The 3 types of models

Descriptive models

A narrative that explains how Trump won the 2016 US election can be useful to ask counter-factual questions of the past, to write a post about, etc. It can be used to explain a phenomena.

But, it probably doesn’t have any predictive power. Our world has changed (

) since then, social media platforms have different rules on data collection and advertising.

Even if the model has described the reality well for that snapshot of time (which is extremely hard to verify, as there could be many confounding variables), it may simply not be useful at all for anyone other than historians.

How Bitcoin has behaved before it became a “risk-on” asset (with a 0.8 correlation with nascent tech stocks) is not predictive on how it’ll behave in this new regime.

Usually, a descriptive model doesn't even pretend it could have existed at the time when it could have had predictive power: the hidden factors may have made the construction of there model impossible. A narrative isa descriptive model.

Predictive models

A predictive model, in contrast is one that has predictive power in our current situation. We may have “verified” its predictive power by looking at similar scenarios where it was useful. Or the “universal” model may have been built to handle a great variety of scenarios, including the one where we want to use it.

For every predictive model, one important component is timeframe. What window of time are we expecting this model to have a predictive power on? If the world’s GDP can be fitted to an exponential curve, do we expect it to increase up until the end of time?

Models with a feedback loop

Predictive models may influence the future, if we act upon their recommendation.

Self-canceling: If a model correctly predicts a pandemic’s growth, which then used to restrict people’s movement and therefore ability to catch the virus - then it’s impossible to verify its accuracy anymore, as we can’t untangle the new policy’s effect on the spread of the virus.

Self-reinforcing: Let’s assume there’s a model that predicts who’s the next world leader gonna be at a young age. We then surround that person with the best teachers, and when he/she grows up, the “best of the best” will desperately wants to work with him/her. That’s a pretty good setup to be the world leader, right? So the prediction of the model has influenced the world, and made the outcome of its prediction more likely. Soros calls this reflexivity.