Causal inference aims to identify relationships so that certain input values imply certain outputs—for example, a certain constellation of macro variables causing the price of a given asset to move in a certain way, assuming all other variables remain constant.

Statistical inference about relationships among two or more variables produces measures of correlation that can only be interpreted as a causal relationship when several other conditions are met—for example, when alternative explanations or reverse causality has been ruled out. Meeting these conditions requires an experimental setting where all relevant variables of interest can be fully controlled to isolate causal relationships. Alternatively, quasi-experimental settings expose units of observations to changes in inputs in a randomized way to rule out that other observable or unobservable features are responsible for the observed effects of the change in the environment.

These conditions are rarely met so inferential conclusions need to be treated with care. The same applies to the performance of predictive models that also rely on the statistical association between features and outputs, which may change with other factors that are not part of the model.

The non-parametric nature of the KNN model does not lend itself well to inference, so we'll postpone this step in the workflow until we encounter linear models in the next chapter.