多元卡尔曼滤波器数组赋值与协变量引入技术问询

Hey there! Let's work through your multivariate Kalman Filter issue and get that model capturing cross-variable dependencies like it should.

First, Why Your Current Model Acts Like Separate Univariate Models #

The main problem with your code is how you've set up the trend component and noise matrices:

• When using SSMtrend(1) with a 6-variate time series, you're telling KFAS to use a single shared trend across all 6 variables, not individual trends for each variable.
• Your Har array is a 6x6x100 matrix filled with 0.01, but since the state dimension is only 1 (from SSMtrend(1)), KFAS can't properly interpret the 6x6 Q matrix. This effectively forces the model to treat each variable independently, hence matching univariate results.

Fixing the Multivariate Structure #

To model dependent trends across variables, we need to:

1. Assign individual trends to each variable
2. Use non-diagonal covariance matrices for state noise (Q) and observation noise (H) to capture cross-variable dependencies

Here's revised code with these fixes:

library("KFAS")
set.seed(123) # For reproducibility
n <- 100

# Generate your original data
dx <- data.frame(
  a1=cumsum(sample(c(-1, 1), n, TRUE)),
  a2=cumsum(sample(c(-1, 1), n, TRUE)),
  a3=cumsum(sample(c(-1, 1), n, TRUE)),
  a4=cumsum(sample(c(-1, 1), n, TRUE)),
  a5=cumsum(sample(c(-1, 1), n, TRUE)),
  a6=cumsum(sample(c(-1, 1), n, TRUE))
)
mx <- as.matrix(dx)

# 1. Create a positive-definite state noise covariance matrix (Q)
# This captures dependencies between the trends of different variables
Q_base <- matrix(runif(6*6, -0.005, 0.005), nrow=6)
Q_cov <- t(Q_base) %*% Q_base + diag(0.01, 6) # Ensure matrix is positive-definite
Q_mat <- array(Q_cov, dim = c(6, 6, n)) # Replicate for time-varying (or use static if preferred)

# 2. Create a positive-definite observation noise covariance matrix (H)
H_base <- matrix(runif(6*6, -0.002, 0.002), nrow=6)
H_cov <- t(H_base) %*% H_base + diag(0.005, 6)
H_mat <- array(H_cov, dim = c(6, 6, n))

# 3. Build the multivariate model with individual trends per variable
model1 <- SSModel(
  mx ~ SSMtrend(rep(1, 6), Q = Q_mat), # rep(1,6) = 1st-order trend for each of 6 variables
  H = H_mat
)

dkf <- KFS(model1)
dkf

Adding Covariates to Capture Dependencies #

If you want to include covariates to model cross-variable relationships, you have two options:

Option 1: Fixed-Effect Covariates

Add covariates directly to the model formula (treats covariate coefficients as fixed):

# Generate example covariates
dx$x1 <- rnorm(n) # Covariate 1
dx$x2 <- rnorm(n) # Covariate 2

# Create covariate matrix (exclude intercept since trend handles level)
cov_matrix <- model.matrix(~ x1 + x2, data = dx)[, -1]

# Build model with fixed-effect covariates
model_cov_fixed <- SSModel(
  mx ~ cov_matrix + SSMtrend(rep(1, 6), Q = Q_mat),
  H = H_mat
)

dkf_fixed <- KFS(model_cov_fixed)
dkf_fixed

Option 2: Random-Effect Covariates

Use SSMregression to model covariate coefficients as random (great if you expect coefficients to vary over time):

# Build model with random-effect covariates
model_cov_random <- SSModel(
  mx ~ SSMtrend(rep(1, 6), Q = Q_mat) + 
    SSMregression(cov_matrix, Q = diag(0.01, ncol(cov_matrix))), # Q for random coefficients
  H = H_mat
)

dkf_random <- KFS(model_cov_random)
dkf_random

Key Takeaways #

• Use SSMtrend(rep(1, n_vars)) to assign individual trends to each variable
• Use non-diagonal Q and H matrices to model cross-variable dependencies
• Covariates can be added as fixed effects (direct formula) or random effects (SSMregression) to enhance model performance over univariate setups

内容的提问来源于stack exchange，提问作者adam.888