Optimizing code for HAC Weight matrix generation - How to be faster?
9 views (last 30 days)
Show older comments
Hi everyone,
I am re-creating Newey-West procedure for a heteroskedasticity and autocorrelation consistent standard errors (HAC) from scratch.
In order to compute the sandwich matrix V(b), I need to generate the weight matrix W
Since I have to run that test thousands of times, I need to optimize the code. Right now, I have this:
%% Weight matrix generation for HAC
T = length(Dependant_Variable);
k = length(Coeff);
df = T-k;
Optimal_Lags = floor(4*(T/100)^(2/9));
Residual_correlation = Residual*Residual';
Weight_correlation = zeros(size(Residual_correlation));
for i = 1:size(Weight_correlation,1)
for j = 1:size(Weight_correlation,2)
if abs(i-j)>Optimal_Lags
Weight_correlation(i,j)=0;
else
Weight_correlation(i,j)=(T/df)*Residual_correlation(i,j)*(1-abs(j-i)/(Optimal_Lags+1));
end
end
end
Question:
How can I get the same result without using loops?
Thanks and best regards,
0 Comments
Accepted Answer
Alan Stevens
on 18 Mar 2024
Try replacing the loops with something like this:
i = (1:size(Weight_correlation,1))';
j = 1:size(Weight_correlation,2);
c = ones(1,size(Weight_correlation,2));
r = ones(size(Weight_correlation,1),1);
lags = abs(i*c-r*j);
ind1 = lags>Optimal_Lags;
ind2 = lags<=Optimal_Lags;
Weight_correlation(ind1)=0;
Weight_correlation(ind2)=(T/df)*Residual_correlation(ind2).*(1-lags(ind2)/(Optimal_Lags+1));
Check carefully, as I can't really test it properly, not having your data.
2 Comments
Alan Stevens
on 20 Mar 2024
Edited: Alan Stevens
on 20 Mar 2024
Try this:
Variables = rand(10,10);
Companies = rand(10,1);
Coeff = (Variables'*Variables)\Variables'*Companies;
Expected = Variables*Coeff;
Residual = Companies-Expected;
T = length(Companies);
k = length(Coeff);
df = T-k;
Optimal_Lags = floor(4*(T/100)^(2/9));
Residual_correlation = Residual*Residual';
Weight_correlation = zeros(size(Residual_correlation));
for i = 1:size(Weight_correlation,1)
for j = 1:size(Weight_correlation,2)
if abs(i-j)>Optimal_Lags
Weight_correlation(i,j)=0;
else
Weight_correlation(i,j)=Residual_correlation(i,j)*(1-abs(j-i)/(Optimal_Lags+1));
end
end
end
i = (1:size(Weight_correlation,1))';
j = 1:size(Weight_correlation,2);
c = ones(1,size(Weight_correlation,2));
r = ones(size(Weight_correlation,1),1);
lags = abs(i*c-r*j);
ind2 = lags<=Optimal_Lags;
Weight_corr = ind2.*Residual_correlation.*(1-lags/(Optimal_Lags+1));
Check = Weight_correlation - Weight_corr;
disp(Check)
(I note that you removed the T/df term which gave the inf values!).
More Answers (0)
See Also
Categories
Find more on Nonspherical Models in Help Center and File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!