Today models are everywhere from big Google-sized problems to tiny optimizations like customer journey conversion rate. But models are not mirrors of reality, “models are opinions embedded in mathematics”. From the data they use to the questions we ask they are reflections of the modelers own values, desired and world view.
In her book Weapons of math destruction Cathy O’Neil talks about how we use mathematical models and proxies to simulate the world and make decisions.
A mathematical model is nothing more than a simulation of the world, it is a simplifications (they have to be), and often they are modeling a context based on data they don’t have .. so they use proxies.
Proxies are statistical correlations that make assumptions about what has a causal effect on something else (if A happens B will happen), but often we use proxies that are spurious at best — we don’t know that A has an effect on B. “A model’s blind spots reflects the judgements and priorities of it’s creators”
“Whether a model works or not is also a matter of opinion”, “models, despite their reputation for impartiality, reflect goals and ideology”.
But we need models, so how do we make them work for us? O’Neil suggests a few factors relevant for a model of any size or purpose:
- Transparency. Since every model is a reflection of someones values and ideas and many models uses proxies, the best way to make them work is to make them transparent so that everyone can participate in the evaluation and discussion of the assumptions the models make
- Dynamic. “That is how trustworthy models operate. They maintain a constant back-and-forth with whatever in the world they are trying to understand or predict. Conditions change and so must the model.”
- Clarity. [This is my own experience] A model must be very clear on what it is modeling and what it is not modeling. Don’t over-market it.