Welcome to the 2022-23 College Basketball Season

First as always, let me welcome you to PPB's coverage of the 2022-2023 college basketball season. It's another chance to accomplish our goal of perfectly predicting a NCAA bracket. Second, I'm going to start this season in the same fashion as past seasons: A review of the lessons learned followed by a grading section. Let's get started!

LESSONS LEARNED FROM 2022

Though I can point to a number of unique and isolated teachable moments (mostly detailed in the grading section), I think the most ubiquitous and detrimental problem was subjectivity, where judgments were being made based upon my own thoughts, intuitions, and feelings rather than the data and models.

In my defense, I relied on my intuition and 20+ year experience to fill in the gaps for the 2021 tournament, when data and models were less reliable due to the experimental nature of that season and that tournament. In 2022, when data and models were more reliable than 2021 (although probably not as reliable as pre-2020 seasons), I should have realized there would be less gaps to fill. Likewise, not all subjective judgements were incorrect. For example, MIA to the E8 based on Meta Analysis, ignoring the Conf OSUS rules for UCLA/USC, and disregarding the SGLT Regression predictions due to data typos (more about this in grading section) were all winning subjective calls. But for every subjective call that worked, there was at least one subjective call that didn't work. The prediction to advance TEX to the E8 based on Meta analysis just like MIA was pure folly. Even though both teams fit the criteria, the probabilities of advancement were not the same. In MIA's octet, there was a vulnerable 2-seed in AUB (Predictive OSUS), a vulnerable 3-seed in WISC (Champ Profile and Hist Seed Comps), a vulnerable 6-seed in LSU (Predictive OSUS and Conf OSUS), and a vulnerable R64 opponent in 7-seed USC (location). In TEX's octet, only 2-seed UK showed vulnerability as a R64 upset-victim (Predictive OSUS). No other data point hinted at an easy advancement for TEX other than the lack of historical comparison for the octet (as well as region). So, yes TEX could have been on the same Meta-identity list as MIA, just not ranked as high on the list. This example brings me to my next point.

Instead of broadly painting the issue as subjectivity, I would more specifically claim the problem to be a failure to thoroughly evaluate it (as was done in the previous paragraph). Yes, I do evaluate and even reject my subjective views (explained below in the QC/SC Grading Section), but more importantly, I'm able to evaluate it because I take very deliberate actions to document my subjectivity (i.e. - articles like Opening Week Thoughts and "Initial Reactions to the Bracket Reveal" which has been copied in the Appendix below for historical reference). In fact, not re-reading the "Opening Week Thoughts" article could be the biggest failure to evaluate subjectivity as it contained numerous profitable insights: KU as a potential NC via 5YOD profile, NOVA as a F4 team, GONZ as a pretender, PUR and MICH as guaranteed 1-win minimums, and OHST beating interior-oriented teams like LOYC but losing to guard-oriented teams like NOVA. In the conclusion of that same article, I, in an ironic-yet-prophetic fashion, stated the following:

"The one thing I'm taking away from it is self-reflection. This gives me an opportunity to record my thoughts and perceptions (nicer terms of describing my biases). When I get to BCW, I don't want my biases clouding my bracket analysis, and knowing my biases is probably the best way to approach it."

So as early as November, I had already stumbled upon (albeit accidentally) what would be the eventual problem of the 2022 predictions. Theoretically speaking, these November premonitions are original, instinctive and devoid of any influence from data, experience, or the bias of others (akin to gut-reactions). Even more, they are formed far outside of the pressure and haste of Bracket Crunch Week. This would seem to give them control-group status when examining the subjectivity arising through the course of the season. It doesn't mean that my predictions in November are the correct ones or even the pure ones, it simply means they are the measuring stick as to whether my subjectivity remained or changed over the season and why did it remain/change? Needless to say, I will be doing at least one article pertaining to Opening Week Thoughts for this reason.

With the major takeaway of last season addressed, let's look at the individual circumstances where predictions went right or wrong and why.

GRADING THE PREVIOUS YEAR

In the grading section, I am going to do things a little differently. I am going to give two grades: one for the model and one for the bracket scientist. The grade for the model will reflect standards such as accuracy and win-value. The grade for the bracket scientist will reflect standards such as accuracy of the work, value of model interpretations, and foresight of recommendations.

Championship Profiles

Model: B+
Scientist: C-/D+

This model is solely for picking the national champion, which means its win-value is no higher than six (6 out of 63). Historically, its accuracy is mostly ballpark-esque, meaning it will get you close to the actual result (two to three potential candidates, but more work to identify the actual one). Nonetheless, it is a great starting point. I could have given it a B or even B-, but I settled on B+ because I blamed the lack of identifying KU more on the scientist than on the model. Upon re-reading the article for NC criteria, I re-discovered an exception for the 6-game in-season winning streak against tournament competition. If a team won five consecutive games against Power Conference teams and this 5-win streak includes conference tournament games, then this exception qualifies as a pass. 2000 MIST qualified under this exception, and unfortunately, I failed to recognize that 2022 KU also qualified under this exception. This exception would put KU in the Tier 2 list with NOVA/TENN/PUR/UK/PROV, with their only disqualifiers being a SO starting PG and lacking PG production. KU would have been the 2nd best Tier-2 choice behind NOVA, and I wasn't high on NOVA as a NC (though I loved them as a F4 pick). Even worse, I re-evaluated the model in the off-season (all criteria, all exemptions, and all disqualifiers) and KU ends up being a 0-DQ candidate, placing them in a tier by themselves above GONZ. You'll have to wait for the article later this month.

OSUS Predictive

Model: B- (great for road-maps, terrible for R64 match-ups)
Scientist: C+

As for the model, it did pretty well with pathing, which I think OSUS models may be the most equipped to do. By pathing models, I mean road-maps as to how the octet/region will play out. For example, it predicted/suggested that GONZ wouldn't advance to F4, either DUKE/TXTC would advance to F4, lack of a historical match for the BAY and UK octets and their results weren't matched historically either, ARI potentially not making F4, NOVA most likely making F4, KU potentially making F4, and AUB getting upset before S16. As for the model being used as a R64 match-up predictor, the accuracy was far less impressive. Statistically, it was on-target with 3-seed WISC not making S16, 4-seed PROV making the S16, and not-so-hot 6-seed LSU losing to 11-seed IAST, but it was dead wrong on OS 12-seeds versus US/AS 5-seeds (incorrectly predicting NMST and RICH to lose) and dead wrong on US 10-seeds vs OS/AS 7-seeds (incorrectly predicting LOYC and USF to win). In the final R64 matchup prediction, it was a 50-50 split with the guaranteed 8-seed lock SDST being wrong and the 75% guaranteed 8-seed UNC being right. Lesson learned, I think this makes two consecutive years of wild and inconsistent results using the OSUS predictive model as a R64-picker, and maybe the model is less reliable in the R64 in wild years since 2021 and 2022 would qualify for this label. Knowing which years the model works and which years it doesn't is probably more important than knowing its results.

OSUS Poll-Based

Model: C+
Scientist: B

As for the model, this is arguably the best year it has had, as it went 4-0 in R64 match-ups and 21-9 in seed-based expected wins. The only reason I can't move the model grade into A territory is because our goal is perfection (which would have to be 30-0 instead of 21-9). As for the grader, I did the only thing I could do by saying "This is a high-risk model, so use it as a last-resort/tie-breaker pick."

OSUS Conference-Based

Model: C+
Scientist: B+

In the final write-up, I admitted that this model was harder to put together without knowing the answers ahead of time. Thinking through the formal logic of all the OS/US possibilities was a challenge in its own right, never mind that circumstantial probabilities also existed with the outcomes of the play-in games. As fate would have it, the one game which all of the circumstantial probabilities existed (DAME/RUT) played on Wednesday night (If it played on Tues, I would have had an extra 24 hours to work through the logic). As for the model, it wasn't as historically accurate as usual (for the BEC, this may be due to an incomplete conference season). I also lowered the model grade because the model was better at protecting from mis-picks than it was at producing actual predictions. I gave the scientist a higher grade than the model because I ignored the UCLA/USC OSUS prediction (and another one) in my final predictions.

QC/SC Analysis

Model: B-
Scientist: C+

As for the model, it was predicting an insane tournament, along the lines of 2014, 2018, and 2021. The big misses of the model was the SC predicting seed-expectation under-performance for 4-seeds (expected eight wins) and 7-seeds (expected two wins), but in actuality, 4-seeds met expectations with eight wins and 7-seeds exceeded expectations with three wins). Of the correct calls, the 5-seeds matching or exceeding seed expectations was the closest to failing if not for the three-win run by HOU (who happened to be a 1-seed in disguise). As for the scientist, the big bulk of the C+ grade comes from trusting the model and abandoning the pre- and in-season hunches regarding mean-reversion. All season long, I thought the 2022 tournament would be a lot calmer (13-15% M-o-M rating) due to mean reversion, but the QC never supported this thesis, so the only choice I had was to abandon the pre-season thesis. A major drag on the scientist grade comes from the scientist posing this question then failing to address it: "What is the significance of the 2022 QC entangling itself with the historical QCs along the #7-#14 spots?" This was a critical question raised in the write-up but never answered by the bracket scientist. An answer could have been huge value considering the fact that two of the 2022 F4 teams (DUKE and NOVA) were among the teams in these eight spots.

Historical Seed Comparisons

Model: B-
Scientist: D

If OSUS Conference Based is the most mentally draining model, Historical Seed Comparisons is the most time-consuming. With only three-and-a-half days to prepare multiple models and write-ups for each, I may or may not do this analysis again because of the amount of time that I would have to dedicate to it. If I remember correctly, it took me an entire day (Tuesday) to complete. A lot of the comparisons were fairly accurate: 1v16, 8v9, 6v11, and 3v14 (the 11-seed analysis was a no-decision due to no matching years). The 5v12 and 4v13 were inaccurate, costing a loss in each seed group. The 7v10 and 2v15 models were also fairly accurate, indicating the highest probability for upset was STPC over UK in R64 and MIA over AUB in R32 (both happened). Unfortunately, yours truly is responsible for over-analysis in the 7v10 match-ups and over-emotionalizing the 2v15 match-up. Doubting that "back-to-back years of 15v2 upsets would be unlikely" is single-handedly the most embarrassing comment in the Final Analysis article, and it's why the scientist's lowest grade comes in the model. I should have at least given proper consideration to the upset possibility since the UK octet (like the BAY octet in the same region) had no historical comparison. If UK was destined to advance to the F4 from that region, it would have had matched historical octets that produced that particular outcome. Again, this one is on me and one shining example of my failure to properly evaluate my subjectivity.

Returning Participants Model

Model: B+
Scientist: A+

The only reason this model falls short of a grade in the A-range is the win value. It predicted a F4 team (2-seed NOVA), which only counts for four wins (4 out of a total of 63). The other value of the model is the confirmation of tournament insanity. The scientist gets an A+ for digging up this information, although an A grade would also be fitting as I still haven't unlocked any patterns or insights from this model.

Seed-Group Loss Table

Model: C
Scientist: B-

To say the least, this is not what I was hoping for. Correct: 1-seed F4 count, 2-seed E8 count, 3-seed WTOT, 3-seed F4 count and 4-seed F4 count. Incorrect: 1-seed E8 count, 2-seed WTOT, 2-seed F4 count, 3-seed E8 count, 4-seed WTOT, and 4-seed E8 count. Overall record: 5-6 with no decision made for 1-seed WTOT. If I remember correctly, I prefaced this section in the final article saying "I want to focus on this tool in the off-season because I want it to work" to subtly remind myself of typos I found and corrected in two of the years (both typos involved the 5- thru 12-seeds, not the 1- thru 4-seeds). From the off-season reflections on this error in the data, I've realized this change would have altered the inputs for the regression model (the regression formulas used in the model would have been constructed from incorrect data inputs), but the matching method should have been unaffected. The use of inaccurate data could be a potential explanation of the poor results. I went B- for the scientist because the model's methods predict round counts in the E8 and F4, but the scientist misused the model to predict R64 and R32 upsets. Most importantly, I am trying for an update article on this model, but for now, updating it to spreadsheet format from pen-and-paper format is the current priority.

Meta Analysis

Model: N/A
Scientist: B-

It's kind of unfair to punish the model when the scientist didn't provide any examples of teams that fit the meta-criteria, so I didn't give the model a grade at all. In fairness, it is a new model, and I have three different spreadsheets with all of the information, so a consolidation job may be required before any useful prognostic adventures can be taken. Though I identified MIA as a Meta-based sleeper pick, I gave myself a B- grade for the lack of follow-through in the actual Meta Analysis section of the Final Article.

THINKING FORWARD TO THE CURRENT SEASON

If you've made it this far in the article, I at least hope your head doesn't hurt like mine. All joking aside, there's always something to learn from the previous year's mistakes. As stated before, I'm definitely writing a Opening Week Thoughts article for every reason listed in this article, and I've already scheduled a date for it as well. I really want to do a second one as well, and December may fit the calendar since pre-season tournaments should be finished and conference-play should be starting. If I'm not mistaken, four of the Power-6 conferences have 20-game conference schedules, so that amounts to ten weeks plus an eleventh for the conference tourney, so December seems like the right fit.

Another idea would be a selected reading plan for the week prior to BCW. I already re-read certain articles like the Jan and Feb QC Analysis so that I can follow-up on questions and ideas that arise from them. It would make sense to re-read all of the "Mistakes Made/Lessons Learned" articles from all past seasons to avoid making the same mistakes this year that derailed previous attempts.

Finally, I think a row/column table that summarizes the projections of each model would help avoid the MIA/TEX Meta-based predictions. The only issue with this idea is presentation, meaning I don't know how such a table would look in or be supported by Blogger. It could be an idea that I try on a dry-erase board at home, and if it succeeds in organizing my work, then the following year I incorporate it into the PPB blog.

Anyways, I'm looking forward to another year of PPB, and as always, thanks for reading my work! Below, you will find the aforementioned appendix. When the 2022 bracket was unveiled, I posted a mini-rant in the "To My Readers" regarding the site selections of the committee (and yes, I am still upset over it). This appendix is simply a re-post of the text for the purposes of documenting my biases/subjectivity for later reference.

APPENDIX - Initial Reactions to the Bracket Reveal

Thoroughly disappointed in the Committee's outright negligence with site distribution. #8 BOST in Portland OR, #5 CONN getting Buffalo NY over #4 ARK with VRMT as the #13, #10 DAVD closer to Greenville SC than #2 DUKE, #7 MURR getting an equidistant trip as #2 UK, #11 MICH safely in over IND and RUT plus playing in Indianapolis over #6 COST, #7 OHST 120 miles closer to Pittsburgh PA than #2 NOVA, #16 TXSO and #16 TXAMCC closer to Fort Worth, TX than #1 KU, #11 IAST going to Milwaukee WI vs #6 LSU, #10 MIA and #15 JAST stay in E.S.T. in Greenville SC where #7 USC travels three times zones eastward from LA, and worst of all if #5 HOU wins two games they play a regional in San Antonio TX closer than the 1-4 seeds.