PredictWise Blog

NFL Week 2

Bookmark and Share

I forgot to post this yeseterday, so I am one game short!

Election Update - 9/11, 54 Days

Bookmark and Share

I launched my state-by-state senatorial and gubernatorial predictions in the first week of August. You can find them on PredictWise updating as the data updates: senatorial and gubernatorial. Most important for the senatorial elections, you can find the expected balance of power on a separate table. I hope to add the district-by-district house election by October 1.

The method is the exact same as I described in this paper, forthcoming in the International Journal of Forecasting; with one big exception as there are no state-by-state betting markets, yet. Thus, the data is simplified to just fundamental data and polling data. The polling data, pulled from Huffington Post’s Pollster and RealClearPolitics, is transformed into an expected vote share and probability of victory using the model outlined in the aforementioned paper. I average the aggregated polling snapshots and then correct for both the anti-incumbency bias (incumbents poll too low compared to the results) and the regression to the mean (leads narrow as Election Day approaches) with a set of linear coefficients determined by past elections. The fundamental data (e.g., previous votes in the district, incumbency, and economic indicators) is transformed into an expected vote share and probability of victory using the model outlined in this paper with Patrick Hummel (forthcoming in Electoral Studies). We tested dozens of reasonable variables and used historical data to form the most efficient relationships between the fundamental data and expected outcomes.

The reason that I have not written yet on this is threefold. First, I find the fixation on the horserace months before the elections distracting. Second, I am working on a lot of other cool things! Third, the senatorial forecasts are not that much different than anyone else.

I could have gotten out dozens of small articles on the state of the senatorial election over the last few months, but, to be frank, when I contemplate politics, I prefer to focus on matters like our impending war with ISIS (or Syria) (i.e., foreign policy), the ongoing persistent, but sluggish recovery (e.g., economic policy), and major policy debates over gay marriage and gun control (e.g., social policy). We have elections to change (or preserve) policy, so it seems foolish that we focus so much time on elections and so little time on policy. Also, the fate of the senatorial elections are largely out of our hands, unless you are big investor of money (or time), but it is always meaningful to learn more about the major policy issues of the day! That brings me point two, which is a totally exciting project that I will announce in a few weeks that focuses more on sentiment and predictions around issues.

I was pretty stoked when I compiled the tables and saw that they looked pretty similar to everyone else (good sign that everything was running well), but, then again, that does not make it too meaningful to you; my senatorial forecasts just confirm what everyone else is already saying. First, binary control over the senate is incredibly close. The Republicans need 51 seats and the Democrats need 50 seats. The most likely outcome, at 21% is the Republicans getting 51 seats; they have a 60% likelihood of getting 51 or more seats. Second, I have 10 competitive races (i.e., currently between 10% and 90%): AK, AR, CO, GA, IA, KY, LA, MI, NC, and NH. The New York Times has 9 competitive races (they have Kentucky at 7% for Grimes (D) and I have her at 12%; they have New Hampshire at 9% for Brown (R) and I have him at 17%; they have Kansas at 24% for Orman (I) and I have him at 8%, but that may change once the real polling starts there and the courts rule on the fate of the Democratic candidate). FiveThirtyEight has all 11 reasonable races in the mix. Third, everyone has the same few key races: NC, AK, LA, and IA, with KS a wildcard.

Here is New York Times and FiveThirtyEight compared with PredictWise:

Tomorrow I will provide an overview of the gubernatorial elections. And, over the next few weeks I will jump into the key senatorial and gubernatorial elections in more depth, and focus on shifts in the elections as they occur.

 Updating Predictions: senatorial, senatorial balance of power, and gubernatorial.

NFL Week 1

Bookmark and Share

I am going to get this live in a few weeks with a lot more detailed game predictions. But, until I do, here are my week 1 probabilities of victory for the NFL. The predictions are primarily derived from betting markets. The home team is favored in 13 of 16 games.

World Cup Recap: Accurate and Calibrated

Bookmark and Share

The PredictWise forecasts was 15-0 in the knockouts games (0-1 in silly consolation games). It certainly was exciting to watch this unfold as the data (heavily driven by markets) endowed me with a rooting interest for all of these games (with, of course, the notable exception of the USA v. Belgium game where I was rooting hard against the data!). Below are all of the games, after the group stage, and morning forecast:

All of these forecasts moved in real-time during the games; this was both a meaningful challenge and an exciting viewing experience. It is meaningful in that it test the infrastructure to deliver low latency, quantifiable market intelligence. It is an exciting viewing experience as the user gets to see an objective measure of the value of a key of play and current situation. I look forward to dissecting the time-granular forecast data in much more detail the near future.

While the error was low, the calibration was off for the knockout games. We had an average likelihood of victory of 68% for the favored team in the 15 knockout games. Which means we expected the favorite to win 10 of the 15 games. With a 15-0 record we underestimated the likelihood of any of the favorites winning the game; we should have had every team at 100%.

That being said, everything evened out nicely over the 64 games. Our forecasts in the knockout rounds did not include the possibility of a draw, so the likelihood of a draw is coded in as 0%. The morning of the game, the favorite team was forecast to win 59% of the games and won 66%, the underdog 23% and won 20%, and a draw 18% which occurred just 14%. Obviously, the 15-0 in the knockout round upped the average for the favored team a little too high; all of the buckets from 50% and above happened a little more often than they should. But, overall, hard to complain with this sort of calibration over 64 matches with 3 possible outcomes:

For a full recap of the World Cup coverage, please visit

All 64 game are listed below:

World Cup, Final: Germany for the Win

Bookmark and Share

Germany is 60% to win today over Argentina 40%. Both teams have played well enough to make it to the final. This was not much of a surprise as they were two and three going into the tournament, with semi-final loser Brazil as first. And, as noted from the beginning, Brazil was first not because they had the best team, but the home-field advantage. Argentina was slightly more favored going into the tournament than Germany for two reasons. First, they had slightly stronger fundamental data. Second, they had an easier group stage (Germany had to contend with USA, Portugal and Ghana).

For the sake of the score keeping I am going to pretend that the consolation match never happened! Seriously, why are they playing that silly match?!? With home team Brazil getting humiliated in the consolation match to Netherlands that makes us 14-0 in meaningful knockout games and 0-1 in silly consolation games.

Updating Predictions: likelihood of any game and likelihood of any team reaching any round. Details on method and full coverage.