Tuesday, December 15, 2015

2016 Election - Odds-maker-based Republican Model

It’s dark every night when I leave work. I know what that means: it’s December and the college football season is almost over. We still have some football left, notably the conference championship games, the bowls, and the college football playoff, but the end is in sight.

Last year, that I meant I stopped posting for 9 months. But this year, this year it means I’m shifting into political mode. After all – the 2016 presidential election is a mere 328 days from now. I’ll be building out a polling model just like I did in 2012 (here’s hoping I can achieve the same result) but this far from Election Day the polls don’t have much predictive value. This far from the election polls do have some predictive power, but other elements like endorsements and experience matter as well.

That being said, I have built a model, but it doesn’t rely on polls. Instead, it relies on odds-maker consensus. Since July 20th, I’ve been capturing betting data from different odds-makers on each of the declared candidates regarding a) winning their party’s nomination and b) become president.

I can’t think of a better day than today, the day the republican candidates are debating in Las Vegas, to make my first 2016 election post.

The process for developing these percentages has been:
  1. Capture the odds at which one could bet on a candidate (to win the nomination or the presidency) from 26 different odds-makers
  2. Average those odds to get a consensus
  3. Translate those odds into an implied likelihood of winning
  4. Normalize the odds across all candidates such that they add up to 100% (removing the house edge)
Repeating that process every few days has let me build the following graphs, tracking each candidate’s fortunes over time. I’m experimenting with two different ways to present the data. The difference should be very clear. In this top graph I put everyone together. It has the advantage of letting you directly compare one candidate’s fortune to another’s, but there is a glut of candidates at the bottom who are obscured.

My second graph, shown below, plots every candidate on their own little graph. Each graph is identical in both axis values and scale. This data visualization technique is called small multiples, and visualizing many different trends is one of its best uses.

This technique has the advantage of giving each candidate their own little space (very socialist) and making each line much easier to parse. Conversely, it's harder to understand what's happening between candidates, and each candidate's line is shrunk a little horizontally. For example TRUMP looks over the last few weeks looks like just a squiggle down below, but in the graph above you can exactly what's been going on with TRUMP.

Both graphs (obviously) tell basically the same story, but do so with emphasis on different things. I think I slightly favor the second presentation, but I'd be interested to hear what people think.

This is just the first post of many on the 2016 election. In the coming weeks and months I expect I'll make or publish:

  • Democratic primary model
  • Primary Election night county based forecast
  • Primary polling based model
  • All sorts of general election stuff
Happy presidential election cycle!

P.S. I know the fonts and colors aren't the same between the two graphs. It's bugging me too. But I'm jumping back and forth between two pieces of software as I sort this out, and I decided it was more important to get started than get every color/font/detail spot on. 

Saturday, December 5, 2015

No updates this week

I've been in LA all week attending the Fellowship Admissions Course. For my non-actuary readers it's the highest level of accreditation you can recieve as an actuary. I kept thinking I would have time to do some math, but it's Saturday morning and time for me to admit I won't.

I'll be watching todays games though! And I'll be back for the bowl games,

Friday, November 27, 2015

College Football Math: Week 13

Rivalry week :). Holidays have me slow moving and I'm writing this after my huskies have already shellacked the cougars, so no matter what happens tomorrow it'll be a good week for me. I've decided to stop the if they win / if they lose graphs for this week; we're in lose and you're out territory for everyone except Clemson and mayyybe Bama.

We've already played this week's #1 game (and it did not disappoint) but there's plenty of excitement on the slate for tomorrow.

On a sappier note, one of the many things I'm grateful is that I have the ability and the tools to do this model and write about it, and that I have readers who find it interesting. I'm not wrapping up the season or anything, I'll be making some bowl posts as well as some retrospective looks at the year (and then it'll be time for politics!), but I thought I'd pause on this week of thanksgiving and express gratitude for what has been another awesome year of college football.

Friday, November 20, 2015

Week 12 College Football Math

I apologize for the late post. I've been running around like a crazy man this week.

It's an elimination week in the Big-12. Baylor, TCU, and Oklahoma all need to win to keep their CFP hopes alive. Oklahoma State, being undefeated, would still have some CFP hopes if they lost, but it would be much much better if they won.

I wrote and published a long post on some of the model's inner workings, go check it out.

Tuesday, November 17, 2015

Dashboards updated through week 11 & CFP

Through week 9

Last week, the CFP Committee Model predicted the rankings would be:
  1. Clemson
  2. Alabama
  3. Ohio State
  4. Baylor
  5. Notre Dame
When the rankings came out they were:
  1. Clemson
  2. Alabama
  3. Ohio State
  4. Notre Dame
  5. Iowa (?)
  6. Baylor
I post that not to brag, but as a lead-in to this week's rankings. The model swapped Clemson and Alabama. The AP poll or the coaches poll would never rank a 1 loss team #1 with so many undefeateds remaining. But the CFP committee? They might; they're a creative bunch. I'm interested to see how they handle Oklahoma. On one hand, they crushed Baylor this week but on the other hand they were ranked 12 last week, and it's just one game. I'm not sure how much appetite the committee will have for moving them up.

Saturday, November 14, 2015

Week 11 CFP Math

Are you ready for football?

I wrote a long post (with lots of pictures!) that gets into the single most powerful ideas behind my model, go check it out.

I'm continuing to experiment with new ways to analyze college football data, especially as it relates to the CFP. I have something new to try this week, and there's more on that below. First though, let's look at the graphic of dessssssssssssssstiny.

  • Controls Destiny: All the undefeated power 5 teams control their own destiny, no surprise
  • Almost Controls: Filled with teams who: have one loss, are blue chip programs (also, there's Iowa). They are almost certain to get in if they win out. They are behind 5 undefeated teams right now, but teams lose, shake-ups happen, and if one of these teams wins out they're in
  • Needs Some Help: Teams who have 1 loss, but started the season with lower expectations, and have smaller names. 
  • See you next year: Intuitively, I disagree with the model here. I'm not sure why it's hating on Michigan State and UNC. Either of these teams winning out (and beating Ohio State or Clemson along the way) would have a strong CFP resume. Houston looks about right though - they need to win out and hope for chaos up top.

Week 11 Games - High CFP Leverage

I was curious which games had serious CFP implications, so I sat down to try and figure it out with math. It turned out that measuring "CFP Implications" was trickier than I expected. I started by adding the simulation output that finds, for each team, what their CFP odds would be if they won the game and what their CFP odds would be if they lost the game. For example: in the graphic below OU's CFP odds will be 33% on average if they beat Baylor this weekend, and 1% if they lose.

However, those statistics is not sufficient to measure CFP leverage. For example: Notre Dame is in a strong position to make the CFP. If they lose to Wake Forest this weekend that will change; their CFP odds will crash to near 0. So why isn't Wake Forest @ Notre Dame included in the list below of high leverage CFP games? Because there's almost no chance Wake beats Notre Dame; the game will be an uninteresting blowout. Were I to include only the "if they win" and "if they lose" numbers, I'd be stuck with games like Wake @ Notre Dame at the top of my list. I needed to do better. 

I added a measure for how close the game is likely to be, and got the list below. I'm happy with that result for the time being. Especially since it pulled out OU @ Baylor. The final formula includes: CFP odds if they win, CFP odds if they lose, and likelihood of losing (closer games are better), and it gave us the following list of games.

Duh. I could have picked this game as the top CFP game without any math at all. Oklahoma and Baylor both have a good shot to make the CFP, the loser will see their CFP odds plummet, and the game is expected to be close. 

Alabama looked like the best team in the country last week, they stopped the Leonard Fournette freight train in its tracks. No rest for weary though, as they head to Mississippi State, and they'd better be up for it, because a single loss could doom them.

For years this has been the flagship game in the Pac-12, and it is again this year. In past years Stanford has spoiled the season for Oregon several times; I'm sure Oregon is looking to return the favor and knock Stanford out of the CFP.


If OK State loses to Iowa State they will be out; as it should be for anyone who loses to Iowa State


LSU had a rough week last week, but they're still in the hunt. They only have one loss and could sneak back into the SEC championship conversation with an Alabama loss. But this should be a good game, the hogs are always tough.

Happy football!

Wednesday, November 11, 2015

Week 11 College Football Math

The CFP  rankings have been released, we've already had some MACtion, let's do some math.

  • I wrote a long mathy post that gets into the single most powerful ideas behind my model, go check it out
  • Dashboards are updated
  • There's another week 11 post coming that focuses on CFP math

Week 11 games

Oh man. It begins and ends with Baylor/OK, doesn't it. Each team controls its destiny in the Big-12, and very likely in the whole CFP, and this is Baylor's first real game of 2015. In other spots we see both UW and WSU football games on the Top Ten list; it feels good.

Week 10 Results

I added two columns to this week's outcome chart. Their focus is the new published about idea: Posterior Win Probability

The first column shows how that game's outcome translated into PWP. For example, Louisiana Tech rolled over North Texas. Their 56-13 margin of victory translated to a 99% PWP. Since Louisiana Tech's prior win probability (aka model confidence) was just 97%, we say the model was wrong by 2%.

The model was much, much more wrong on New Mexico State @ Texas State. Texas State was supposed to crush NM State, but instead they went down 31-21, a 10-point loss that translates to a PWP of 23%. It's fun to see when the model is super wrong, but there's another reason I like the PWP addition here.

In past weeks, we've looked at whether the model called the right winner. But predictions that accurately get the winner aren't all equal. You can be right, barely right, or even really right. For example, the model correctly picked both Florida and Temple, but while the Temple prediction was awesome, the Florida game turned out much closer than the model expected.

So not only can we now see right pick vs. wrong pick, we can see how right and how wrong.