Hello, I have divergence issue and I think I need some reparameterization. I would like to perform Bayesian inference with stock price. Attached are posterior outcome from weekly, monthly and yearly data.
Weekly (size=7)
Is my posterior dist. valid with the prior values given by the example?
Parameters from example:
σ∼exp(50)
ν∼exp(.1)
si∼N(si−1, σ^−2)
log(yi)∼ t(ν,0,exp(−2si))
Thanks and Regards.
I tried to adjust the parameter values of lambda to 1./1 for nu or v and sigma to get uniform prior after knowing that the exponential distribution is a special form of gamma distribution with alpha = 1 from PyMC2 documentation. Since the beta corresponds with the lambda, so I adjusted the value of beta or lambda to 1. I got no divergence after the adjustment except for the weekly data possibly because of low amount of samples. Anyway, I dealt with that by increasing target_accept. I hope my solution should be reasonable enough.
I gather that the number of data point is different in these cases? and there are more data point in Yearly? (is that what the size= indicates?)
My suggestion would be try changing the prior of sigma and nu to HalfNormal or HalfCauchy.
Thanks for the reply. Yup, the size indicates the number of data point. Anyway, I have stuck with Yearly data because it is more accurate. By the way, may I know is the stochastic volatility model shown in example a GARCH model?
Just realized that it isn’t the GARCH model, hopefully there will be documentation for GARCH model.
We have a small example on GARCH model https://github.com/pymc-devs/pymc3/blob/master/pymc3/examples/garch_example.py
