Entropy maximizing models (and closely related information theory) permit the construction of probability distributions with known mean and variance. You could constrain you model to match some high level of variance. I assume this is in the context of uncertainty? See text by TRIBUS on Rational Description .....

GARCH type model deals with the changing variance of data. But it depends on your purpose for prediction. ANN, SVM are also able to deal with complex system. Or a hybrid method that combining multiple models. It may helps If you provide more details about your problem.

I deal with safety issues and crash data have a nature of variance being greater than mean a condition that we call overdispersion. For that case, negative binomial modelling is found to be good!!! 

If data are count, discrete, random and cross-sectional then Negative Binomial (NB), Poisson-LogNormal, Conway Maxwell Poisson, Zero-Inflated NB would be appropriate. For time series continuous data, I agree with the suggestion by Yanru Zhang (see above) but for time series count data, the appropriate model would be: Integer-valued AR(p) Poisson or NB.

Poisson and NB models deals with count data like Number of Crashes. Having continuous data like travel time its better to use Linear regression Models like Multiple Linear regression, Multivariate regression, ANOVA or MANOVA and others which can deal with continuons outcomes but with any form vof independent variables. I have attached a vlink for your guidance. Thanks

I didnt attach the link . Its down here

https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=1&cad=rja&uact=8&ved=0CB0QFjAA&url=http%3A%2F%2Fwww-rohan.sdsu.edu%2F~renglish%2F470%2Fnotes%2Fchapt15%2Fchapter15.ppt&ei=8mzFU4rQOKLjsASVt4DYBQ&usg=AFQjCNHpliZxdiWt2hxky9n7JeCU2BGB_Q&sig2=XYf-NQGOJOwTnKDC-_Vtkg