Re: Question regarding EndCriteria parameter (Old Subject: Garch11 exmaple code needed)

Posted by Andres Hernandez-2 on Jul 23, 2017; 8:02pm
URL: http://quantlib.414.s1.nabble.com/Re-Question-regarding-EndCriteria-parameter-Old-Subject-Garch11-exmaple-code-needed-tp18435p18437.html

maxIterations_

Maximum number of iterations. 

Are these iterations through the TimeSeries data or future iterations of how many results will be returned.

maxIterations is used to tell the optimization method, regardless of the method, the maximum number of times that it should iterate. If the method does not determine success (however that is defined by the method), then it should stop after this many iterations. In general, it should mean the maximum number of times that the objective function should be evaluated, although this is not necessarily true for some methods, e.g. swarm methods for which each iteration evaluates N times the objective. 

maxStationaryStateIterations_

Maximun number of iterations in stationary state. 

What is the meaning of stationary state.

Stationary state refers here to the best answer until now. Meaning that if the optimization method does not find a better answer in the next maxStationaryStateIterations, it should stop looking. 

rootEpsilon_

root, function and gradient epsilons 

I assume this is the value from which the process starts searching.

No, all three epsilons relate to the number accuracy that the method should concern itself with. rootEpsilon refers to the number accuracy in the parameter space, e.g. when comparing whether two points are the same, they would be so considered up to the given accuracy. For most applications, the full machine accuracy (e.g. ca. 10^-15 for doubles) is not needed, so giving a wider tolerance will allow the method to stop earlier. 

functionEpsilon_

What's this?

This epsilon refers to the accuracy at the objective/target space. If the method compares whether a new solution is as good as the best current solution, then it only cares up to the accuracy given here. 

gradientNormEpsilon_

What's this?

Ditto, but for the gradient. 

It is the obligation of the optimization method to pay attention to these parameters, but some might not be relevant. If the method does not use a gradient, then gradientNormEpsilon_ will of course not be important. There might be other parameters that might be ignored by the method, e.g. like maxStationaryStateIterations_, which for a local optimization method is also not important. 

cheers,

Andres

On Sun, Jul 23, 2017 at 8:04 PM, Andrew Leach <[hidden email]> wrote:

Hi,

Ok, I've progressed and I'm willing to share my code here for anyone else who is having difficulties.

This is the main method of my class:

int CGarch::GarchOnArray(const std::vector<double> &iPrices, std::vector<double> &oGarch)

{

if (iPrices.empty()) {

QL_FAIL("ERROR: input array (ts) is empty");

return -1;

}

if(iPrices.size < 3) {

QL_FAIL("ERROR: minimum (3) individual prices not present in ts array");

return -1;

}

oGarch.clear();

Date ds(7, July, 1962);

TimeSeries<Volatility> ts(ds, iPrices.begin(), iPrices.end());

Garch11 g11(ts, Garch11::MomentMatchingGuess);

LevenbergMarquardt lm;

g11.calibrate(ts, lm, EndCriteria(3, 2, 0.0, 0.0, 0.0));

g11.calibrate(ts);

TimeSeries<Volatility> tsOut = g11.calculate(ts);

tsOut.find(ds+ts.size()+1);

oGarch = tsOut.values();

return 0;

}

However, I still don't understand the EndCriteria part of the code and haven't been able to find any answers in the available documentation.

I'd appreciate if someone could help me explain what each parameter does.

maxIterations_
Maximum number of iterations. 

Are these iterations through the TimeSeries data or future iterations of how many results will be returned.

maxStationaryStateIterations_
Maximun number of iterations in stationary state. 

What is the meaning of stationary state.

rootEpsilon_
root, function and gradient epsilons 

I assume this is the value from which the process starts searching.

functionEpsilon_

What's this?

gradientNormEpsilon_

What's this?

I appreciate this all probably related to Levenberg-Marquardt algorithm but I'm not a maths expert and haven't done stochastic calculus so struggling a little bit with this.

Really appreciate any help.

Thanks in advance.

On 11 July 2017 at 00:26, Andrew Leach <[hidden email]> wrote:

I've been reading the Book on Implementing QuantLib that I bought on LeanPub and the test case examples and I've come up with this outline of code that I need to perform Garch11 calculations.

QuantLib::Garch11 cgarch2(iArray, Garch11::MomentMatchingGuess);

LevenbergMarquardt lm;

cgarch2.calibrate(iArray, lm, EndCriteria(3, 2, 0.0, 0.0, 0.0));

cgarch2.calibrate(iArray);

oArray = garch.calculate(iArray);

where iArray and oArray are QuantLib::TimeSeries<Volatility> objects.

Is this outline correct? I appreciate someone giving some insight as I'm relatively new to this.

Also, I don't understand how to define the EndCriteria part of the calibrate method. I've read the docs and that doesn't help me either. Could someone explain in simple English please?

What I want to do is finish after 3 interations but as far as the other parameters I'm none the wiser.

Anyone can help me?

Thanks in advance.

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
QuantLib-users mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/quantlib-users

Virus-free. www.avg.com

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
QuantLib-users mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/quantlib-users