Facebook has an excellent open source time series analysis tool called Prophet (for example to predict event attendance). Here they use Bayesian modeling to infer various seasonal patterns combined with unpredictable changepoints and wrap them in a Generalized Additional Model. This talk focusses on bolts and nuts of this algorithm and shows how you can implement their model with PymC3.
A lot of time series models only focus on predicting relatively short time intervals. By inferring Bayesian Generalized Additive Models are able to predict over longer horizons in the future. And because it is Bayesian we also get a hold on uncertainty intervals of our predictions.
During this talk we'll discuss:
Prior knowledge of bayesian inference is assumed.
The talk is based on this blog post