while spread dynamics are less persistent.
  - The short end of the yield curve is more volatile than the long end.
  - The level of the yield curve is highly persistent, but exhibits small variation relative to its mean. In fact the level is more persistent than any single yield.
  - The slope is less persistent than any single yield but highly variable relative to its mean.
  - The curvature is the least persistent of all factors and displays the largest variability relative to its mean.

### Nelson-Siegel model fit and key statistics
- A three-factor Nelson-Siegel Model (NSM) fits U.S. monthly Treasury yields well (432 cross sections estimated).
- Descriptive sample details and selected exact statistics preserved from tables:
  - Yield-only model, summary for yields (selected entries from Table 1):
    - 3M: Mean 36.101 Std. Dev. 12.969 Minimum 0.860 Maximum 15.999 ρ(1) 0.981 ρ(12) 0.777 ρ(30) 0.454
    - 120 (level): Mean 7.514 Std. Dev. 2.484 Minimum 3.372 Maximum 15.065 ρ(1) 0.991 ρ(12) 0.873 ρ(30) 0.728
    - Slope: Mean 1.413 Std. Dev. 1.447 Minimum -3.223 Maximum 4.140 ρ(1) 0.941 ρ(12) 0.419 ρ(30) -0.154
    - Curvature: Mean -0.016 Std. Dev. 0.782 Minimum -2.062 Maximum 3.012 ρ(1) 0.838 ρ(12) 0.332 ρ(30) 0.094
  - Residuals (Table 2, selected):
    - 3M residuals: Mean -0.059 Std. Dev. 0.092 Minimum -0.628 Maximum 0.226 ρ(1) 0.655 ρ(12) 0.274 ρ(30) 0.108
    - 120 residuals: Mean 0.032 Std. Dev. 0.069 Minimum -0.159 Maximum 0.311 ρ(1) 0.739 ρ(12) 0.356 ρ(30) 0.032
  - Estimated factors (Table 3):
    - Level: Mean 7.723 Std. Dev. 2.377 Minimum 3.934 Maximum 14.399 ρ(1) 0.992 ρ(12) 0.879 ρ(30) 0.734
    - Slope: Mean 1.693 Std. Dev. 1.903 Minimum -5.136 Maximum 5.370 ρ(1) 0.949 ρ(12) 0.456 ρ(30) -0.134
    - Curvature: Mean -0.472 Std. Dev. 2.040 Minimum -5.953 Maximum 6.007 ρ(1) 0.884 ρ(12) 0.439 ρ(30) 0.062
  - Correlations between estimated factors and empirical proxies:
    - Corr(level estimates, empirical level) = 0.99
    - Corr(slope estimates, empirical slope) = 0.99
    - Corr(curvature estimates, empirical curvature) = 0.97
  - Factor persistence (VAR own-lag coefficients):
    - Estimated own-lag coefficients: level 1.00, slope 0.92, curvature 0.84 (noting stationarity assured by eigenvalues with modulus less than 1).

### Yield-macro model and interactions
- Extended NSM to include three macroeconomic variables (manufacturing capacity utilization CU, federal funds rate FFR, annual price inflation INF) in a yield-macro framework where yields load only on yield curve factors.
- Yield-macro model descriptive statistics (Table 5, estimated factors):
  - Level: Mean 7.706 Std. Dev. 2.396 Minimum 3.921 Maximum 14.503 ρ(1) 0.992 ρ(12) 0.877 ρ(30) 0.733
  - Slope: Mean 1.674 Std. Dev. 1.875 Minimum -5.023 Maximum 5.279 ρ(1) 0.946 ρ(12) 0.442 ρ(30) -0.134
  - Curvature: Mean -0.601 Std. Dev. 2.125 Minimum -6.259 Maximum 5.756 ρ(1) 0.890 ρ(12) 0.465 ρ(30) 0.050
- Measurement errors for yields in the yield-macro model (Table 6, selected):
  - 3M: Mean -0.059 Std. Dev. 0.084
  - 120: Mean 0.038 Std. Dev. 0.072
  - Average standard deviation for maturities 6 to 60 months ≈ 6.4 basis points (noted in source).
- VAR parameter estimates (Table 7, selected highlights and exact coefficients preserved where representative):
  - The 6x6 coefficient matrix includes blocks showing interactions among level, slope, curvature, CU, INF, FFR.
  - Selected entries (values, standard errors in parentheses, t-statistics in brackets as in source):
    - Level(-1) row: 0.97 (0.02476) [39.2045], -0.09 (0.04906) [-1.78343], 0.21 (0.08044) [2.64645], 0.29 (0.04287) [6.84723], 0.56 (0.04298) [12.9222], -0.04 (0.02178) [-1.78560]
    - Slope(-1) row: 0.00 (0.02398) [-0.17922], 1.00 (0.04753) [20.9513], -0.18 (0.07793) [-2.32893], -0.21 (0.04153) [-5.04640], -0.52 (0.04164) [-12.5521], 0.03 (0.0211) [1.22017]
    - Curvature(-1) row: -0.03 (0.00872) [-2.91824], -0.06 (0.01728) [-3.75630], 0.83 (0.02834) [29.4552], 0.02 (0.0151) [1.26505], 0.05 (0.01514) [3.27703], -0.02 (0.00767) [-2.02061]
  - Interpretation: diagonal elements significant; 12 out of 30 off-diagonal elements appear insignificant; lower-left 3x3 block (macroeconomic→yield factors) has five insignificant coefficients; upper-right 3x3 block (yield factors→macroeconomic) has three insignificant coefficients.

### Impulse response findings (qualitative and exact linkages preserved)
- Four groups of impulse responses analyzed: (i) macro variables to macro shocks; (ii) macro variables to yield shocks; (iii) yield curve to macro shocks; (iv) yield curve to yield shocks.
- Responses of macro variables to macro shocks:
  - Macroeconomic variables show significant persistence except for FFR (ܴܨܨ).
  - An increase in FFR lowers CU over following years; FFR rises with CU and INF consistent with an estimated monetary policy reaction function.
  - INF increases with CU and has a negative delayed response to FFR.
- Responses of macro variables to yield curve shocks:
  - Macro variables show mostly insignificant responses to curvature shocks.
  - An increase in the slope produces an important decline in FFR, indicating a close relation between the slope factor and the monetary policy instrument.
  - An increase in the level raises CU, FFR, and INF; during the period Fed accommodated only a fraction of expected inflation by increasing nominal funds rate, reducing CU, and limiting increase in INF.
- Responses of the yield curve to macro shocks:
  - Macroeconomic shocks have little influence on curvature.
  - Positive shock to FFR quickly increases the slope (making the yield curve steeper or less negatively sloped).
  - Positive shocks to CU, and to a lesser degree INF, produce inverse and delayed responses in slope, suggesting bond market anticipates monetary tightening.
  - Shocks to INF and CU generate a prolonged increase in the level factor, suggesting long-term inflation expectations may not be firmly anchored.
  - Positive shock to FFR induces a small temporary increase in the level factor.
- Responses of the yield curve to yield shocks:
  - The three yield curve factors display significant persistence.
  - A shock to the level increases the slope (associated with lowering of the short end relative to long end and a loosening of monetary policy).
  - A surprise increase in the slope reduces the level, shifting the entire curve downwards but with a relatively stronger decline in the short end.

### Conclusions and proposed software extensions (preserved conclusions)
- Main empirical conclusions:
  - The U.S. yield curve (1972:1-2007:12) exhibits large variation across maturities that can be characterized by three factors: level, slope, curvature.
  - A three-factor Nelson-Siegel model reproduces stylized facts and provides a framework for dynamic evolution, but meaningful interpretation requires linking yield curve factors with macroeconomic variables.
  - A six-factor Nelson-Siegel model fits the U.S. yield curve well and provides information on macro-financial linkages; market yields contain relevant predictive information about the Fed’s policy rate.
- Software notes:
  - Fund staff developed software (MCM Term Structure Software) to estimate at least nine term structure models, focusing on Nelson-Siegel and Affine-Term Structure models; uses C# for solving models and a friendly Excel interface.
- Proposed extensions:
  - Multi-country/global yield curve analysis: construct hierarchical dynamic factor models with global and country-specific factors (Diebold, Li, and Yue (2008) approach), using state-space and Bayesian MCMC methods to estimate large-parameter multi-country models.
  - Introduction of no-arbitrage restrictions: extend Nelson-Siegel models to arbitrage-free (affine arbitrage-free Nelson-Siegel) specifications (Christensen, Diebold, and Rudebusch (2007)) to integrate absence of arbitrage while maintaining Nelson-Siegel factor structure.

*Source: _wp10258 - Appendix I summarizes the (IMF staff estimates and text from the provided PDF content).*

### 1. Q is equivalent to P in the sense that, for any event E,

### _wp10258 - 1. Q is equivalent to P in the sense that, for any event E,

### Equivalent martingale measure: definition and implications
- Q is equivalent to P in the sense that, for any event E, ( ) 0() 0PEQE .  
  - In other words, P and Q assign zero probability to the same events.
- A measure Q with the two properties below is called an equivalent martingale measure:
  - 1) Q is equivalent to P;
  - 2) Discounted security prices are martingales under Q.
- In a continuous-time setting, the existence of an equivalent martingale measure guarantees the absence of arbitrage (the “First Fundamental Theorem of Asset Pricing”).

### Martingale property of discounted security prices (pricing equations)
- Definitions and processes:
  - Let tS be the price at time t of a security that does not pay any dividend or coupon.
  - Let tr be the short rate at time t.
  - Define the discount factor process tD as (2) 0 exp t u u Drdu       .
  - The discounted security price process is tt DS.
- Martingale property under Q:
  - Under Q, tt DS is a martingale. Therefore, for any 0 (3) ( ) Q tt ttt EDS DS   .
  - Using definition of tD, equality can be rewritten as (4) 00 exp exp tt Q tutut uu Erdu Srdu S              .
  - Hence (5) exp t Q tutt ut Erdu SS            .
  - Alternative expression: (6) ( ) 0 Q ttt EdDS    .
- When tr and tS follow Ito processes, the martingale condition can be expressed as (7) ( ) ( ) 00 QQ tt ttt tt tttt ErDSdtDdSED rSdtdS         .
- Since 0 tD, we obtain (8) ( ) Q tttt EdSr S dt  .
  - Therefore, under Q the instantaneous expected rate of return on any security equals the short rate; Q is often called the “risk-neutral” measure.
- Equations (5) and (6) provide general pricing equations under AOA and will be applied to zero-coupon bonds to determine the term-structure of interest rates.

### Existence of an equivalent martingale measure (construction via Girsanov)
- Setup and assumptions for construction:
  - Fixed horizon T.
  - Uncertainty generated by a single one-dimensional Brownian motion P W.
- Conditions on the process :
  - Consider process  such that (9) 2 0 2 0 a.s. 1 exp 2 T t t T P t t dt Edt                .
- Martingale candidate under P:
  - The process (10) 2 00 1 exp tt P tuuu uu dWdu         is a martingale under P.
  - Define Q by (11) T dQ dP   (Radon-Nikodym derivative).
- Girsanov’s Theorem and Brownian motion under Q:
  - Define process Q t W via (12) 0 t QP ttu u WWdu    .
  - When  satisfies the integrability condition above, Q t W is a Brownian motion under Q (Girsanov’s Theorem) as in (13) 0 t QP ttu u WWdu    .
- Application to a geometric Brownian price S under P:
  - Under P, (14) P tStSt t dSS dtS dW .
  - Discounted price process under P derived in (15) (equation chain) leads to expression with drift and diffusion terms.
  - Girsanov’s change yields discounted dynamics under Q in (16) ( ) ( ) Q ttS t ttS tttt Q St St ttSttt dDSr DSdtDS dWdt rDSdtDSdW                 .
- Martingale (zero-drift) condition under Q:
  - Require drift of DS to be zero under Q, leading to condition (17) ( ) 0 St St tt rDSdt        .
  - Equivalently, for any security price process S, the market price of risk  must satisfy (18) 0 St St Stt S r r              .
- Summary of construction:
  - The probability measure Q defined by (19) 2 00 1 exp Tt P Tttt tt dQ dWdt dP             , where  is the market price of risk process, is equivalent to P and makes discounted security prices martingales under Q.

### Equivalent martingale measures and Monte-Carlo simulations
- Interpretation and practical guidance:
  - Q does not represent real-world uncertainty; it is an artificial risk-neutral probability measure.
  - Filtration generated by P W represents real-world information revelation; filtration generated by Q W represents artificial risk-neutral information revelation.
  - Monte-Carlo simulations under Q are appropriate when computing expectations under Q (pricing applications), because arbitrage-free security prices are expectations under Q (equation (5)).
  - Monte-Carlo simulations under Q are not appropriate for computing VaR of a portfolio:
    - Q and P are only equivalent in the sense of assigning zero probability to the same events; distributions of portfolio values under Q and P differ.
    - VaR is a quantile of the real-world distribution; the relevant measure for VaR is P, not Q.

### Single-factor term-structure models and no-arbitrage PDE
- Model setup:
  - Let ( ) , Bt  be the price at time t of a zero-coupon bond that pays one unit at time T t   .
  - Let tx be a scalar stochastic process. A one-factor term-structure model is a function b such that (20) ( ) ( ) ( ) , , , , , 0 t B t b x t t e t          , where  is a parameter vector and ( ) , e t  is a pricing error.
  - Pricing errors assumed null for remainder: ( ) , 0 e t   for any t and .
- Dynamics under Q:
  - Assume x under Q follows Ito process (21) , , Q t x Q x Q t dxdtdW .
  - Applying Ito’s lemma to b yields (22) ( ) 2 2 , , , 2 1 , , , 2 Q tx Q x Q x Q t t t b b b db x t T d t dW txxx                         .
- No-arbitrage (Backward Kolmogorov) PDE:
  - Under Q the instantaneous expected rate of return on any security equals tr, so drift of b must equal rt b: (23) ( ) 2 2 , , , 2 1 , , , 2 x Q x Q t t b b r b x t T txx                 .
  - Single-factor models often take x to be the short rate: (24) tt x r t  .
  - Then PDE becomes (25) ( ) 2 2 , , , 2 1 , , , 2 r Q r Q t t b b r b x t T tr r                 .
  - This stochastic PDE is the Backward Kolmogorov equation (also called the “no-arbitrage” pricing equation); solution b is the term structure function under AOA.
- Boundary condition and solution representation:
  - Boundary condition: (26) ( ) , , , 1 t b r T T   .
  - Feynman-Kac representation: (27) ( ) , , , exp T Q t t uu ut br t T E r d             .
  - Equation (27) is equation (5) applied to a security paying one unit at time T.
  - Depending on the r process, closed-form solutions for zero-coupon bond prices may exist (e.g., mean-reverting or mean-reverting square-root processes as in the CIR model).
- Steps to derive single-factor arbitrage-free models:
  - 1) Assume existence of an equivalent martingale measure Q;
  - 2) Existence of Q guarantees AOA;
  - 3) Assume the single factor driving term structure follows an Ito process under Q;
  - 4) Use Ito’s lemma to derive drift of zero-coupon bond price under Q;
  - 5) Impose that instantaneous expected return under Q equals short rate to get a second expression for the drift;
  - 6) Combine the two drift expressions to obtain the no-arbitrage PDE.

*Italic: Source: _wp10258 (PDF chapter/section provided).*

### 7. The solution to the no-arbitrage PDE is the term structure function under AOA.

### 7. The solution to the no-arbitrage PDE is the term structure function under AOA.

### VI. THE COX-INGERSOLL-ROSS MODEL
- Model assumption:
  - Single factor: the short rate.
  - Short rate follows a mean-reverting square root process.
- Short rate process under P (equation (28)):
  - dr_t = κ(θ - r_t) dt + σ sqrt(r_t) dW_t^P (notation as in source: κ, θ, σ; equation number (28) shown in text).
- Short rate path from time 0 with r_0 > 0 (equation (29)): integral representation given in the source.
- Positivity conditions:
  - If κ > 0 and θ > 0, the short rate is non-negative.
  - If 2·κ·θ ≥ σ^2, the short rate is strictly positive at any time.
- Short rate under martingale measure Q (equation (30)):
  - dr_t = [κ(θ - r_t) - λσ sqrt(r_t)] dt + σ sqrt(r_t) dW_t^Q (λ is the market price of risk).
- Parameter vector to estimate in CIR:
  - β = (κ, θ, σ, λ).
- Closed-form bond price / term structure solution:
  - Bond price has exponential-affine form (equation (31)):
    - b(t, r, τ; β) = exp[A(τ; β) - C(τ; β) · r_t]
  - Functions A(·) and C(·) specified in (32) (detailed expressions in source).
  - Note: in CIR, b does not depend on calendar time t explicitly and can be written b(r, τ; β).
- Continuously compounded yield y(t, τ; β) (equations (33)–(34)):
  - Defined via exp[ - y(t, τ; β) · τ ] = b(t, r, τ; β).
  - Equivalent expression: y(t, τ; β) = [A(τ; β)/τ] + C(τ; β) · r_t (see equation (34)).
- Classification:
  - CIR is an affine term-structure model because yields are affine functions of r_t.

### VII. AFFINE MULTI-FACTOR MODELS
- Two-factor bond price notation:
  - b_t = b(x_{1,t}, x_{2,t}, τ) = b_t^{1,2}(x_{1,t}, x_{2,t}, τ) (see (35)).
- Factor dynamics under Q (equation (36)):
  - dx_{1,t} = μ_1^Q(x_{1,t}, x_{2,t}) dt + σ_1^Q(x_{1,t}, x_{2,t}) dW_{1,t}^Q
  - dx_{2,t} = μ_2^Q(x_{1,t}, x_{2,t}) dt + σ_2^Q(x_{1,t}, x_{2,t}) dW_{2,t}^Q
  - W_1^Q and W_2^Q independent Brownian motions under Q.
- No-arbitrage PDE for two-factor bond prices (equation (39)):
  - PDE derived from martingale condition on discounted bond price D_t b_t, coefficient of dt must be zero; full PDE shown in (39).
- Affine diffusion assumption (equation (40)):
  - μ_1, μ_2, σ_1^2, σ_2^2 are affine functions of factors:
    - μ_1 = ε_1 + δ_{11} x_1, μ_2 = ε_2 + δ_{22} x_2
    - σ_1^2 = ω_{11} + λ_{11} x_1, σ_2^2 = ω_{22} + λ_{22} x_2
- Short rate affine in factors (equation (41)):
  - r_t = γ_0 + γ_1 x_{1,t} + γ_2 x_{2,t}
- Exponential-affine bond price conjecture (equation (42)):
  - b(x_1,x_2,t,T) = exp[α(τ) + β_1(τ) x_1 + β_2(τ) x_2], where τ = T - t.
  - α(·), β_1(·), β_2(·) deterministic functions of τ.
- Reduction of PDE to ODEs:
  - Substituting exponential-affine form into PDE reduces it to three ordinary differential equations (equations (43)–(46)).
  - The ODEs for β_1 and β_2 are Riccati equations.
  - Boundary conditions (equation (47)):
    - α(0) = 0, β_1(0) = 0, β_2(0) = 0 (ensuring b(x_1,x_2,T,T) = 1).
- Two-factor CIR special case (equation (48)):
  - Each factor follows mean-reverting square-root (CIR) diffusion.
  - Parameters κ_i, θ_i, σ_i (i = 1,2) with positivity constraints as in one-factor CIR.

### APPENDIX II. ESTIMATION TECHNIQUES — IMPLEMENTATION OVERVIEWS
- High-level approach:
  - MCMTS uses a one-stage Powell method solution to estimate Nelson-Siegel and CIR models.
  - Iterative optimization proceeds until tolerance constraints are met (Powell’s method referenced; see Powel (1964) in source footnote).
- Nelson-Siegel family (implementation overview):
  - Key issue: decay parameters may be fixed or time-varying.
  - MCMTS uses fixed decay parameters to provide stable state-space framework; requires joint optimization (cross-section and time series).
  - Estimation steps (one-stage with Powell):
    - Cross-Sectional Optimization:
      - Tool: OLS
      - Input: Cross-sectional yield data
      - Output: Time series of factors
    - Time Series Optimization:
      - Tool: Auto-regression or Vector auto-regression (AR(1) or VAR(1))
      - Input: Time series of factors
      - Output: Parameters of state-space model
    - Joint Analysis:
      - Tool: Kalman Filter
      - Input: Initial time series factors and state-space model parameters
      - Output: Adjusted time series factors and likelihood value
  - Covariance treatment:
    - Covariance matrix of measurement errors computed by OLS from cross-sectional step.
    - Covariance matrix of state innovation errors computed by AR(1) or VAR(1) step.
    - Result: unnecessary to estimate H and Q directly, reducing parameter count and aiding convergence.
  - Empirical fit:
    - Example: RMSE typically less than 5 basis points when estimating Nelson-Siegel models (statement from source).
- Nelson-Siegel algorithm description (high-level steps, preserved notation):
  - For given λ, compute factor loadings x_{i,τ} for each maturity τ:
    - x_{1,τ} = 1
    - x_{2,τ} = (1 - exp(-λ τ)) / (λ τ)
    - x_{3,τ} = x_{2,τ} - exp(-λ τ)
  - Form measurement matrix X(τ) from x_{i,τ}.
  - For each t ∈ {1,...,T} run OLS on Y_t = X_t β_t + ε_t, with ρ = 0 in source notation, obtain β_t and ε_t; compute H from ε_t series.
  - Run AR(1)/VAR(1) on β_t to get μ, Φ, ν_t; compute Q from ν_t series.
  - Kalman Filter to compute adjusted factors and log-likelihood L:
    - L = Σ_{t=1}^T [ -1/2 ln | f_{t|t-1} | -1/2 η_{t|t-1}' f_{t|t-1}^{-1} η_{t|t-1} ] (source notation used for filter quantities; see step (4) and likelihood expression).
- CIR model estimation overview:
  - MCMTS uses one-stage Powell method for CIR.
  - Iterative steps:
    - Cross-Sectional Optimization:
      - Tool: OLS
      - Input: cross-sectional yields
      - Output: time series of factors
    - Compute parameters of state-space model (depends on P)
    - Joint Analysis:
      - Tool: Kalman Filter
      - Input: initial factors and state-space parameters
      - Output: adjusted factors and likelihood value
  - Covariance treatment:
    - Measurement error covariance H computed by OLS.
    - State innovation covariance Q dynamically calculated at each Kalman time step because Q_jj depends on β_{j,t-1}.
- CIR algorithm description (two-factor example, preserved notation):
  - Parameter vector P = [λ_1, σ_1, θ_1, κ_1; λ_2, σ_2, θ_2, κ_2]′ (source notation shows ordering).
  - Measurement constants and coefficients:
    - ρ_i = -Σ_{j=1}^K ln a_{ij}(τ)
    - x_{ij}(τ) = ln b_{ij}(τ)
    - Form X(τ) matrix and run OLS Y_t = X_t β̃_t + ε_t to obtain β_t and ε_t; compute H from ε_t.
  - State equation for given P:
    - β_t = μ + Φ β_{t-1} + ν_t with μ_j = (1 - exp(-Δ κ_j)) θ_j and Φ diagonal with elements exp(-Δ κ_j).
    - Var(ν_{j,t}) = (1/2 κ_j) σ_j^2 [1 - exp(-2 Δ κ_j)] β_{j,t-1} (source gives precise function and notes dependence on β_{j,t-1}).
  - Kalman Filter with dynamically updated Q_jj; compute log-likelihood L as in Nelson-Siegel case.
- Estimation analysis summary (Appendix Table 1 values preserved):
  - Nelson-Siegel model with one decay parameter:
    - Number of Parameters: 1
    - Parameters: λ
    - Number of Iterations before Convergence: <10
  - Nelson-Siegel model with two decay parameters:
    - Number of Parameters: 2
    - Parameters: λ_1, λ_2
    - Number of Iterations before Convergence: <50
  - One-factor Cox-Ingersoll-Ross model:
    - Number of Parameters: 4
    - Parameters: λ, σ, θ, κ (notation in source: λσθκ ,,,)
    - Number of Iterations before Convergence: <200
  - Two-factor Cox-Ingersoll-Ross model:
    - Number of Parameters: 8
    - Parameters: λ_1, σ_1, θ_1, κ_1 ; λ_2, σ_2, θ_2, κ_2 (source notation: 1111 ,,,, λσθκ and 2222 ,,,, λσθκ)
    - Number of Iterations before Convergence: <2,000
  - Source: Fund staff estimates (as stated in source).

### IV. MONTE CARLO SIMULATIONS (state-space framework)
- Simulation framework:
  - Factor dynamics driven by AR(1) or VAR(1).
  - User sets number of simulated paths N and number of steps S.
  - After generating paths, MCMTS computes summary statistics (mean, median, standard deviation, percentiles) of interest rates at final time and draws histograms and displays simulated paths.
- Nelson-Siegel simulation approach:
  - Generate K i.i.d. standard normal random variables ω ∼ N(0, I).
  - Obtain joint normal with covariance Q via Cholesky: C = Cholesky(Q), set q = C ω so q ∼ N(0, Q).
  - Use draws q to drive state-space evolution and produce simulated factor and yield series.
- CIR simulation approach:
  - Use Euler discretization of square-root diffusion:
    - β_{j,t+Δ} = β_{j,t} + κ_j(θ_j - β_{j,t}) Δ + σ_j sqrt(β_{j,t} Δ) z_{j,t}, with z_{j,t} ∼ N(0, Δ) (source gives discretized Euler scheme in notation jt...; preserved qualitative form).
  - Simulated interest rates computed from measurement equation.

### APPENDIX III. DESCRIPTION OF THE MCM TERM STRUCTURE SOFTWARE (MCMTS)
- Capabilities:
  - Create data sets (yield curves and macro variables).
  - Specify time horizons.
  - Select term-structure models (Nelson-Siegel family, CIR family, Macro-Factor Models).
  - Estimate model parameters and evaluate estimation performance.
  - Run Monte Carlo simulations.
  - Estimate impulse responses and variance decomposition using a Macro Factor Model.
- Implementation notes:
  - MCMTS implemented as an Excel add-in with C# algorithms (source states: algorithms developed in C#).
- User workflow highlights:
  - Data required:
    - "Yield Curve" sheet: zero-coupon yields by maturity over time.
    - "Macro Factors" sheet: time series of macro variables (needed for macro-factor models).
  - Frequency options: daily, weekly, monthly.
  - Simulation horizon defaults:
    - Daily: one year after end date
    - Weekly: three years after end date
    - Monthly: ten years after end date
  - Nelson-Siegel model options:
    - Five alternatives: three-factor base; Bjork and Christensen four-factor; Bliss three-factor; Svensson four-factor; Adjusted Svensson four-factor.
    - Factor relationship options: Independent Factors, AR(1) or Correlated Factors, VAR(1).
    - Optimization methods: No Optimization; Minimize Measurement RMSE; Maximum Likelihood with Kalman Filter.
  - CIR model options:
    - One-factor or two-factor CIR.
  - Model evaluation features:
    - Cross-sectional yield fitting (compare observed vs estimated on a chosen date).
    - Time-series yield fitting (compare observed vs estimated for a chosen maturity).
    - Chi-square test of overall fit.
  - Monte Carlo and diagnostics:
    - User chooses number of simulations and paths to display; MCMTS presents distributions and summary statistics.
    - Impulse-response and variance-decomposition analysis available for macro-factor models; user chooses shocks, responding variables, and horizon.
- Practical notes:
  - Two-factor models may fit yields better but take longer to estimate (statement preserved from source).

*Source: _wp10258 - 7. The solution to the no-arbitrage PDE is the term structure function under AOA.*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2010/_wp10258.pdf_