R/kim.R

In guilbran/dynfactoR: Dynamic factor model estimation for nowcasting

#' Implementation of Kim (1994) filter, an extension to Kalman filter
#' for dynamic linear models with Markov-switching. Documentation
#' is incomplete, rudimentary and needs to be rechecked!
#'
#' @param x0 Initial condition for state vector
#' @param P0 Initial condition for state variance
#' @param y Data matrix (Txn)
#' @param F System matrix for measurement equation
#' @param A Transition matrix for state equation
#' @param R Error covariance for measurement equation
#' @param Q Error covariance for state equation
#' @param p Transition probability matrix
#' @return Filtered states and covariances with associated probability matrices.
KimFilter <- function(x0, P0, y, F, A, R, Q, p)
{
  # Define all containers for further computations. Notations for variables and indices,
  # where appropriate, carefully follow Kim (1994). State vector is denoted as 'x', its
  # covariance as 'P'. Appended letters explicit whether these are updated, approximated
  # or smoothed.

  if (is.vector(y) || (ncol(y) <= length(x0)))
  {
    stop("Number of factors should be strictly lower than number of variables. \n
          Increase number of variables or estimate a VAR model instead.")
  }

  T <- nrow(y)
  n <- dim(F)[1]
  J <- length(x0)
  s <- dim(p)[1]

  ## x:   x^(i,j)_(t|t-1): predicted state vector - (2.6)
  ## xU:  x^(i,j)_(t|t): updated state vector - (2.11)
  ## P:   P^(i,j)_(t|t-1): predicted state covariance - (2.7)
  ## Pu:  P^(i,j)_(t|t): updated state covariance - (2.12)
  ## eta: eta^(i,j)_(t|t-1): conditional forecast error - (2.8)
  ## H:   H^(i,j)_(t): conditional variance of forecast error - (2.9)
  ## K:   K^(i,j)_(t): Kalman gain - (2.10)
  ## lik: f(y_t, S_(t-1)=i, S_t = j | t-1): joint conditional density - (2.16)
  ## loglik: log of (2.16)
  x <- array(NA, c(T,J,s,s))
  xU <- array(NA, c(T,J,s,s))
  P <- array(NA, c(T,J,J,s,s))
  Pu <- array(NA, c(T,J,J,s,s))
  eta <- array(NA, c(T,n,s,s))
  H <- array(NA, c(T,n,n,s,s))
  K <- array(NA, c(T,J,n,s,s))
  lik <- array(NA, c(T,s,s))
  loglik <- array(NA, c(T,s,s))
  ## Pr[S_(t-1) = i, S_t = j | t-1 ]: (2.15)
  ## Pr[S_(t-1) = i, S_t = j | t ]: (2.17)
  ## Pr[S_t = j | t-1 ]: used only for the smoothing part
  ## Pr[S_t = j | t ]: (2.18)
  jointP_fut <- array(NA, c(T,s,s))
  jointP_cur <- array(NA, c((T+1),s,s))
  stateP_fut <- array(NA, c(T,s))
  stateP <- array(NA, c(T,s))

  ## x^(j)_(t|t): approximate state vector conditional on S_j - (2.13)
  ## P^(j)_(t|t): approximate state covariance conditional on S_j - (2.14)
  xA <- array(NA, c(T,J,s))
  Pa <- array(0, c(T,J,J,s))
  result <- array(0, c(T,1))

  # Some initial conditions to get started
  for (i in 1:s) { xA[1,,i] <- x0 }
  for (i in 1:s) { Pa[1,,,i] <- P0 }
  jointP_cur[1,,] <- matrix(c(0.25,0.25,0.25,0.25), ncol=2)

  for (t in 2:T)
  {
    for (j in 1:s)
    {
      for (i in 1:s)
      {
        x[t,,i,j] <- A[,,j] %*% xA[(t-1),,i]
        P[t,,,i,j] <- A[,,j] %*% Pa[(t-1),,,i] %*% t(A[,,j]) + Q
        eta[t,,i,j] <- y[t,] - as(F[,,j], "matrix") %*% x[t,,i,j]
        H[t,,,i,j] <- F[,,j] %*% as(P[t,,,i,j], "matrix") %*% t(F[,,j]) + R
        K[t,,,i,j] <- P[t,,,i,j] %*% t(F[,,j]) %*% solve(H[t,,,i,j])
        xU[t,,i,j] <- x[t,,i,j] + K[t,,,i,j] %*% eta[t,,i,j]
        Pu[t,,,i,j] <- (diag(1,J) - K[t,,,i,j] %*% F[,,j]) %*% P[t,,,i,j]
        jointP_fut[t,i,j] <- p[i,j]*sum(jointP_cur[(t-1),,i]) # is everything alright here?
        lik[t,i,j] <- (2*pi)^(-n/2) * det(H[t,,,i,j])^(-1/2) *
                      exp(-1/2*t(eta[t,,i,j]) %*% solve(H[t,,,i,j]) %*% eta[t,,i,j]) *
                      jointP_fut[t,i,j]
        loglik[t,i,j] <- log(lik[t,i,j])
        jointP_cur[t,i,j] <- lik[t,i,j]
      }
      # Technically, there should be sum(lik[t,,]) term but it cancels out and is computed later
      stateP[t,j] <- sum(jointP_cur[t,,j])
      stateP_fut[t,j] <- sum(jointP_fut[t,,j])
      # Compute probability-filtered state process and its covariance
      xA[t,,j] <- xU[t,,,j] %*% jointP_cur[t,,j] / stateP[t,j]
      for (i in 1:s)
      {
        Pa[t,,,j] <- Pa[t,,,j] +
                     (Pu[t,,,i,j] + (xA[t,,j] - xU[t,,i,j]) %*% t(xA[t,,j] - xU[t,,i,j])) * 
                     exp(log(jointP_cur[t,i,j]) - log(stateP[t,j]))
      }
    }
    jointP_cur[t,,] <- exp(log(jointP_cur[t,,]) - log(sum(lik[t,,])))
    stateP[t,] <- exp(log(stateP[t,]) - log(sum(lik[t,,])))
    result[t,1] <- log(sum(lik[t,,]))
  }

  return(list("result"=sum(result), "xA"=xA, "Pa"=Pa, "x"=x, "P"=P, "stateP"=stateP, "stateP_fut"=stateP_fut))
  
}

#' Smoothing algorithm from Kim (1994) to be used following a run
#' of KimFilter function.
#'
#' @param xA Filtered state vector to be smoothed
#' @param Pa Filtered state covariance to be smoothed
#' @param x State-dependent state vector
#' @param P State-dependent state covariance
#' @param A Array with transition matrices
#' @param p Markov transition matrix
#' @param stateP Evolving current probability matrix
#' @param stateP_fut Predicted probability matrix
#' @return Smoothed states and covariance matrices. This is the equivalent
#' of Kalman smoother in Markov-switching case.
KimSmoother <- function(xA, Pa, A, P, x, p, stateP, stateP_fut)
{
  # Define all containers for further computations. Notations for variables and indices,
  # where appropriate, carefully follow Kim (1994). State vector is denoted as 'x', its
  # covariance as 'P'. Appended letters explicit whether these are updated, approximated
  # or smoothed.

  T <- dim(xA)[1]
  J <- dim(xA)[2]
  s <- dim(xA)[3]

  ## Pr[S_t = j, S_(t+1) = k | T]: (2.20)
  ## Pr[S_t = j | T]: (2.21)
  jointPs <- array(NA, c(T,s,s))
  ProbS <- array(NA, c(T,s))

  ## xS: x^(j,k)_(t|T): inference of x_t based on full sample - (2.24)
  ## Ps: P^(j,k)_(t|T): covariance matrix of x^(j,k)_(t|T) - (2.25)
  ## Ptilde: helper matrix as defined after (2.25) 
  xS <- array(0, c(T,J,s,s))
  Ps <- array(0, c(T,J,J,s,s))
  Ptilde <- array(NA, c(T,J,J,s,s))

  ## xAS: x^(j)_(t|T): smoothed and approximated state vector conditional on S_j (2.26)
  ## Pas: P^(j)_(t|T): smoothed and approximated state covariance conditional on S_j (2.27)
  ## xF: x_(t|T): state-weighted [F]inal state vector (2.28)
  ## Pf: P_(t|T): state-weighted [f]inal state covariance
  xAS <- array(0, c(T,J,s))
  Pas <- array(0, c(T,J,J,s))
  xF <- array(0, c(T,J))
  Pf <- array(0, c(T,J,J))
  # Initial conditions for smoothing loop
  ProbS[T,] <- stateP[T,]

  for (t in seq(T-1,1,-1))
  {
    for (j in 1:s)
    {
      for (k in 1:s)
      {
        jointPs[t,j,k] <- ProbS[(t+1),k]*stateP[t,j]*p[j,k] / stateP_fut[(t+1),k]
        Ptilde[t,,,j,k] <- Pa[t,,,j] %*% t(A[,,k]) %*% solve(P[(t+1),,,j,k])
        xS[t,,j,k] <- xA[t,,j] + Ptilde[t,,,j,k] %*% (xA[(t+1),,k] - x[(t+1),,j,k])
        Ps[t,,,j,k] <- Pa[t,,,j] +
                       Ptilde[t,,,j,k] %*% (Pa[(t+1),,,k] - P[(t+1),,,j,k]) %*% t(Ptilde[t,,,j,k])
        xAS[t,,j] <- xAS[t,,j] + jointPs[t,j,k]*xS[t,,j,k]
        Pas[t,,,j] <- Pas[t,,,j] + jointPs[t,j,k]*(Ps[t,,,j,k] +
                                          (xAS[t,,j] - xS[t,,j,k]) %*% t(xAS[t,,j] - xS[t,,j,k]))
      }
      ProbS[t,j] <- sum(jointPs[t,j,])
      xAS[t,,j] <- xAS[t,,j] / ProbS[t,j]
      Pas[t,,,j] <- Pas[t,,,j] / ProbS[t,j]
    }
  }
  for (t in 1:T)
  {
    for (j in 1:s)
    {
      xF[t,] <- xF[t,] + xAS[t,,j]*ProbS[t,j]
      Pf[t,,] <- Pf[t,,] + Pas[t,,,j]*ProbS[t,j]
    }
  }

  return(list("xF"=xF, "Pf"=Pf, "ProbS"=ProbS))
}