Chan Karolyi Longstaff Sanders JOF 1992

[TomDoan]

The following is a replication of the estimation of various interest rate models by GMM from Chan, Karolyi, Longstaff and Sanders(1992), "An Empirical Comparison of Models of the Short-Term Interest Rate", Journal of Finance, vol 47, no 3, 1209-1227. Note that the model, while estimated with monthly data, uses a parameterization which annualizes the rates. This includes two versions for the estimation; one which uses the weighting scheme from the original paper, the other uses a common weighting scheme computed under the just-identified specification.

ckls.rpf

Original estimation method

(5.4 KiB) Downloaded 1046 times

ckls commonsw.rpf

Alternative estimation method

(5.61 KiB) Downloaded 1123 times

rates.txt

Data file

(5.4 KiB) Downloaded 1342 times

[Paradox4412w]

Mr Doan kindly recreated the Chan, Karolyi, Longstaff and Sanders (1992) nested model test for short-term interest rates.

I want to include a dummy variable test for monetary regime changes.

Any suggestions??????

[TomDoan]

Doesn't ckls.rpf include a dummy variable shift?

[Paradox4412w]

It does! I am going blind running CDS models!

[Paradox4412w]

Tom:

I tried to replicate CKLS structural break test using your code, but it is different (can't verify CKLS finds on page 1223, Table 5. Using your data and code.

Any suggestions?

[TomDoan]

Thanks for pointing that out. The variance formula wasn't right. It should be

frml varianced = epsd(t)^2-(sigmasq+sigmasqd*dummy)/12.0*r^(2*(gamma+gammad*dummy))

(it was using EPS(T)^2 rather than EPSD(T)^2). I fixed the two attachments in the original post.

[Paradox4412w]

Getting closer.

For the unrestricted model, the second dummy BETAD in ckls is -0.0751 while RATS generates -0.751436998. Also, the papers ChiSQ is 2.2939 while RATS generates 2.8494.

[Paradox4412w]

Tom, do you think your code revisions are correct for the dummy variable test? If so, this implies that the dummy variable test (from Gauss) in the original CKLS JF paper is incorrect. Correct?

[TomDoan]

They both could be "correct". Non-linear IV with iterated weight matrices can end up with different weights, which will produce different test statistics.

[Paradox4412w]

Excellent point! Thanks. Just ran several with different weightings and confirmed that.