Fast Kalman filter

Share:

Description

This function allows for fast and flexible Kalman filtering. Both, the measurement and transition equation may be multivariate and parameters are allowed to be time-varying. In addition “NA”-values in the observations are supported. fkf wraps the C-function FKF which fully relies on linear algebra subroutines contained in BLAS and LAPACK.

Usage

1
fkf(a0, P0, dt, ct, Tt, Zt, HHt, GGt, yt, check.input = TRUE)

Arguments

a0

A vector giving the initial value/estimation of the state variable.

P0

A matrix giving the variance of a0.

dt

A matrix giving the intercept of the transition equation (see Details).

ct

A matrix giving the intercept of the measurement equation (see Details).

Tt

An array giving the factor of the transition equation (see Details).

Zt

An array giving the factor of the measurement equation (see Details).

HHt

An array giving the variance of the innovations of the transition equation (see Details).

GGt

An array giving the variance of the disturbances of the measurement equation (see Details).

yt

A matrix containing the observations. “NA”-values are allowed (see Details).

check.input

A logical stating whether the input shall be checked for consistency (“storage.mode”, “class”, and dimensionality, see Details).

Details

State space form:

The following notation is closest to the one of Koopman et al. The state space model is represented by the transition equation and the measurement equation. Let m be the dimension of the state variable, d be the dimension of the observations, and n the number of observations. The transition equation and the measurement equation are given by

alpha(t + 1) = d(t) + T(t) alpha(t) + H(t) * eta(t)

y(t) = c(t) + Z(t) alpha(t) + G(t) * epsilon(t),

where eta(t) and epsilon(t) are iid N(0, I(m)) and iid N(0, I(d)), respectively, and alpha(t) denotes the state variable. The parameters admit the following dimensions:

a[t] in R^m d[t] in R^m eta[t] in R^m
d[t] in R^(m * m) d[t] in R^(m * m)
y[t] in R^d c[t] in R^d epsilon[t] in R^d.
Z[t] in R^(d * m) G[t] in R^(d * d)

Note that fkf takes as input HHt and GGt which corresponds to H[t] %*% t(H[t]) and G[t] %*% t(G[t]).

Iteration:

Let i be the loop variable. The filter iterations are implemented the following way (in case of no NA's):

Initialization:
if(i == 1){
at[, i] = a0
Pt[,, i] = P0
}

Updating equations:
vt[, i] = yt[, i] - ct[, i] - Zt[,,i] %*% at[, i]
Ft[,, i] = Zt[,, i] %*% Pt[,, i] %*% t(Zt[,, i]) + GGt[,, i]
Kt[,, i] = Pt[,, i] %*% t(Zt[,, i]) %*% solve(Ft[,, i])
att[, i] = at[, i] + Kt[,, i] %*% vt[, i]
Ptt[, i] = Pt[,, i] - Pt[,, i] %*% t(Zt[,, i]) %*% t(Kt[,, i])

Prediction equations:
at[, i + 1] = dt[, i] + Tt[,, i] %*% att[, i]
Pt[,, i + 1] = Tt[,, i] %*% Ptt[,, i] %*% t(Tt[,, i]) + HHt[,, i]

Next iteration:
i <- i + 1
goto “Updating equations”.

NA-values:

NA-values in the observation matrix yt are supported. If particular observations yt[,i] contain NAs, the NA-values are removed and the measurement equation is adjusted accordingly. When the full vector yt[,i] is missing the Kalman filter reduces to a prediction step.

Parameters:

The parameters can either be constant or deterministic time-varying. Assume the number of observations is n (i.e. y = y[,1:n]). Then, the parameters admit the following classes and dimensions:

dt either a m * n (time-varying) or a m * 1 (constant) matrix.
Tt either a m * m * n or a m * m * 1 array.
HHt either a m * m * n or a m * m * 1 array.
ct either a d * n or a d * 1 matrix.
Zt either a d * m * n or a d * m * 1 array.
GGt either a d * d * n or a d * d * 1 array.
yt a d * n matrix.

If check.input is TRUE each argument will be checked for correctness of the dimensionality, storage mode, and class. check.input should always be TRUE unless the performance becomes crucial and correctness of the arguments concerning dimensions, class, and storage.mode is ensured.
Note that the class of the arguments if of importance. For instance, to check whether a parameter is constant the dim attribute is accessed. If, e.g., Zt is a constant, it could be a d * d-matrix. But the third dimension (i.e. dim(Zt)[3]) is needed to check for constancy. This requires Zt to be an d * d * 1-array.

BLAS and LAPACK routines used:

The R function fkf basically wraps the C-function FKF, which entirely relies on linear algebra subroutines provided by BLAS and LAPACK. The following functions are used:

BLAS: dcopy, dgemm, daxpy.
LAPACK: dpotri, dpotrf.

FKF is called through the .Call interface. Internally, FKF extracts the dimensions, allocates memory, and initializes the R-objects to be returned. FKF subsequently calls cfkf which performs the Kalman filtering.

The only critical part is to compute the inverse of F[,,t] and the determinant of F[,,t]. If the inverse can not be computed, the filter stops and returns the corresponding message in status (see Value). If the computation of the determinant fails, the filter will continue, but the log-likelihood (element logLik) will be “NA”.

The inverse is computed in two steps: First, the Cholesky factorization of F[,,t] is calculated by dpotrf. Second, dpotri calculates the inverse based on the output of dpotrf.
The determinant of F[,,t] is computed using again the Cholesky decomposition.

Value

An S3-object of class “fkf”, which is a list with the following elements:

att A m * n-matrix containing the filtered state variables, i.e. att[,t] = E(alpha[t] | y[,t]).
at A m * (n + 1)-matrix containing the predicted state variables, i.e. at[,t] = E(alpha[t] | y[,t - 1]).
Ptt A m * m * n-array containing the variance of att, i.e. Ptt[,,t] = var(alpha[t] | y[,t]).
Pt A m * m * (n + 1)-array containing the variances of at, i.e. Pt[,,t] = var(alpha[t] | y[,t - 1]).
vt A d * n-matrix of the prediction errors given by vt[,t] = yt[,t] - ct[,t] - Zt[,,t] %*% at[,t].
Ft A d * d * n-array which contains the variances of vt, i.e. Ft[,,t] = var(v[,t]).
Kt A m * d * n-array containing the “Kalman gain” (ambiguity, see calculation above).
logLik The log-likelihood.
status A vector which contains the status of LAPACK's dpotri and dpotrf. (0, 0) means successful exit.
sys.time The time elapsed as an object of class “proc_time”.

The first element of both at and Pt is filled with the function arguments a0 and P0, and the last, i.e. the (n + 1)-th, element of at and Pt contains the predictions
at[,n + 1] = E(alpha[n + 1] | y[,n]) and
Pt[,,n + 1] = var(alpha[n + 1] | y[,n]).

Author(s)

David Luethi

Philipp Erb
Zurich Insurance Company (http://www.zurich.com)

Simon Otziger
Clariden Leu (http://www.claridenleu.com)

Maintainer: Philipp Erb <erb.philipp@gmail.com>

References

Harvey, Andrew C. (1990). Forecasting, Structural Time Series Models and the Kalman Filter. Cambridge University Press.

Hamilton, James D. (1994). Time Series Analysis. Princeton University Press.

Koopman, S. J., Shephard, N., Doornik, J. A. (1999). Statistical algorithms for models in state space using SsfPack 2.2. Econometrics Journal, Royal Economic Society, vol. 2(1), pages 107-160.

See Also

plot to visualize and analyze fkf-objects, KalmanRun from the stats package, function dlmFilter from package dlm.

Examples

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
## <--------------------------------------------------------------------------->
## Example 1: ARMA(2, 1) model estimation.
## <--------------------------------------------------------------------------->
## This example shows how to fit an ARMA(2, 1) model using this Kalman
## filter implementation (see also stats' makeARIMA and KalmanRun).
n <- 1000

## Set the AR parameters
ar1 <- 0.6
ar2 <- 0.2
ma1 <- -0.2
sigma <- sqrt(0.2)

## Sample from an ARMA(2, 1) process
a <- arima.sim(model = list(ar = c(ar1, ar2), ma = ma1), n = n,
               innov = rnorm(n) * sigma)

## Create a state space representation out of the four ARMA parameters
arma21ss <- function(ar1, ar2, ma1, sigma) {
    Tt <- matrix(c(ar1, ar2, 1, 0), ncol = 2)
    Zt <- matrix(c(1, 0), ncol = 2)
    ct <- matrix(0)
    dt <- matrix(0, nrow = 2)
    GGt <- matrix(0)
    H <- matrix(c(1, ma1), nrow = 2) * sigma
    HHt <- H %*% t(H)
    a0 <- c(0, 0)
    P0 <- matrix(1e6, nrow = 2, ncol = 2)
    return(list(a0 = a0, P0 = P0, ct = ct, dt = dt, Zt = Zt, Tt = Tt, GGt = GGt,
                HHt = HHt))
}

## The objective function passed to 'optim'
objective <- function(theta, yt) {
    sp <- arma21ss(theta["ar1"], theta["ar2"], theta["ma1"], theta["sigma"])
    ans <- fkf(a0 = sp$a0, P0 = sp$P0, dt = sp$dt, ct = sp$ct, Tt = sp$Tt,
               Zt = sp$Zt, HHt = sp$HHt, GGt = sp$GGt, yt = yt)
    return(-ans$logLik)
}

theta <- c(ar = c(0, 0), ma1 = 0, sigma = 1)
fit <- optim(theta, objective, yt = rbind(a), hessian = TRUE)
fit

## Confidence intervals
rbind(fit$par - qnorm(0.975) * sqrt(diag(solve(fit$hessian))),
      fit$par + qnorm(0.975) * sqrt(diag(solve(fit$hessian))))

## Filter the series with estimated parameter values
sp <- arma21ss(fit$par["ar1"], fit$par["ar2"], fit$par["ma1"], fit$par["sigma"])
ans <- fkf(a0 = sp$a0, P0 = sp$P0, dt = sp$dt, ct = sp$ct, Tt = sp$Tt,
           Zt = sp$Zt, HHt = sp$HHt, GGt = sp$GGt, yt = rbind(a))

## Compare the prediction with the realization
plot(ans, at.idx = 1, att.idx = NA, CI = NA)
lines(a, lty = "dotted")

## Compare the filtered series with the realization
plot(ans, at.idx = NA, att.idx = 1, CI = NA)
lines(a, lty = "dotted")

## Check whether the residuals are Gaussian
plot(ans, type = "resid.qq")

## Check for linear serial dependence through 'acf'
plot(ans, type = "acf")


## <--------------------------------------------------------------------------->
## Example 2: Local level model for the Nile's annual flow.
## <--------------------------------------------------------------------------->
## Transition equation:
## alpha[t+1] = alpha[t] + eta[t], eta[t] ~ N(0, HHt)          
## Measurement equation:
## y[t] = alpha[t] + eps[t], eps[t] ~  N(0, GGt)

y <- Nile
y[c(3, 10)] <- NA  # NA values can be handled

## Set constant parameters:
dt <- ct <- matrix(0) 
Zt <- Tt <- matrix(1)
a0 <- y[1]            # Estimation of the first year flow 
P0 <- matrix(100)     # Variance of 'a0'

## Estimate parameters:
fit.fkf <- optim(c(HHt = var(y, na.rm = TRUE) * .5,
                   GGt = var(y, na.rm = TRUE) * .5),
                 fn = function(par, ...)
                 -fkf(HHt = matrix(par[1]), GGt = matrix(par[2]), ...)$logLik,
                 yt = rbind(y), a0 = a0, P0 = P0, dt = dt, ct = ct,
                 Zt = Zt, Tt = Tt, check.input = FALSE)

## Filter Nile data with estimated parameters:
fkf.obj <- fkf(a0, P0, dt, ct, Tt, Zt, HHt = matrix(fit.fkf$par[1]),
               GGt = matrix(fit.fkf$par[2]), yt = rbind(y))

## Compare with the stats' structural time series implementation:
fit.stats <- StructTS(y, type = "level")

fit.fkf$par
fit.stats$coef

## Plot the flow data together with fitted local levels:
plot(y, main = "Nile flow")
lines(fitted(fit.stats), col = "green")
lines(ts(fkf.obj$att[1, ], start = start(y), frequency = frequency(y)), col = "blue")
legend("top", c("Nile flow data", "Local level (StructTS)", "Local level (fkf)"),
       col = c("black", "green", "blue"), lty = 1)

Want to suggest features or report bugs for rdrr.io? Use the GitHub issue tracker.