Home > Articles > Thinking in Bets for Data Scientists

Thinking in Bets for Data Scientists

Data scientists are uniquely positioned to provide leadership on their teams around risk and uncertainty. We are trusted by our coworkers to have an understanding of experimentation and data-driven decision making. This trust can be leveraged to improve processes, decisions, and (ultimately) the output of our teams.

In Thinking in Bets, Annie Duke describes how to make good decisions, informed by her time as a professional poker player. Life (she argues) and data science (I argue) are like a game of poker. In some games, like chess, each player has perfect information. They know where all the pieces are, what moves those pieces can make, and the conditions for winning. But poker is a game of imperfect information. Each player knows only the cards in their own hand. While they can intuit things from the body language or play style of other players, that intuition is not perfect. Some players are really good at bluffing. Some behaviors are easy to misread.

Poker, thus, requires players to make decisions in a system with imperfect information (often with high dollar amounts on the line). Doesn't this sound like life? The stakes are high, and we don't know what the future holds, but we have to make some decision. By Duke's definition, that's what a bet is: "a decision about an uncertain future".

In the face of uncertainty, we data scientists resort to experimentation. We can determine the best color for a button or the best copy on a page or any number of other things by running a well-designed experiment. While this is valuable work, it only scratches the surface in terms of where we can apply good decision making.

Running experiments is only useful insofar as they help us unlock real value (i.e. moving our OKR's or KPI's). Think about it like driving a car. When you press the accelerator, the tachometer shows an increase in the RPM at which your engine is turning. This turning is then put through a system of gears and eventually spins the wheels [0]. The ability to run experiments quickly is akin to being able to turn the engine quickly. Without a good system for choosing the right experiments and leveraging their results, we limit our ability to impact the business. When we view decision making as a process of choosing the right bet, we can choose strategies that help us make better decisions and have greater impact.

Red Teams

The strategy from Duke's book that I found most directly applicable to my work as a data scientist is Red Teaming. Established after 9/11, these teams have as their express goal "arguing against the intelligence community's conventional wisdom" [1]. By "spotting flaws in logic and analysis," red teams help drive intelligence agencies closer to both the truth and a proper understanding of the uncertainty in analyses [2].

Within weeks of reading this book, my team and I happened to be working on understanding our KPI's better, which involved some new analysis. This seemed like a perfect time to apply a "red team" strategy -- as someone would posit a result, I would see if I could disprove it. Whether I could or couldn't, I reported on both. And then other members of the team would try to prove or disprove my result! By this collaborative process, we came to understand the truth of the situation where we could have easily misled ourselves.

If you want to try this out, here's a few techniques I've found useful:

  1. Explicitly try to disprove an analysis. If you have sufficient time, working to show how an analysis is wrong can be really valuable, even if it withstands scruitiny. You will likely find some small issues in the way something is calculated, some ambiguous terms or metrics that could be misunderstood, or an invalid assumption. These findings can lead to further quantification of their impact on the original analysis. In this way, we can gain a better understanding of how confident we should be in said analysis.

  2. Try to reproduce an analysis. Avoiding looking at the code for the original analysis, try to get the same result via a slightly different pathway. If the original author used raw log data, see if you can answer the question using the data warehouse. Come up with new metrics that should move in the same direction as those used in the first analysis. If two people come to the same conclusion independently, we gain confidence in that conclusion.
    You've got to be careful with this one! Knowing the hypothesis that is being tested can skew the analysis you do (even unconsciously) [3].

  3. Answer an adjacent question. Sometimes there just isn't enough time to fully reproduce or disprove an analysis. In these cases, we can test an upstream cause or a downstream effect instead.
    For instance, if our analysis finds that sales of trucks decrease when gas prices are high, we could look up years in which gas prices were high and see if truck sales were down.

While these techniques are most effective when applied by a separate person who hasn't been influenced by the same process/data as the original author, I have found value in explicitly shifting my perspective to "red team" myself. Working specifically to disprove my own analysis, I end up understanding the results in greater depth.

Be humble

Humility is a key element of truth-seeking. We must remember that the point of our work is not to prove to our teammates that we are geniuses; we're trying to produce some positive result for our employer. We are more likely to find the truth when we seek it rather than pursuing our own glory.

I believe that sometimes our drive to compete gets in the way of our humility. I am not a particularly competitive person by nature, but if you're reading this and you are competitive, Duke has some advice for you:

Keep the reward of feeling like we are doing well compared to our peers, but change the features by which we compare ourselves: be a better credit-giver than your peers, more willing than others to admit mistakes, more willing to explore possible reasons for an outcome with an open mind, even, and especially, if that might cast you in a bad light or shine a good light on someone else. In this way we can feel that we are doing well by comparison because we are doing something unusual and hard that most people don’t do. That makes us feel exceptional.

When red-teaming my own analysis, I sometimes find things that cast doubt on it. Sharing my results, I'm tempted to leave these observations out. The desire to present my findings in the best light possible is (I believe) a natural one, yet one I must work against. Duke writes that "if we have an urge to leave out a detail because it makes us uncomfortable… [it is] exactly the detail we must share."

Bringing reasons to doubt to the table along with the analysis itself helps know what information we need to get. Sometimes a little bit of additional analysis can alleviate the doubt. Sometimes the concern will turn out to reflect a small enough risk that we don't need to address it. And sometimes the only way to get more information is through an experiment. The important thing is bringing uncertainty to the table so we can address it directly.

Communicating uncertainty

We communicate about uncertainty all the time. When asked if we're going to an after-work social event, for instance, we say that we "might go" (which typically means we are definitely not going) or that we will "probably go" (it's a bit of a tossup). These phrases are examples of words of estimative probability, or WEP's. In colloquial usage, these casual WEP's are just fine, but they are less helpful when we're trying to make a good decision.

For one, different words mean different things to different people. Andrew Mauboussin's research shows that words like "maybe", "probably", and "usually" are interpreted to correspond with wide ranges of probabilities depending on the audience. For instance, when someone says that an event "might happen", her or his audience could interpret that as an event with probability between 25% and 55%. That's a huge range!

By using WEP's in communication, we run the risk that our audience will misinterpret the likelihood we think a certain event has. But there are ways to overcome this. Mauboussin advocates for explicitly giving a percentage alongside words of estimative probability. This approach is used in the medical research field, where institutional review boards require researchers to inform people of the risks in treatments using WEP's [4]. These words should be accompanied by a percentage; for instance, a researcher might inform a participant using language like, "This side effect is rare (will happen to less than 1% of subjects)".

Parting words

Certainty is alluring. But as data scientists, we should know better! The world is filled with uncertainty, and only by defining and quantifying it can we drive toward an accurate understanding of reality. This understanding, then, enables us to make higher quality decisions. And this improvement in decision-making doesn't have to stop at the individual; we can bring this idea to our teams, departments, and companies.


  1. I think this is how it works; I'm really not much of a car person.
  2. Neal K. Katyal. 1 July 2016. "Washington Needs More Dissent Channels", The New York Times
  3. Ibid.
  4. Duke here references Richard Feynman, but I can't find a direct citation. Still, this seems to jive with my own experience.
  5. University of Tennessee, Chattanooga.