I’ve recently been working through an excellent resource on causal inference
from Matheus Facure. Machine learning models are great at predicting things (well, in some cases…*cough* Zillow *cough*). You throw some data at them, make some predictions and try to find a model which gives you the best accuracy (or whatever measure it is you’re trying to optimize). Some models may be explainable as to why they make their predictions and some are black box models which can be incredibly difficult to interpret which in turn runs the risk of a lack of trust in the predictions. However, consider a machine learning model trained on airline prices…it would see that ticket sales a relatively low when the prices are low (during term time for example) and ticket sales go up as prices go up (during school holidays). A naive model may suggest increasing prices to increase sales…