Question Models, not Policy
We shut down society because of a model. How much do you know about the model governing your state? Though I’m glad that science is driving policy, all of us ought to learn more about these models.
Marc Andreessen famously predicted that software is eating the world. In a nod to the Father of Netscape, Steve Cohen and Matthew Granade argued that models would run the world in an op-ed on August 18, 2017.
Mr. Cohen founded Point72, a quantitative hedge fund that, indeed, is run by models with the singular goal of making money. By contrast, his previous fund was a fundamental fund, with investment decisions being made by human beings. Cohen realized, with enough data and computing power, machines are better at predicting the future than humans.
Seven years was what Mr. Cohen predicted before models would run organizations. I thought it would take at least a decade, even though I cofounded a company building models of global commodities supply chains. Most organizations still rely on humans to make decisions and, for both social and technical reasons, replacing human judgment is difficult.
Thanks to COVID-19, our predictions were wrong. Turns out, every nation is being run by pandemic models.
In some ways this is a fantastic development. Models are a kind of “digital twin” of the world around us, allowing for better predictions of the future. Companies like Google, Facebook, and Amazon are already really good at using data, which makes them most valuable companies in the world. Unfortunately, most companies—and certainly most governments—don’t use data to make better decisions.
Maybe the case of COVID-19 will cause us to become enlightened by the power of models to drive smart policy, instead of just taking our best guess.
The problem is, models are not oracles with perfect clairvoyance into the future. Models are a guess, just a more well-founded guess than gut alone. The output of a model can suggest policy, but an official still needs to make a judgment call.
Before we can decide whether our officials are making the right call or not, we first need to understand their models.
Models have a lot of quirks since they are only as good as the data fed into the them. A related problem is that it’s hard to decide whether, based on the output of a model, someone made the right decision. Models don’t output a “yes” or “no,” or a “buy” or “sell.” They just give probabilities around a conclusion. Running a society with models is much more like poker than chess. You might know the probabilities, but you still need a human at the other end to make a judgment.
That’s one of the reasons that it’s notoriously hard to measure how good a model is, other than test it against other models.
Let’s say I have a model that picks whether a stock is going up or down 90% of the time. You get unlucky a few times in a row and dump that model for another model that gets the direction right only 60% of the time, which then makes you money for the next week. Gleefully, you decide you made the right decision, never knowing the “truth.”
Those aren’t scare quotes on truth, I really mean that “truth” is hard to define. Is the 90% model “better” than the 60% model? I suppose so, but if you had a model that was 91%, it would be better still…in most cases.
A way to compare models is to back-test them. You take a few years of historical market data and run the two models through it with some metric (in this case “did you guess the direction right?”) and then have a coherent way of comparing the models.
If you’re an elected official, you didn’t have the chance to backtest models, so you had to choose one and go with it. For COVID-19, we had to make assumptions about transmission rate, virulence, how long it was contagious, even how it was spread. Because governments lacked data, the error bar on the outputs was enormous. When some (scientifically valid) models are predicting millions of deaths, even if other models are predicting a much lower death rate, it’s perfectly reasonable to order a swift and immediate shut-down. And that’s exactly what many governors did.
6 weeks later, people are starting to question whether that action was right. Such reasoning done in hindsight is unfair. We should deal with the hand we were dealt today, not lament about how we could have played an earlier hand better.
We should, however, hold officials accountable to their models. We should get educated about the models they’re using to run society by asking questions, not about the policy conclusions, but about the models themselves:
- What kind of model are you using?— all models and the data that feed them ought to be open-source, free to the public.
- What are the goals of the model?—different goals like “zero life lost” vs. “minimize death” vs. “overwhelm hospitals,” you’ll get very different policy choices.
- Where is data incomplete?—every model will have to make assumptions to make up for lack of data. That’s fine, it’s important to know what they are.
- How are you going to collect more data?—better data means better models and governments should take an active role in creating better datasets.
- Has your model changed?—models are not static. Every time more data is injected, models will get better. We can ask how they got better, what areas improved the most based on new data?
No matter where you stand philosophically on policy and whether your state is opening up or shutting down, you should understand the implications on the model. If you’re remaining in lockdown, how many lives does the model predict we save? If the model is uncertain, how will that uncertainty be reduced with data? If your state has decided to open up, how many deaths does the model predict? How will you know if you need to lock down again?
Sweeping policy decisions might very well be the only way we save ourselves from this and future pandemics; or at least, that’s what the models tell us. Rather than just trusting in the “science,” we’d best start educating ourselves about how the models work.
