We use probabilistic modelling to transfer information from root causes, through a network of cause-effect steps, to the business relevant outcome.
Imagine a coffee roaster evaluates its supply chain and is interested in the climate change exposure given from individual coffee plants. Data and models suggest with reasonable probability that a specific coffee plant suffers from ever higher draught risk in future. Higher draught risk could manifest in more volatile or overall lower volumes of coffee beans to harvest. The lower coffee supply is coupled, together with a wealth of other variables, to higher coffee commodity prices for roasters. Due to competition, roasters may be able to transfer only parts of this additional cost to end consumers. Thus, climate change induced draughts impose an inherent business risk to roasters that can be quantified by propagating conditional probabilities over the entire coffee value chain.
We use probabilistic modelling in situations where models are either very complex, or cause-effect chains are not parameterized well (which is the vast majority of cases).
System dynamic modelling
If instead there exist robust, parameterized equations to describe dynamic effects, we turn to system dynamic modelling.
Imagine the same coffee roaster as above intends to hedge short-term draught risk by means of futures on coffee beans. Future prices follow a well parameterized behavior, similar to the standard Black-Scholes options model. Prices can mostly be described by differential equations taking a number of effects into account. Thus, the problem is well suited for system dynamical modelling.
Finally, we employ machine learning in several important ways.
First, above described modelling requires to learn many parameters from data and we use machine learning algorithms to make robust estimations of current, and predictions of future parameters. Further, we create independent statistical models if more appropriate than fitting probabilistic or system dynamic models. Especially models involving deep learning such as for image segmentation are best solved with convolutional neural networks and related architectures.