Handling missing observations while using fmincon

Hello,

I have data that consists of 70 variables that I observe for 100,000 observations but I don't observe a few of those 70 variables for some random observations and MATLAB codes them as NaN. I'm using fmincon to find the minimum of a function with 70 parameters. My function is of the form f(). When I run the algorithm, I notice that MATLAB does not compute the index function when there is a missing value for one of the k variables for some observation i. How should I handle such missing observations?

Thank you in advance,

Selcen

Answer 1

[Walter Roberson]

Use this kind of structure:

XY = [X, Y];
XY = rmmissing(XY);
Xm = XY(:,1:end-1);
Ym = XY(:,end);

objfun = @(PARAMS) sum((YourFunction(PARAMS,Xm)-Ym).^2);

bestPARAMS = fmincon(objfun, PARAMS0);

Answer 2

[Matt J]

They should be zeros rather than NaNs, shouldn't they? With zeros, they will make no contribution to the linear part of the prediction.

X(isnan(X))=0;