如何在Dynamic Nelson Siegel模型中纳入宏观经济因素？——R语言代码实现及数据格式求助

Hey there, I totally get the frustration of translating theoretical econometric models into code—especially when you're new to R and working with something as nuanced as the Dynamic Nelson-Siegel (DNS) model with macroeconomic variables. Let’s break this down step by step to get you unstuck.

First, let’s align on the data structure you need—this is half the battle:

• Yield data: A time-series panel where each row is a time period (e.g., monthly), and each column is a yield for a specific maturity (1m, 3m, 1y, 5y, 10y, etc.). Ensure no missing values (use interpolation or na.omit() if needed).
• Macro variables: A time-series aligned perfectly with your yield data (same time periods, same frequency). Common choices include inflation, GDP growth, or policy rates. Always standardize these variables (using scale() in R) because their magnitudes can differ drastically from DNS factors, which will throw off model estimation.

Here’s a quick example of how your data should look in R:

library(tibble)

# Simulated yield data (replace with your real data)
yield_data <- tibble(
  date = seq.Date(as.Date("2000-01-01"), by = "month", length.out = 240),
  m1 = rnorm(240, 3, 0.5),  # 1-month yield
  m3 = rnorm(240, 3.2, 0.6), # 3-month yield
  y1 = rnorm(240, 3.5, 0.7), # 1-year yield
  y5 = rnorm(240, 4.2, 0.8), # 5-year yield
  y10 = rnorm(240, 4.5, 0.9) # 10-year yield
)

# Simulated macro data (replace with your real data)
macro_data <- tibble(
  date = yield_data$date,
  inflation = rnorm(240, 2, 0.3),
  gdp_growth = rnorm(240, 2.5, 0.4),
  ff_rate = rnorm(240, 3, 0.5)
)

# Merge and standardize
full_data <- merge(yield_data, macro_data, by = "date")
full_data[, -1] <- scale(full_data[, -1]) # Standardize all non-date columns

Diebold et al.’s core idea is straightforward once you map it to code:

1. Expand the state vector: The basic DNS state vector is [Level_t, Slope_t, Curvature_t]. Add your macro variables to make it [Level_t, Slope_t, Curvature_t, Macro1_t, Macro2_t, ..., MacroK_t].
2. Update the VAR components:
   • The mean vector μ grows from 3×1 to (3+K)×1 (add means for each macro variable).
   • The transition matrix Φ grows from 3×3 to (3+K)×(3+K) (include cross-effects between DNS factors and macro variables, e.g., how inflation affects the level factor over time).
   • The disturbance covariance matrix Σ_η grows to (3+K)×(3+K) to account for variance in both DNS and macro shocks.

Since you mentioned the statespacer package, this is the easiest way to avoid writing Kalman filter code from scratch. Here’s how to set up the extended DNS model:

library(statespacer)

# Fixed lambda parameter (standard value from Diebold et al.)
lambda <- 0.0609

# Maturities converted to years
maturities <- c(1, 3, 12, 60, 120) / 12

# Build loading matrix for yields: maps state vector to observed yields
yield_loadings <- t(sapply(maturities, function(tau) {
  c(
    1,
    (1 - exp(-lambda * tau)) / (lambda * tau),
    (1 - exp(-lambda * tau)) / (lambda * tau) - exp(-lambda * tau),
    0, 0, 0 # Zeros for macro variables (yields only depend on DNS factors)
  )
}))

# Build loading matrix for macro variables: each macro variable maps directly to its state
macro_loadings <- diag(c(0, 0, 0, 1, 1, 1)) # Positions 4-6 are macro states

# Combine all loadings (rows = observed variables, columns = state variables)
Z <- rbind(yield_loadings, macro_loadings)

# Initial guess for transition matrix Phi (start with persistent diagonal values)
Phi <- diag(6)
Phi[1:3, 1:3] <- matrix(0.9, 3, 3) # High persistence for DNS factors
Phi[4:6, 4:6] <- matrix(0.8, 3, 3) # Persistence for macro variables

# Initial guesses for other parameters
mu <- rep(0, 6) # State mean vector
Sigma_eta <- diag(rep(0.1, 6)) # State disturbance covariance
Sigma_epsilon <- diag(c(rep(0.01, 5), rep(0.001, 3))) # Measurement error (smaller for macro)

# Convert data to matrix (drop date column)
Y_matrix <- as.matrix(full_data[, -1])

# Estimate the model
model <- statespacer(
  y = Y_matrix,
  Z = Z,
  T_matrix = Phi,
  R = diag(6),
  Q = Sigma_eta,
  H = Sigma_epsilon,
  a1 = mu,
  P1 = diag(rep(1, 6)),
  estimate = TRUE # Let the package estimate all parameters
)

# Check results
summary(model)

• Convergence issues: If the model fails to converge, start with simpler initial guesses (e.g., a diagonal Phi matrix) or add one macro variable at a time instead of all at once.
• Missing data: Use tidyr::fill() or linear interpolation for small gaps, but avoid over-manipulating yield data.
• Standardization: Skipping this step is the #1 cause of wonky parameter estimates—always scale your data!

If you want to understand the Kalman filter mechanics, Huang’s manual implementation is a great reference:

• Start with his basic DNS code, then modify the state vector to include macro variables.
• Update the transition equation to account for cross-terms between DNS factors and macro variables.
• Adjust the Kalman filter’s prediction and update steps to handle the larger state vector.

For a beginner, though, statespacer is the best starting point—it handles the low-level Kalman filter work so you can focus on the model’s economic structure.

内容的提问来源于stack exchange，提问作者Gino Ramos