Is it possible to adapt a vector autoregressive model’s parameters continuously without the need to rerun vgxvarx?

Is it possible to first estimate a VAR model's parameters using vgxvarx and then somehow continuously adjust the model parameters as new data of the situation that is being modelled arrives?

Running the vgxvarx command to estimate the parameters of a vector autoregressive model sometimes takes a very long time to complete when dealing with many variables and lags, is there any Matlab function that can allow the model to tune itself to new data without the need to redo it from scratch? Something similar to the way “ adapt ” works in the neural networks toolbox perhaps?

Answer 1

[Hang Qian]

Hi Peta,

The VGXVARX function cannot adapt model parameters, but the idea you proposed can be implemented. VAR models are typically estimated by equation-wise least squares, which can be performed iteratively instead of using all observations at once.

If you do not want to write your own codes for recursive least squares, or using some third-party MATLAB codes, you may consider the state-space model supported by the Econometrics Toolbox. Put the VAR coefficients as the state variables (static transition, of course) and use each observation to update the state distribution whenever a new observation arrives. This will yield the recursive least squares results.

- Hang Qian

[Peta]

I’m trying to understand how to implement your suggestion of state-space models, but I’ve never worked with them before so it’s not crystal clear.

I take it you mean that I should use the ssm function from the econometrics toolbox and use the numbers stored in VARSpec.AR (autoregressive coefficients) as input?

But the ssm requires a minimum of three things to be defined; a state-transition matrix A, a state-disturbance-loading matrix B, and a measurement-sensitivity matrix C. If my VAR coefficients are A, what would B and C be in this case?