Over the last couple days I've made several improvements to my prediction system, which has fully-automated my predictions and also allows me to roll it back to any arbitrary vote in the past, and see what the prediction would've been from that point in time. I used this to plot out the predictions of both Model A and Model B over the entire time that Ack's been holding votes for his stories, going back to mid-December last year. I'd hoped to see which model was more accurate so I could pick between one, but...
...turns out they're not very different.
I tracked two averages: the number of votes a model predicted correctly, and the number of votes it predicted right in a row. For example, predicting from Vote 73, model B correctly predicted the next two stories, got the next two wrong, then got the next 5 right. Going by the total number of correct predictions it got 7 right, but by the correct predictions in a row it only got 2.
So, the totals are as such:
Average correct predictions:
- Model A: 5.213
- Model B: 5.096
Average correct predictions in a row:
- Model A: 3.160
- Model B: 3.255
Based on these results... I think I'm going to continue using Model A going forward. The slight edge Model B has at predicting correctly in a row is less than the edge Model A has for total correct predictions. Additionally, since I started tracking the margin of victory for each prediction, I can guess a little more accurately when something might upset the calculated winner.
I'm still going to keep Model B on the books, and every so often revisit these averages, just in case it cleans up its act. But I'll only report on the results of Model A here in the thread.
With that in mind, here are this round's predictions:
Vote 84 is probably going to Security! unless someone mounts a determined campaign to stop it. It's been projected to win by a margin of around 30 votes for a while now, and it hasn't shifted much as we get close. Enough people taking it from 6 to 0 could change this, but personally I don't expect that to happen; it's been a pretty solid vote-getter for as long as I've been tracking them.
The next several votes are all by pretty hefty margins. I, Panacea in Vote 85 by a margin of 60+, MirrorVerse in vote 86 by a margin in the 70s, Recoil in Vote 87 by more than a hundred and Wyvern in Vote 88 by almost as much.
Then we get to Vote 89, where Really Bad End is estimated to pull ahead by the skin of its teeth -- winning by 1 vote. NSW is right behind it, though, and if it can't manage an upset will win Vote 90 by a margin of more than two hundred. Everyone's favorite Thinker team is back in Vote 91: Alea Iacta Est is expected to win by 110+ votes.
Finally, we have another close race and a potential upset. All Alone is currently projected to win Vote 92 by a mere 4 votes, while recently-updated Trump Card is right behind it. As it stands Trump Card has to work its way back up from a vote reset, while All Alone has nearly 668 votes currently... but Trump Card has been earning twice as many votes on average, so it'll close the gap rapidly. We'll see if All Alone can stay ahead as we get closer.
Past that point, estimated rounds-to-win are as follows:
Alea Iacta Est: 9
All Alone: 24
Bait & Switch: 27
Confrontation II: 4
Danny & Taylor: 3
Hostage Situation: 14
I, Panacea: 5
Junior Hero: 9
Meet the Heberts: 2
MirrorVerse: 18
NSW: 19
Nemesis: 16
One More Trigger: 11
Really Bad End: 44
Recoil: 5
Security!: 8
Slippery Slope: 9
Trump Card: 11
War Games: 6
Wyvern: 6
Again, that's from Vote 93, not the current vote.
Accuracy this round was 97.25% overall, with an average difference of 9 votes between projected and actual results. Most accurate was our newest winner, War Games, which came in exactly as predicted at 100% and 0 votes off. Least accurate proportionally was previous winner Trump Card, at only 83.78%. Least accurate in raw count was a tie between Trump Card and Recoil; both stories came in 18 votes shorter than expected.
I think for my next upgrade I'm going to try and account for the post-victory drop-off in votes...
My vote:
[X] Meet the Heberts - 6
[X] NSW - 5
[X] War Games - 4
[X] One More Trigger - 3
[X] Security! - 2
[X] Recoil - 1
