Hot and cold streaks are inescapable in any baseball broadcast (or basketball, or perhaps any other sport for that matter.) When a player significantly over (or under) performs for a sustained period of time, it catches our attention. The announcer deems them “hot,” or “on fire.”

Why do we say this? We may just care about player streakiness to describe the recent past. More likely, though, is that we’re seeing a pattern that we hypothesize will continue into the future. If Aaron Judge has been on a tear in recent games, we have a heightened expectation of seeing a positive outcome in his next plate appearance. What does the data say about this? Should we care when the announcer says someone is on a hot or cold streak? Let’s find out.

Before we get started, I’ll also shamelessly plug Baseball IQ (the newsletter you’re reading right now.) If you’re interested in baseball research and analysis, consider signing up to get studies like this one delivered straight to your inbox!

Prior Work

This problem has been studied by many analysts across many sports. Should-be Nobel laureate Amos Tversky first studied the hot hand effect in basketball in 1985, finding “no evidence for a positive correlation between the outcomes of successive shots” (so, no hot hand). Some follow-up studies have found positive effects, but usually with small magnitude (so maybe there is a hot hand effect, but it’s small enough where we can debate whether it really matters.)

Similar studies exist in baseball. This analysis focuses on the approach taken in The Book, by Tango, Lichtman, and Dolphin, though many other studies of hot streaks in baseball exist. Tango et al. look at sustained periods of high or low weighted on base average (wOBA), and see if the following one or several games for these players are similarly hot or cold. The study finds that both extremely-hot and extremely-cold players revert almost entirely to their mean expected performance following a streak, suggesting a hot-hand effect that exists, but is so small that we shouldn’t really care about it in practical terms.

This post is going to replicate the methods used in this 2006 study on a much larger dataset, extend its methodology by controlling for additional factors such as park effects, and see if its findings still hold today.

Methodology

The data used in this analysis come from Chadwick