At Kolonial.no we built and implemented our own custom Bayesian forecasting model in PyMC3 - and learned a lot in the process. In this talk, we share our thoughts on when to do what we did, and more important, when not to do it. We also share how we stumbled into numerous pitfalls on the way but managed to climb out again.
At Kolonial.no, Norway’s largest online groceries retailer, sales forecasts are crucial for an efficient and scalable warehouse and distribution operation. To fill this need, we built and implemented our own custom Bayesian forecasting model with hundreds of random variables in PyMC3- and learned a lot in the process. In this talk, we share our thoughts on when to do what we did, and more important, when not to do it. We also share how we stumbled into numerous pitfalls on the way but managed to climb out again.
This talk should be interesting for data scientists who consider running their own large Bayesian models, managers of data scientists proposing something like this and engineers interested in how we managed to put this thing into production and how we work to improve it over time.
No specific background knowledge is required to appreciate the main take-aways of this talk, but prior knowledge of Bayesian modelling and PyMC3 will probably make you appreciate the talk even more.