Analyzing NBA Timeouts
The practice of NBA coaches calling timeouts to stop runs has been a subject of analysis in basketball discussions. Play-by-Play (PBP) data from 2019–20 to 2023–24 has been utilized to determine whether there is any evidence supporting that some coaches have suboptimal timeout strategies.
Introduction
Timeouts are an important component of in-game management during NBA games. Coaches use timeouts to organize a possession, change or keep “momentum”, reorganize offensively or defensively, challenge a referee’s decision, etc.
It is not uncommon to read or hear comments of fans criticizing a coach because of the timing of a timeout, a couple of Twitter searches with the words “Ham timeout” and “Malone timeout” yield these tweets:
This discussion has led me to ponder whether some coaches excel at calling timeouts, or if there exists an optimal moment for taking a timeout to stop opponent runs.
Understanding Play-by-Play (PBP) Data and Its Significance
What is Play-by-Play (PBP) Data?
Play-by-Play (PBP) data is a detailed log of all the events that occur during an NBA game. This data is recorded in real-time and includes information about every play, such as:
- Timestamp: The exact time when each event occurs.
- Event Type: The type of action (e.g., shot attempt, foul, substitution, timeout).
- Players Involved: The players who are involved in each play.
- Event Details: Specifics of the event, such as the outcome of a shot, the type of foul, or the reason for a timeout.
- Score Updates: Changes in the score as a result of the events.
I analyzed Play-by-Play (PBP) data obtained from nba.com using the Python package nba_api.
Timeout Analysis
As this analysis is mainly concerned with timeouts taken by coaches to stop runs, only the first 3 quarters of each game are considered in the dataset, as timeouts taken in late game situations are used mostly for different purposes.
The dataset contains a total of 37,434 timeouts taken, 32,830 were called by the non-run team, and of these, 21,649 successfully halted the opposing team’s run.
Calling a timeout stops the opponent’s run ~66% of the time (21649/32830).
For the timeout analysis I will introduce a new variable, calculated from PBP data: runs.
Runs are uncontested scoring stretches from one team. If we gather all the runs (excluding 4th quarter and OT) for the NBA 2023–24 season, most runs are just a single scoring possession that is eventually answered by the other team:
Team timeout performance
Let’s analyze the relationship between wins and runs after taking a timeout. Specifically, I examined how well teams perform in the runs after a timeout when not on a run.
The previous figure illustrates a positive correlation between wins and the mean run length of a team’s subsequent run after a timeout (ATO). The Pearson’s correlation coefficient for win percentage and mean run length post-timeout is 0.44, suggesting a moderate correlation. This implies that superior teams execute better after-timeout plays than inferior teams with Detroit’s 2019 team looking like an outlier.
We then define Standard runs as all runs where the team didn’t take a timeout and After TO as a team’s subsequent run after a timeout. The Pearson’s correlation for win percentage and mean standard run length is .60, indicating a stronger relationship than ATO’s runs.
The table below highlights teams that have better performance in runs after timeouts (ATO) compared to their standard runs.
The table indicates that the Miami Heat is listed three times (19, 20, 21) among the top 10 teams surpassing their standard mean run length, a testament to Erik Spoelstra’s consistent ATO calls and timely timeouts.
Halt vs non-halt timeouts
I hypothesized that better teams would halt runs more frequently than poorer teams. To test this, I analyzed the relationship between win percentage and halted runs, which yielded a Pearson’s correlation coefficient of .30, suggesting a weak correlation between the variables.
In analyzing the effectiveness of timeouts in the NBA, it’s notable that approximately 66% of these timeouts successfully halt an opponent’s scoring run. This statistic highlights that, under typical circumstances, timeouts can indeed be a valuable tool for provoking momentum shifts.
However, the question arises: Is the situation of halt timeouts vs non-halt timeouts different?
To examine this question, I created a set of features trying to capture the situation of the previous 5 runs:
- Previous Run Length
- Average Opponent Run Length
- Opponent Run Length Max and Min
- Average Own Run Length
- Opponent Run Length Max and Min
- Net Diff
- Scoring Rate (Note that this rate is positive for home team’s runs and negative for visitor team’s runs)
df['scoring_rate'] = (((df['home_score']-df['previous_run_home_score']) -(df['away_score']-df['previous_run_away_score']))/df['time_spent_run']).fillna(0)
Next, let’s examine whether there is a difference in the features’ average between timeouts that successfully halted runs and those that did not:
The analysis shows that the length of previous runs, whether by the opposition or our team, does not differ across halt and non-halt timeouts, as well as home and away timeouts. However, there is a variation in scoring rates and the net point differences, particularly indicating that calling a timeout when the net difference is small can be effective in stopping a run.
To determine whether there are significant differences between two groups for various features, I utilized the Mann-Whitney U test, a non-parametric test suitable for comparing distributions without assuming normality. This test was applied to home timeouts and away timeouts to take into account the differences in values (since net difference and scoring ratio are positive to the home team) and to each feature separately for two groups, timeouts that halted a run and timeout that didn’t halt a run.
Given the large number of features tested, I addressed the issue of multiple comparisons by applying the Bonferroni correction. This method adjusts the significance level by dividing it by the number of tests performed, thereby controlling the risk of Type I errors (false positives). The combination of the Mann-Whitney U test and the Bonferroni correction provided an adequate framework for identifying statistically significant differences between the groups, ensuring that our findings are robust and reliable.
The table below presents the results of the statistical tests conducted for all features. The “Rejection” column indicates whether the null hypothesis — that the distributions of the features are the same for both groups — was rejected.
We can observe that scoring ratios and net differences distributions are statistically different between both groups (halt and non-halt timeout) and for both home and away timeouts.
Methods
Since coaches may take multiple timeouts, each timeout is not independent of the others. Each coach introduces their own variability based on their unique style of calling timeouts. It’s important to note that we are only considering timeouts called by the team not currently on a run. I employed a multilevel logistic model to account for the random effects of the same coach calling multiple timeouts, and logistic because our outcome variable — whether the timeout stopped the run or not — is binary.
After deciding on the random effects variable, I began choosing fixed effects variables. I chose to use many of the variables we conducted exploratory data analysis on, including Net Diff in the run, scoring ratio for the previous 2 runs and the current run and win percentage of each timeout team.
I plan to use the variance of the random effects (coach) to determine which coach has a bigger effect on halted runs. The fixed effects coefficients will tell us which variables at the timeout moment increase, decrease, or have no effect on halted runs.
To account for the difference in home and away timeouts, I estimated two models:
βX represents the fixed effects in the model, while Qq represents the random effects for coaches.
Results
Fixed Effects
This plot shows the estimates for all the fixed effects in both models, colored by season. Being on the log-odds scale, estimates greater than zero means that the variable corresponds to an increase in halted runs, estimates less than zero means that the variable corresponds to a decrease in halted runs, and estimates equal to zero show no impact on batted passes. Variable whose intervals contain zero might not impact halted runs at all. We can see that only scoring ratio in the current run and net differences are different than zero for both models and all years.
Random Effects
We can observe that the coaches’ random effect is consistent across years and its intervals do not contain zeros, so there might be variance explained by each coach timeout taking strategy.
This table offers a summary of the top and bottom six coaches for the 2023–24 season. It presents intriguing scenarios: Michael Malone’s Denver Nuggets, despite being a good team, had a low estimate. Conversely, Jacque Vaughn’s Brooklyn Nets, one of the poorest performers in the 2023–24 season, are listed at the top.
We can observe that top coaches in the league like Kerr and Popovich consistently rank at the bottom of the estimate, so it may be worth for these coaches to take a look at the timeout strategy they are using and maybe hire Mike D Antoni to be an assistant coach.
Conclusion
Our analysis of NBA timeout data from the 2019–20 to 2023–24 seasons has provided valuable insights into the effectiveness and strategic use of timeouts in professional basketball:
- Timeout Effectiveness: Approximately 66% of timeouts successfully halt an opponent’s scoring run, indicating that timeouts can be an effective tool for shifting momentum in a game.
- Team Performance: We found a moderate positive correlation (0.44) between a team’s win percentage and their mean run length after timeouts, suggesting that better teams execute more effectively following timeouts.
- Situational Factors: The analysis revealed that the scoring rate and net point difference at the time of the timeout are statistically significant factors in determining whether a timeout will successfully halt a run.
- Coach Impact: Our multilevel logistic model demonstrated that individual coaches have a discernible impact on timeout effectiveness, even when controlling for team quality and game situation.
These findings suggest that there is indeed an art to calling effective timeouts, and that even successful coaches may have room for improvement in this aspect of game management. The data indicates that the optimal moment for taking a timeout to stop an opponent’s run may be when the scoring rate is high and the net point difference is still relatively small.
Future research could explore the long-term impact of timeout strategies on team success, analyze specific after-timeout plays, or investigate how timeout effectiveness varies in high-pressure situations like playoff games.
This study provides a data-driven foundation for NBA coaches and analysts to reassess and potentially optimize their timeout strategies, potentially leading to marginal gains that could make a significant difference in close games.
Thanks for reading
