One of the most common refrains heard from broadcast crews (or in living rooms) of any basketball game is that winning the first and last few minutes of each quarter is crucial for success to establish the flow of the game, keep/steal momentum, etc. I have taken that conventional wisdom for granted as truth without ever questioning its veracity...until now.
Entirely defined by myself, critical periods are the first and last two minutes (120 seconds of gameplay) of each quarter. This will include possessions that have finished within that 120 range, but not those that started but did not finish with in the 120. For example, a possession that ends at 10:05 on the clock (out of 12:00) would fall under the critical period range, but one that starts at 10:05 but ends at 9:52 would not.
The number of critical periods won is positively associated with winning the game
This is a very simple hypothesis, but one that must be investigated first prior to continuing the analysis. On the surface, it should be true, as it seems reasonable that winning any lengths of time in a game is likely predictive of a win. Figure 1 presents the count of critical period wins by game outcome for the Atlantic Division in the 2017-2018 season.
We can see that there is a general increasing trend for winning as the critical period win count increases. Table 1 shows the likelihood of winning by each critical count win value. The way to read this table is - “You are 0.45 times as likely to win as lose the game if you win 1 critical period”.
Figure 2 is Table 1 in graphical form for those so inclined.
Based on the statistical evidence and common sense, I am comfortable stating that the more critical periods you win, the more likely you are to win the game. Keep in mind that this is unadjusted for literally any other factor.
Winning critical periods is more predictive of winning than winning random two minute sections of playtime not within the critical periods.
The second hypothesis takes the results from the first hypothesis and builds on it by comparing the critical periods to other non-critical periods. For this analysis two random two minute sections of play time (per quarter) that do not overlap with critical periods will be chosen and rated as won/lost in the same manner as the critical periods. Let’s look at the random two minute sections using the exact same graphics as the first hypothesis.
If you are having trouble differentiating the two sets of figures, you are not alone. Based on visual inspection only, the information is remarkably similar. However, any time you take a random sample, you could always just have gotten lucky/unlucky, and so we need to repeat the random period generation a large number of times to smooth out any weirdness. Figure 5 is a duplicate of Figure 2/4 with the critical periods and the random periods on the same graph, but this time the random drawing of non-critical period start times (still two per quarter non-overlapping with critical periods) was done 100 times per game, instead of just once. Because what we are measuring is a ratio, this does not throw off the values.
After performing tests for difference in slope and difference in intercepts, we get the same results as a visual inspection — the relationships appear to be nearly identical.
Conclusions & Next Steps
This was fairly interesting, as I expected that the “critical periods” would have increased importance compared to the random periods. This starts to suggest that the importance of these times are overstated.
However, there are at least three massive caveats for this analysis. First, the count of the won periods does not distinguish between winning by +20 or +1. Second, the relationship may not be identical for close games vs. blowouts, or even for different teams. For example, I could see a scenario where the Warriors win fewer of these periods than the Sixers, let’s say, but that the Warriors win those periods by huge margins. Style and quality of play can certainly have an influence on these sort of large scale analyses. Last, I chose 120 seconds as the cutoff point, but the cut points could also be one minute, or five minutes, etc. The shorter the length of time you would think would be more 50/50, whereas the longer times might favor the eventual winning team.
The next installation on this topic will break out by team (Atlantic Division) and by specific critical periods to see if the signal to noise is the same across various ways of splitting the data. If you have any thoughts on directions to take this, let me know in the comments below.