# Kalman Filtering, Smoothing and Parameter Estimation for State Space Models in R and C#

I am in the process of writing an open source State Space Analysis suite in C# (for fun). I have implemented a number of different Kalman-Based Filters (Kalman Filter, Information Filter and the Square Root Filter) and the State and Disturbance Smoothers that work with these implementations. My tests on these filters (using the Nile data from Durban and Koopman's (DK) book "Time Series Analysis by State Space Methods" and other more complex data) show that the filters and smoothers work and they produce very similar results (as you would expect for a local univariate model). The smoothed output for the basic Kalman Filter for the Nile data looks like (because everyone likes a graphic or two :])

The dotted line is the 90% confidence interval.

Now, my problem; I am now attempting to write the first part of the parameter estimation code and for the parameter estimation of the covariances $\mathbf{H}_{1}$ and $\mathbf{Q}_{1}$ in the linear Gaussian State Space model

$$y_t = \mathbf{Z}_{t}\alpha_{t} + \epsilon_{t},$$ $$\alpha_{t + 1} = \mathbf{T}_{t}\alpha_{t} + \mathbf{R}_{t}\eta_{t},$$

where $y_{t}$ is our observation vector, $\alpha_{t}$ our state vector at time step $t$ and

$$\epsilon_{t} \sim NID(0, \mathbf{H}_{t}),$$ $$\eta_{t} \sim NID(0, \mathbf{Q}_{t}),$$ $$\alpha_{1} \sim NID(a_{1}, \mathbf{P}_{1}).$$

where $t = 1,\ldots, n$.

We have an implementation of the Expectation Maximisation (EM) algorithm and this is estimating $\mathbf{H} = 19550.37$ and $\mathbf{Q} = 7411.44$ with the loglikelihood converging in 32 iterations to -683.1 with a convergence tolerence of $10^-6$.

The DK book quotes $\mathbf{H} = 15099$ and $\mathbf{Q} = 1469.1$ and from Tusell's paper and this walkthrough using R he matches this prediction getting $\mathbf{H} = 15102$ and $\mathbf{Q} = 1468$. I have debugged my code and it really seems like my implementation of the EM algorithm is correct. So get to the bottom of what is going on I want to run the same data set using R and KFAS...

Using R and the R Package KFAS, I have attempted to reproduce these estimates for $\mathbf{H}$ and $\mathbf{Q}$, but my R knowledge is weak. My current R script is as follows

install.packages('KFAS')
require(KFAS)

# Example of local level model for Nile series
modelNile<-SSModel(Nile~SSMtrend(1,Q=list(matrix(NA))),H=matrix(NA))
modelNile
modelNile<-fitSSM(inits=c(log(var(Nile)),log(var(Nile))),model=modelNile,
method='BFGS',control=list(REPORT=1,trace=1))$model # Can use different optimisation: # should be one of “Nelder-Mead”, “BFGS”, “CG”, “L-BFGS-B”, “SANN”, “Brent” modelNile<-SSModel(Nile~SSMtrend(1,Q=list(matrix(NA))),H=matrix(NA)) modelNile modelNile<-fitSSM(inits=c(log(var(Nile)),log(var(Nile))),model=modelNile, method='L-BFGS-B',control=list(REPORT=1,trace=1))$model

# Filtering and state smoothing
out<-KFS(modelNile,filtering='state',smoothing='state')
out


How can I extend this R script using KFAS in order to estimate the parameters $\mathbf{H}$ and $\mathbf{Q}$? and Should each method of parameter estimation be hitting the same values for $\mathbf{H}$ and $\mathbf{Q}$?

Im not sure why KFS does not return V_eta or V_eps, but why not use modelNile$H and modelNile$Q? They are also kept in the return object from KFS, out$model$H and out$model$Q. You will find what the objects contains by using the names() function as in names(modelNile) or names(out$model). Afer perusing the documentation for KFAS, it seems to me that KFS() will return what you want in components V_eta and V_eps of the object you name out. (This is the case because you are dealing with a univariate time series, so the only diagonal term of V_eps is the variance you want.) You should expect about the same values from your code and any of the packages. Discrepancies may crop up, however, on account of: i) Different initial conditions (exactly diffuse, as KFAS implements, or setting the variance of the initial state to a very large number, as you probably have done), and ii) Different criteria to stop the likelihood climbing. Discrepancies due to different algorithms (like plain Kalman filter iteration versus square root algorithms, either covariance or information filters) may also appear with longer series and/or more complicated models, but with not the Nile data. Please note my name's spelling: Tusell. • Thanks for your reply and sorry about the spelling of your name; I was typing very quickly. I have attempted to check the variables out$V_eta/out$V_eps but they both return NULL. I have also attempted this after running modelNile<-fitSSM(...) by using modelNile$V_eta/modelNile\$V_eps but this also returns NULL. Again, thanks for your time... – MoonKnight May 19 '14 at 7:43