turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Forecasting
- /
- garch, multivariate garch, nonconvergence?

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

03-19-2013 06:46 AM

I have 5 sorted portfolio (time series) returns in first difference (for stationarity) - R1-R5. I have checked all 5 returns, first difference is stationary, only has some AR effect.

My first regression is garch(1,1) for all 5 separately:

proc **autoreg** data=combine_ret1;

model R1 = IST_R LR1 win loss fb bjk gs/ garch=(p=**1**,q=**1**);

**run**;

no problem at all.

My second regression is including explanatory variables in the variance equation for all 5 separately:

proc **autoreg** data=combine_ret1;

model R1 = IST_R LR1/ garch=(p=**1**, q=**1**);

hetero win loss;

**run**;

only 20% converge and have nonzero coefficients. For the rest, either not converge, or the variance equation coefficients (win and loss) are zero or missing. Why does it happen? nonconvergence? like below:

Intercept | 1 | -0.0485 | 0.0442 | -1.10 | 0.2725 |
---|---|---|---|---|---|

IST_R | 1 | 0.6497 | 0.0229 | 28.40 | <.0001 |

LR5 | 1 | -0.4936 | 0.0133 | -37.23 | <.0001 |

ARCH0 | 1 | 0.0160 | 0.008433 | 1.89 | 0.0584 |

ARCH1 | 1 | 0.0692 | 0.004690 | 14.76 | <.0001 |

GARCH1 | 1 | 0.9315 | 0.004024 | 231.49 | <.0001 |

HET1 | 1 | 0 | 0 | . | . |

HET2 | 1 | 0.8206 | 0.1607 | 5.11 | <.0001 |

My third regression is multivariate garch:

proc **varmax** data=combine_ret1;

model r1 r4= IST_R win loss fb bjk gs/ p=**1**;

**q=1** form=ccc;

run;

IST_R is market return, the rest are dummies. I'm using R1 and R4, not R1 and R5 (lowest and highest) because it doesn't converge. I actually want to do R1-R5. Of course, it doesn't converge. What can I do?

thank

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to econfkw

03-19-2013 02:50 PM

Hi econfkw,

Thanks for the question. The developer and I have chatted and think we have a couple ideas.

First, might it be possible that in your HET equation, you have perfectly collinear variables? I would think that in each state, there is either a win or a loss. When you specify the HETERO option with a GARCH model the link function is linear. This might explain your issue. You could try to drop one of the vars in the HET equation.

The developer provides an alternative explanation.

"There is another general reason why introducing extra variables

in GARCH is not easy: there is no guarantee that h_t>0. For example,

h_t = omega + alpha* e_{t-1}^2 + gamma * h_{t-1} + het_1 * WIN +

het_2 * LOSS

When het_1 or het_2 is negative, h_t might be negative. More

severe, when some e_t is zero (no matter what’s the reason), the log-likelihood

on observation t, -(log(h_t) + e_t^2/h_t) = -log(h_t), could go to positive

infinity when h_t keeps positive but goes to 0!!! So far, to my knowledge,

there is no solution for this problem, no matter in theory or in practice. That

might be why GARCH converges, but GARCH with extra variables cannot converge."

As to your second question, perhaps you have perfect collinearity in those dichotomous regressors again. Try the following,

proc **varmax**

data=combine_ret1;

model r1 - r5 = /

noint p=**1**;

**garch q=1** form=ccc;

run;

And if you rule out a data problem, try playing with the convergence settings of the PROC. Also, which version of SAS are you working with?

Hope this helps-Ken

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

Posted in reply to ets_kps

03-19-2013 09:56 PM

The win and loss are not perfectly collinear. There is also a draw dummy(30% of totoal obs), which is omitted. The code provided at the bottom of last repl (multiariation GARCH without explanatory variables) do not converge. Thanks for your advice, I'll try to change the convergence criteria. BTW, the 5 portfolio returns only exhibit something like AR(1) process, it should be fine ha?

The dataset is not large. I'm attaching it and the codes that I tried. The choice of independent variables is changed a little to avoid singularity.

The first 2 codes work, not the last three.