Hello,
As suggested at the bottom of notebook, I tried to use multiple models like EGARCH and GJR-GARCH.
So starting to play with it, I started by change arch_model vol parameter to EGARCH which return the following results : AIC model is best for p=5, and q=1 and got MSE = 1.4. I tweaked the model with p=1 and q=1 and got MSE=0.9, dist = skewt for both.
My question is : which one is the best model? What other distribution can we use here?
For GJR-GARCH, after reading Shepard doc (https://arch.readthedocs.io/en/latest/univariate/generated/arch.univariate.GARCH.html)
which enumerate the parameter "o" as aysmetric innovation,
how to implement this model in our case? Is adding "o=1" parameter in the arch_model is sufficient or should we instanciate the GARCH class ? Should the parameter o be estimated as p and q?
For a general question : in live trading and risk management what is the reasonnable precision of such model for 20 days forecast or more.
Thank you.
Eric
Hi Eric,
We are looking into your query. Will get back to you soon. Thanks!
Hello Eric,
Firstly, I'd like to tell you that I'm glad to see you experimenting with different volatility models.
It's an effective approach to getting comfortable with the underlying ideas.
Some of the questions you ask are far from settled in the econometrics literature.
Let me try and address it as best as I can.
Q1: The short answer is it depends. The model we use is closely related to the data we have and the purpose of the exercise. There is no one-size-fits-all, I'm afraid. As researchers, we should be aware of the limitations and strengths of the one we choose. For your reference, I suggest you look at the paper titled "Comparing the performances of GARCH-type models in capturing the stock market volatility in Malaysia", Lim and Sek, 2013. It compares different GARCH models (symmetric and asymmetric). You can skim through it to get a sense of how you can choose from competing models.
Q1: The short answer is it depends. The model we use is closely related to the data we have and the purpose of the exercise. There is no one-size-fits-all, I'm afraid. As researchers, we should be aware of the limitations and strengths of the one we choose. For your reference, I suggest you look at the paper titled "Comparing the performances of GARCH-type models in capturing the stock market volatility in Malaysia", Lim and Sek, 2013. It compares different GARCH models (symmetric and asymmetric). You can skim through it to get a sense of how you can choose from competing models.
Q2: The documentation here has many working examples of implementing variants of the GARCH models.
Please take a look at it.
Q3: Excellent question. I'd say the best forecasts are one time-period ahead. So if you're looking for predictions, one month hence, you can work with monthly data and forecast the next period. The study, "Longer-Term Time-Series Volatility Forecasts", JFQA, Ederington and Guan, 2010 talks about the problems faced when you use the GARCH models for longer-term volatility forecasting.
Q3: Excellent question. I'd say the best forecasts are one time-period ahead. So if you're looking for predictions, one month hence, you can work with monthly data and forecast the next period. The study, "Longer-Term Time-Series Volatility Forecasts", JFQA, Ederington and Guan, 2010 talks about the problems faced when you use the GARCH models for longer-term volatility forecasting.
Hope this helps.
Hello Vivek,
Really interesting post here. Thank you for sharing. I'm going to dive deep in these ressources.
It is a long journey to the quant-trading and quantra make it easy.
Keep it rolling.
Eric