Here are the slides I prepared for a talk I was invited to give at zeroG (zerog.aero)—a subsidiary of the Lufthansa Group specialised in data science. (Some slides have been modified for copyright and/or confidentiality reason.) Bear in mind that slides alone are not at all self-sufficient, they are meant to support the talk. The targeted audience was rather broad, spanning from data scientists and business analysts, to senior executives.
After observing that many projects fail in spite of a promising spreadsheet-based forecast, we highlight one of the fundamental problems in planning under uncertainty today. Namely, a single statistic—typically the mean—often fails to properly describe an uncertain number. Furthermore, forecasting a single statistic is very hard; and even in the event where it is accurately forecasted, the underlying fundamentals, i.e., the real world, might decide on a vastly diverging outcome.
There exist ways to mitigate the luck factor. Unfortunately, those solutions are often ignored by the vast majority of corporate people (for various reasons that are not discussed). In many cases, those solutions consist in predicting a statistical distribution rather than a single point. We briefly present some of those solutions.
In particular, it is probably worth recalling that many machine learning techniques today—extensively relied on in various industries to support business decisions—are actually yielding a single point estimate. We briefly introduce the so-called Bayesian Neural Networks, which aim at predicting distributions.
Management Science, Risk Management, Luck, Bayesian Neural Networks, Deep Learning, Bootstrap Confidence Intervals