Using our dataset of:
- 38094 rallies over
- 888 games
- 428 timeouts occured in just 269 of those games
(i.e. 70% of games featured no timeouts)
By comparing the rally win rates of teams before and after they call a timeout, this analysis has established that, on average, timeouts improve performance by roughly 7%.
That relative change in rally win rate is captured by the "performance" column in the table below.
Because the serving side is at a disadvantage (winning only 43% of all rallies), this was considered by constraining rallies to only those when the timeout-taking side was serving and returning (as shows in the "serving" and "receiving" columns, respectively).
Another way of incorporating the server/receiver disparity is to directly adjust the metric depending on the share of rallies they enjoyed the receiving-side advantage, which is captured by the "advantage adjusted" column. For example, if the team was mostly receiving before the timeout, and mostly serving after, but their rally win rate remains constant, then that 0% performance boost is adjusted up to become positive (the scale of which depends on the relative serving rates and winning rates). As shown in the table, this disparity does not have a significant impact on the final results.
Summary statistics grouped by skill level:
The raw dataset was generated by users of by Alex Spancake's "pklmart" platform. (www.pklmart.com):
Data was pulled into a python3 jupyter notebook "timeout_analysis.ipynb" directly from the pklmart postgres database in AWS:
Each timeout is individually scored based on the difference between the team that called it's rally win % before and after the timeout.
The rallies included in the calculation for any given timeout are rallies in the game (not match) before and after the timeout, bounded by any other timeouts. This bounding was applied because other timeouts are opportunities to change strategy, and thus counteract the impact of the timeout at hand.
Example calculations are provided for a sample game below, with rallies for one particular timeout highlighted in yellow in the figure on the left.
The figure on the right is just a more detailed record of the game's progress, including the server #, sideouts, etc. Whereas the figure on the left is a simplified view with only the rally winners depicted:
Histogram of rally win rate increases across all timeouts:
If we filter rallies to only those in which the timeout-calling team was serving:
Again, but filtering by receiving rallies instead of serving:
The following 3 charts corresond to the same set of results as the 3 preceding charts, grouped by player skill level (using DUPR rating for amateurs).
The same results depicted in the 3 immediately preceding charts from the perspective of box plots:
Serving rallies:
The charts below quantify the effect of timeouts on unbroken point scoring streaks, i.e. consecutive rallies won by serving team.
Note that rate of 0-length servin win streaks, depicted in the far left columns, essentially reflects the disadvantage of being server, considering this amounts to the percentage of all rallies won by the serving team, and it turns out to be below 50%.
The table below ranks pro doubles teams effectiveness of timeouts, where timeout score is the product of their [timeouts per game] and their [average rally win rate improvement after a timeouts], normalized from 0 to 100: