What Will It Take To Score at Women’s NCAA’s – A Review

2025 NCAA Division I Women’s Swimming and Diving Championships

• March 19-22, 2025
• Weyerhaeuser King County Aquatics Center — Federal Way, Washington
• Short Course Yards (25 yards)
• Meet Central
• Official Psych Sheets
• SwimSwam Preview Index
• Live Results
• Full Results (PDF)

Prior to the kick-off of NCAAs last Wednesday we put together a model to try to predict the times required to make the finals, and it turned out that the model did a pretty good job. Positive differences here indicate that we predicted times slower than the actual time, and negative indicate faster predictions than the actual time.

There was a fair difference between the actual/prediction differences for 8th and 16th, with the ‘A’ final cut off predictions being better on average. We’ll see why a little later on.

500 free

With how quick this event had been this season, it was a bit of a shock that the times required to final were almost identical to last year. The 8th seed in 2024 was Abby McCulloh with a 4:37.18 – this year it was Julie Brousseau in 4:34.59, and McCulloh’s time would have entered her only 18th.

A faster psych sheet should indicate a faster set of results. This is the premise that the model is based on, and can be a reductive one. However, the 500 was the only event consistently predicted to be faster by the model – there was something here that the model was not able to pick up on.

100 butterfly

The 100 fly was the other way around from the 500 free. We predicted the event to be slower at both positions (and in general) by not much less than we’d predicted the 500 to be faster.

The psych sheet times were a little slower this year, but the times to final were almost identical – it took 51.07 and 51.34 last year. There was a big change with the 100 fly scheduling this year however, as it has swapped with the 400IM and is now the first event on Day 3. This makes it a little easier for those with doubles, and may explain the uptick.

100 backstroke

This is only with regards to the ‘A’ final time – we were bang on with the 16th place finisher. The time for 8th however was absolutely rapid. It was two tenths quicker than the 8th seed, and a full seven tenths under the time required last year.

We saw Miranda Grana’s 1:48 in the 200 backstroke ‘B’ final, but the speed of the 100 prelims was the bigger shock. Four of the top five returning swimmers from last year only made the ‘B’ final, and only two of those would have made the ‘A’ final this year with their 2024 time.

So it looks like the difference here was an event getting significantly faster, rather than a limitation of the model – certainly not an unwelcome development.

How Accurate Were We Overall?

The model didn’t just predict the two positions for each event that we’ve looked at here. We used it to predict the top 50, and we’d like to see how good it was across the board. To do so, we’ll take a look at something called the residuals.

The residuals are the difference between a prediction and an actual value for a given ‘x’ value (in our case the ‘x’ value is the position). In a good model these will be randomly distributed – what a statistician might call ‘noise’.

If there is a pattern in the residuals, say something like the following, there’s likely something in the data that isn’t taken into account in the model. We’ve got residuals for four of the events here, showing two different types of patterns in the residuals between them.

100 backstroke

100 fly

100 breaststroke

500 freestyle

Each of the other events has a similar pattern to one of these, indicating that there is something in the data not accounted for, and does explain why we seemed to have faster predictions at lower positions. Even so, the model was pretty accurate. The goodness of fit values were above 0.9 for almost every event, and none of the residuals were particularly big – most were within 3-tenths of a percent.

For those of you interested in the data, you can see the dataset at the google sheet here. You may notice that a cubic fit for the residuals does appear to give a closer fit than either linear or quadratic, but it seems not by enough to make up for the complexity it introduces.