The RATS Software Forum

Dear Tom,
I am having issues with convergence. I have attached the model as well a sample of the data. The data consists of min-by-min stock returns on AXP (American Express) and 10 mood indicators that have been transformed to eliminate the zero observations. The mood indicators were extracted with an algorithm through tweet data and are matched with the stock return data. I would like to investigate which indicator has the largest effect on the stock return volatility but I have not been successful in obtaining convergence.

Thank you,

Pieter

Why are you including those as dependent variables? Wouldn't a univariate GARCH with the "mood" variables as variance shifts be more appropriate?

Tom,

I modeled the GARCH with the "mood" variables as variance shifts per your suggestion and that worked splendidly. I am hoping you could point me towards a procedure that would allow for out-of-sample forecasting and test the "strength" of those forecasts with another sample.

Best,

Pieter de Jong

paldejong wrote:Tom,

I modeled the GARCH with the "mood" variables as variance shifts per your suggestion and that worked splendidly. I am hoping you could point me towards a procedure that would allow for out-of-sample forecasting and test the "strength" of those forecasts with another sample.

Best,

Pieter de Jong

This would produce forecasts of the variance (which, BTW, isn't much different from forecasting a simple univariate GARCH model). How are you planning to evaluate how that works?

I have an additional question about the shift variables. Do I have to code them as dummies (i.e., "angry" tweets per minute is a 1; zero otherwise) or could I leave them in the transformed form, which will be an aggregate of all the angry tweets in that one minute? So far, I have only seen literature describing shifts as dummy variables.

Thank you,

Pieter

No. They can be continuous. Obviously, you would need to decide whether (for instance) 20 should have twice the effect of 10 or whether you need to do some transformation of the continuous shift.

Thank you, I opted for a transformation but I am confused about the interpretation of the parameter signs. Would a statistically significant positive sign on a "happiness" coefficient imply an expansion of the conditional volatility, i.e., tweets about a particular stock with predominantly "happy" words would add to the volatility??

Best,

Pieter

It would. Have you allowed for the possibility that those have an effect on the mean as well?