A lot of people understand the scikit-learn models of todays world but feel uneasy about the whole MCMC method of training. Why are these algorithms different? How is it that you don't use a gradient method but a sampler instead? It can feel a bit misterious if you've not properly been introduced to this other way of thinking.

In this talk I will explain the idea of sampling to get to your model and I will demonstrate it with examples. The goal is to start with a for loop and to end with understanding how MCMC algorithms work. As a consequence the audience will also get a proper introduction to PyMC3. In particular I will discuss the following;

• why markov chain sampling can be equivalent to direct sampling
• how to build your own MCMC sampler with a for loop
• how this for loop can be run faster by using PyMC3 instead
• the key idea of inference and how i was briefly able apply it in a santa kaggle competition
• how to analyse timeseries with MCMC in PyMC3

Parts of this talk are readily available on my blog;

• http://koaning.io/switching-to-sampling-in-order-to-switch.html
• http://koaning.io/elimination-via-inference.html

Let me know if there are any questions. I am submitting multiple talks that I think are interesting and relevant to the PyData crowd, I'll gladly leave it to the committee which (or if any) of them are relevant to the local community.