Two key criteria for developing good predictions are: (1) quantitative evaluation of the prediction and (2) accountability. With that in mind, I take a look at how my banzuke prediction performed.
Upper San’yaku was “chalk” as expected. In the lower San’yaku, I (and other predictions on this site) correctly had Yoshikaze filling the komusubi slot vacated by Shodai. I don’t understand the order of the three sekiwake ranks, as it appears unchanged despite the very different performances at Haru that had all of us predicting the order as Takayasu 1E, Kotoshogiku 1W and Tamawashi 2E.
In the maegashira ranks, of the 31 predictions, I had 11 “bulls-eyes” (correct rank and side” and 3 more correct rank predictions. This is way fewer than I expected or would have liked. The 17 misses were mostly not too bad: 13 missed by one rank, 3 missed by two ranks, and I had Osunaarashi (J1) moving up to M16 and Myogiryu (M15) dropping to J1.
There are three parts to the prediction: the computed ranks, tie-breaking among rikishi with identical ranks, and the departures I make from the computed ranks based on past banzuke patterns. Let’s look at these in turn.
The computed ranks were quite accurate: the official banzuke departs from these in only a couple of places. The computed rank would have Takarafuji at M3, but because of his make-koshi at that rank at Haru, the prediction and the banzuke moved him down to M4. Shodai (one of my two-rank misses) should be down at M7, and I still feel like the NSK cut him way too much slack after his 4-11 performance. And Arawashi and Ishiura would switch sides (but not ranks).
My tie-breaker was higher rank at Haru. This largely resulted in both of my other two-rank misses, as Takanoiwa should have been ranked above Tochiozan (and Aoiyama) by this rule. Presumably his 6-9 record at Haru led to his being dropped further down, although this is not necessarily consistent with past banzuke patterns. In a number of other cases, the tie-breaker got the relative order right, and I will need to look closely to see if the tie-break part of the prediction can be improved.
So, on to the departures from the computed rank order. One rule that resulted in many of my misses was to drop rikishi with 7-8 make-koshi records one spot from their rank at Haru, even if the computed rank would have them retaining their rank. This has often (but not always) been done in past banzuke. Although this rule correctly placed Takarafuji at M4, it placed Kagayaki, Tochinoshin, Ishiura and Daishomaru one slot too low, which also led to one-rank misses in the other direction for Ura, Arawashi, Kotoyuki and Onosho. It seems that the NSK is inconsistent in this scenario, and I’ll have to see if any pattern can be identified.
So overall, I am happy with my computed ranks, need to think more about the tie-break procedure, and need to be more careful with subjective departures from the computed ranks (this also includes demoting Myogiryu in favor of promoting Osunaarashi, even though Myogiryu had a better computed rank).
Others can chime in with how they fared. There will be another opportunity to predict the Nagoya banzuke after Natsu is the in books, and in the meantime we’ll have some actual sumo to watch!
4 thoughts on “Natsu Banzuke Prediction Post-mortem”
So the one thing that we can’t really get to the bottom of computationally, is that the actual banzuke decisions are made by people who take a number of factors into consideration. My favorite example is your citation of Shodai, who should have been pushed down quite a bit more. We all know there is a strong desire to groom Shodai for higher rank. I would expect that may be behind his “levitation”.
Using my computation method, Shodai had a rank velocity of -4.31 in the original run. That would have put him at M5 – but there were so many problems with that run that it was thrown out. Suffice to say it’s back to the drawing board for my approach as well.
My tally was 13 correct, 7 correct rank, wrong side. So pretty dire. On Maegashira, only 4 completely correct, 7 correct rank, wrong side. In other words, a blood bath.
Outliers like Shodai aside, I would’ve done better if I just stuck with straight computed ranks and not tried to adjust, though I think I have a slightly better idea of how to break ties next time. Most of my misses were one-rank switches of two rikishi, so the model isn’t wildly off…
Looking through my maegashira rankings and ignoring the east/west divide I had 13 direct hits, nine out by one, three 2-rank misses, one out by 3 (Tokushoryu) and one predictable absolute howler. Yes, I knew Ura wouldn’t really land at M5 but that’s what my formula threw up! I think the problem is that I give double credit to the 8-7, 7-8 divide compared to the other differences in wins.
Luckiest man on the banzuke is Shodai, but on the other hand he might have been better off rebuilding at M7 than M5.
Shodai is definitely the most over-ranked, and good point that NSK may not have done him any favors placing him that high!