It’s your chance to win some quatloos betting on September’s UAH. Remember: Even though ice has been melting, August’s UAH temperature anomaly was down from July. A number of entrants took the opportunity to steal Paul K2’s chance of decreeing nails in coffins. La Nina may or may not be deepening, the sun may or may not becoming more active. Leprechaun’s may or may not interfere with the monthly temperature. Detailed statistical analyses, reading values off Channel 5 or consulting Paul K2 or Dr. Jay Cadbury may or may not help.
Just wager your qualtoos and take a chance!
[sockulator(../musings/wp-content/uploads/2011/UAHBets5.php?Metric=UAH TTL?Units=C?cutOffMonth=9?cutOffDay=21?cutOffYear=2011?DateMetric=September, 2011?)sockulator]
Enter your bet before Midnight Sept 20.
Here’s my best, based on my new latest-and-greatest statistical model, using seven independent variables. I’m picky, so the greatest p value of the seven is .0047.
KAP– 7 independent variables? You are sure to win then!
Of course, there’s more than a possibility that some of the seven are cross-correlated, so they’re not really independent.
And yeah, I know seven is a big number. But solar and El Niño are near the bottom, and I’m not going to leave them out.
KAP are you using something like his adapted to UHI:
I see: MEI, volcanic, time, sunspots, ‘residual annual cycle’ (2nd order Fourier), plus the parameters to fit the residuals to an AR() process.
Clearly, a method with this level of gory detail should win all the qualtoos! Provided you remember to adapt it to UHA and not enter predictions for GISTemp.
But someone else might come along, throw in the AMO, NAO, and most recent census of Leprechaun population and bet you. You never know!
Well, *something* like that. But I’m not using linear, sinewave, or any other mathematical inputs. All my dependent variables are based on fairly standard climate data you can find on the web (tho some are fairly obscure), sometimes massaged, each with a lag of at least 1 month (so that predictions can be made). I won’t reveal the details until I’ve got a mass of those valuable quatloos in hand and my adoring fans demand to know how I did it. But FWIW I will say that my hindcast (model vs. actual UAH) has an r²=.778 and Akaike of -642.
Kap– Whoo hoo! Using R’s AIC for minimum extent, I only get Akaike’s of 788! (That said, if I divide the extent by 10^6 and do the fits, I get much smaller Akaike’s. Such is the beauty of Akaike! 🙂
Are the massaged Swedish? Or some other sort of massage?
Well, there’s a standard index that is posted rather late online, so I compute it myself from raw data. My computations match the posted ones to the first four decimal places; I know the source of the remaining error but I’m too lazy to go to the rather considerable effort to get the match any more exact. I also do my own version of SATIRE for pre-satellite TSI, but since UAH is all post-satellite era, it’s not needed here.
I actually have an alternate model with a lower Akaike and lower Hannan-Quinn and if I had the guts I would use it instead. But I’m not using it because (a) it had 1 additional parameter and I was concerned about overfitting, in spite of lower scores; and (b) it drove up the p values of several of my other independent variables that I have some confidence in as being “real” (more so than the added variable). The two models give quite similar predictions (at least for September) and if I could bet twice I would put 2.5 quatloos on each. Sadly, Lucia’s Blackboard is not the racetrack! (Imagine how many quatloos I could win with an exacta.)
You could look around and see if there are any true online gambling for UAH– ones that use actual $$s.
Uhmm…. you actually can bet twice. You just need to create a sock puppet with an email. Then you need to remember the sock puppets name. Oh, and the sock puppet needs to know how to add. Of course, doing so is dishonorable. But the script doesn’t know.
Well, if I created a sockpuppet, I would need a clever name so that nobody would know it was me. And to be the least dishonorable as possible, I would only bet 2.5 quatloos and then go back and change my original bet to 2.5 quatloos too.
I think there’s at least a make-able case that if a player keeps his total bets to 5 quatloos or less, multiple bets are legitimate. But I’m willing to trust the hive mind on this one (or the site owner) if anyone cares to chime in.
So if you really think it’s out of bounds, please delete my sockpuppet bet (assuming that you can penetrate my brilliantly disguised sockpuppet name).
KAP,
If you whisper your overfitted prediction to me, I’ll sockpuppet that bet for you. 5 quatloos… we can split the winnings!
P.S. I may be a spambot, but I know how to add.
KAP–
Your self imposed rules for sock-puppetry sound fine to me.
feabqtcy– Are you trying to end your losing streak?
Cheer up, feabqtcqy. The two-sigma error bar on these models is .21° C, so there’s a lot of room for a random guesser to get in there and steal those quatloos. In fact, given the number of bettors in the typical monthly pool, a (narrow) loss is the most expected outcome for either model.
KAP–
Spambots have a tendency to bet ‘0’. feabqtcy has advanced relative to other bots and can enter non-zero values. That makes him unique among bots.
My last modification (at least for this month) is now up. I tried to delete my sockpuppet bet by betting 0 on -275, but it insisted I bet 1. So I hope you’ll have the good sense to delete that cheating sockpuppet, who clearly should not be allowed to play.
The reason for this change was that I found a new significant predictor, which I added to the model. This improved some of the odd effects I got while including the 8th independent variable, described above. So I’m including them all, 9 independent variables all significant to 99% or better. (The worst is still El Niño, and I’m still not gonna leave it out.) And yes, I tried cosmic rays! And no, they didn’t help!
For the statistically minded, r²=.786, Akaike=-655.
In spite of all this, however, the standard error only went down by a couple of thousandths of a degree. So the most likely result is still a narrow loss, given the number of players in a typical month.
Last change I promise. Found a different version of variable 9, with a slightly lower Akaike (-657, r²=.787). Prediction now virtually identical with the 8-variable version.
…. aaaaand …
Researching variable 9 pointed me to variable 10, which also improves the model. Eight of the ten are 99% significance or better, with the other two very close. (El Niño is still at the bottom). Ok, I promise, no more modelling until next month (just like I promised last time). Current r²=.791, Akaike=-661.