Nate Silver’s book, The Signal and the Noise had been on my radar for years, but I never got around to reading it. Since he is about to release his new book, however, I decided it was finally time to dive in. It did not disappoint. Silver is best known for his sports and political forecasts (in 2012, he successfully predicted the results of the US presidential elections in all 50 states). The book discusses predictions in a wide range of fields, highlighting successes and failures, and his thoughts on how predictions can be improved. In this post, I discuss my favorite parts of the book and my takeaways.

One of the things that struck me reading this book, is how recent many of the developments in the field of predictions have been. It’s been just over a decade since Silver became famous for his political predictions. The Oakland As gained notoriety for using sabermetrics in 2002 (Moneyball was published in 2003), even though baseball statistics have been recorded since the National League was founded in 1876 (and even before that in some cases, although with varying accuracy).

The fourth chapter was probably my favorite of the entire book. If you’re like me, you often hear your colleagues, friends and family complain about how weather predictions are “always wrong”, or “getting worse”. As Nate Silver argues, weather forecasting is one of the greatest successes in the field of predictions. Meteorologists have essentially determined how to computationally predict the weather, and as computing power has increased, so has the accuracy of their predictions. The main source of uncertainty in weather predictions comes from imprecise measurements since very small variations in measurement can have a huge impact on the weather forecast. This, Silver explains, is a key idea of chaos theory. Silver also explains that even though scientists can produce excellent weather predictions, the predictions that get reported in popular media (as opposed to those produced by government agencies) tend to be biased toward rain. This is because people tend to get upset when it rains despite a low prediction, so the media are incentivized to skew their forecasts.

I also especially enjoyed the chapter discussing the data-rich field of baseball, a field where Silver has made a name for himself by creating his PECOTA system, a system for predicting player performance. In this chapter, Silver makes the point that the best predictions come from a mix of data and human judgment. You might think that with all the data available in baseball, scouts have become less relevant, but they are more relevant than ever because they can see things that the statistical models can’t. For example, they can get a sense of a player’s personality and temper, and judge whether he’d be a good fit in the team. This point is repeated in other chapters of the book, such as in weather forecasting, where meteorologists combine data from their models with visual observations to validate them.

The other thing that struck me while reading this book was Silver’s take on the Bayesian vs frequentist approaches to statistics. He is a firm proponent of the Bayesian approach, going as far as qualifying the frequentist approach as “When Statistics Backtracked from Bayes”. His main objection to frequentism is that frequentists claim the only source of uncertainty comes from surveying samples of a population as opposed to the entire population, which, according to Silver, is an extreme oversimplification. He also points out that defining a sample population is not always straightforward (“What ‘sample population’ was the September 11 attack drawn from?”) According to him, the reason frequentism came about was because frequentists were looking to eliminate the subjectiveness of Bayes’ prior probability. However, Silver points out that no matter what prior probability you start from (except for 0% or 100%), given sufficient evidence, probabilities will eventually converge.

Interestingly, even though I learned about Bayes’ theorem in school, I did not become aware of the Bayes vs frequentism perspectives until much later, through this article by Cassie Kozyrkov. Cassie has a very different approach. She advocates that you should not shun Bayesian or frequentist thinking, but rather, you should decide on which approach to take based on what type of decision you’re making. If you have a no default action - choose Bayesian, otherwise, choose frequentist.

In fact, I’ve found many articles and posts online that point out that despite his criticisms, Nate Silver is actually a frequentist, as he uses several frequentist techniques in his forecasts. (See here, here or here) These errors do not change the fact that this is an excellent and interesting book and are easily overlooked.

The entire second half of the book is centered around applications of Bayes’ theorem. I especially enjoyed these examples, as they help explain Bayes’ theorem in a very applied way, which contrasts with how it is typically taught in schools (at least in my experience). My favorite was the chapter on chess, where Silver discusses how Deep Blue was able to beat Garry Kasparov.

Nate Silver’s The Signal and the Noise remains relevant in 2024, more than a decade after it was released. I would recommend it to anyone interested in predictions. I was fascinated by how these statistical techniques can be applied to so many different fields, and the examples provided made me realize that in many fields, there is still much room for improvement when it comes to applying statistics. It’s also an excellent reminder that data models are not the ultimate solution and that oftentimes a combination of human judgment and statistical techniques provide the best predictions.

Rating: 8/10