Notes from the Sports Nerds

Site Update: Enhanced 2013 MLB Pick Logic Released

2013-04-17T21:37:22Z

Researching and improving our prediction models is an ongoing process, and one to which we devote significant time. The past year has been particularly active, though, as we’ve been rolling out significant pick logic updates to all of our sports. These updates are primarily designed to eliminate a long-standing source of user confusion, conflicting picks, but they should also help to improve our long term pick accuracy.

MLB is the latest (and final) sport to join the club. So, what’s new?

Creation Of A Master Game Prediction

For a long time, we’ve had completely separate models for game winner picks, run line (or point spread) picks, money line picks, over/under picks, and final score predictions. Why? Because it’s typically more effective to optimize a model to achieve one goal. If your primary goal is to pick point spread winners, that’s all you care about. You should use a model that is the best at predicting against the point spread, even if it isn’t one of the best models for predicting straight up game winners.

If we set out to design one model to predict both game winners and point spread winners, from the get-go, we’d have trade-offs to make. Some calibrations of the model might be pretty good at predicting both game winners and point spread winners, other approaches could end up being great at predicting one of those two things, but not so great at predicting the other. As a result, we stuck with the multiple, focused models approach.

One down side, though, is that independent models don’t always agree. In the past, this inevitable reality has led to some conflicts. For example, in MLB, our game winner model might predict one team has an edge, while the TR money line value pick is finding value on the other side, in a way that doesn’t make sense if you consider the game winner prediction odds. In basketball, our game winner model logic may favor a 1-point underdog to win the game outright, while our point spread model is siding with the 1-point favorite to cover.

Those days are over. In all of our sports, including MLB, we now combine information from all of our model types to create one master prediction for each game, which is then used to create picks against each type of line. Here’s a flow chart for those of you who are visual thinkers (using MLB as the example):

One of the main benefits of this change is that there will no longer be any conflicting picks across pick types.

Incorporation of New Predictive Factors & Weights

Besides eliminating pick conflicts, this new master logic should also help improve pick accuracy. We’ve done more analysis of our historical pick data to find which models perform best under certain conditions, and given them more weight when combining the models. In some cases, that means relying 100% on a single model that has performed well, while in other cases we mix together information from several models.

The general point here is that a “model of models” approach can often be very effective. Even if no individual base model (i.e. Decision Tree, Similar Games, Power Ratings) is outstanding at making, say, MLB over under picks, looking at the set of their predictions as a whole and potentially incorporating other contextual variables about a game can produce profitable long term performance for the TR picks.

No Change To Picks Pages

This is purely a behind-the-scenes change to our prediction model logic. The visual look of the picks pages won’t change.

The MLB game winner picks, MLB money line picks, MLB run line picks, and MLB over-under picks pages will still show our official TR Pick, along with an Odds or Value column showing how likely a team is to win or cover, or how much of an edge they have against a line. And all of those picks pages will still show the base predictions of the individual models for that pick type (Decision Tree, Similar Games, and/or Power Ratings).

The MLB predictions grid page will also remain unchanged.

Logic Updates Are Already In Action

This blog post is actually coming post-facto, as we rolled out this new logic during the evening on Thursday, April 11th. So if you’ve been following our MLB picks over the last few days, you’ve already been seeing the new logic changes in action.

It’s still way too early to make any conclusions — we’ll see where we end up at the end of the 2013 season — but the results have been good so far. From when the new logic was released until this morning (so, games on 4/12/2013 through 4/16/2013), our playable MLB picks of all pick types have gone 56-47 for a gain of +7.2 units.

In comparison, our best individual raw model (Decision Tree) was 55-68 over that span, for a loss of approximately -20 units (we didn’t run an exact unit count there, as that win-loss record kind of speaks for itself).

This is a very small sample size, and you shouldn’t assume the picks will perform the same over the next five day stretch, or the one after that. But it’s at least good to confirm that there doesn’t seem to be any grievous errors in the new logic that wiped out a bunch of good results from our previous approach.

If you’ve got any questions about the change, please fire away in the comment section below. Also, as a final note, now that it’s been two weeks since opening day and we’ve rolled out our prediction model enhancements for the 2013 MLB season, our free MLB trial period for new TR users will be ending tomorrow (Thursday 4/18).

The post Site Update: Enhanced 2013 MLB Pick Logic Released appeared first on Notes from the Sports Nerds.

NCAA Tournament 2013 Recap: How Our Brackets & Betting Picks Did

2013-04-11T20:48:39Z

See more at TeamRankings.com

Now that the 2013 NCAA tournament has concluded in thrilling fashion, let’s review how our algorithmic NCAA tournament betting picks and computer-optimized brackets did this year.

In short, our betting picks were profitable and once again, our brackets did very well. Our bracket for small pools finished among the top 5-7% of the nation, our upset bonus brackets did phenomenally well in early round games, and many BracketBrains subscribers won prizes in their pools, including customer reports of first place finishes in contests as big as 500 people.

Finally, one of our larger pool brackets finished in the 99.99th percentile on ESPN.

Tournament Betting Picks Results

Once again, our 2013 NCAA tournament betting picks ended up nicely profitable, and our against the spread picks did particularly well.

For the 2013 NCAA tournament, our playable (2- and 3-star) spread picks went 13-9-1 (59%) and all ATS picks went 39-27-1 (59%)
Combined, playable spread and totals picks went 31-25-2 (55%) for +3.2 units of profit assuming -110 juice
If you bet all of our spread and totals picks, including the 1-stars, the result would have been 71-61-2 (54%) for +3.5 units of profit for spread and totals picks, at -110 juice
Finally, our money line value picks returned +1.2 units of profit overall

Considering the $50 upgrade fee to add betting picks to our BracketBrains Pro package, if you were wagering at least $20 a play you would have made money betting our tournament picks, even if you only bet the playable picks.

If you were betting $100 a play on all of our playable spreads and totals, plus the money line value picks, you would have earned a 780% return on your $50 upgrade fee in a little over month’s time. Even if you only bought BracketBrains for the betting picks and couldn’t care less about brackets, at a $99 fee the return would have been over 340%. That beats the heck out of the stock market.

Our Official 2013 Brackets

For the newbies, here at TeamRankings we’ve built sophisticated bracket contest simulation technology that enables us to model bracket contests based on their size (in terms of number of brackets entered), scoring system, and other variables.These variables have a profound impact on optimal bracket strategy, and as a result, our technology allows us to pinpoint which bracket picks offer the best chance to win a specific bracket contest.

The optimal bracket can vary quite a bit by situation; the best picking strategy to win a 10-person pool, for example, looks a lot different than the best picking strategy to win a 1,000 person pool.

This year we published 140 official brackets on Wednesday, March 20, the day before the start of the Round of 64. (Despite the emergence of the “First Four” round of games on Monday and Tuesday of that week, most bracket contests ignore those games.) That breaks down as follows:

5 different brackets (a “Best Bracket” and four alternates) for…
7 different pool size ranges and…
4 different base scoring systems

5 * 7 * 4 = 140 unique brackets. In terms of scoring systems, we covered the traditional 1-2-4-8-16-32 system as well as 1-2-3-4-5-6, and also did brackets for upset bonus scoring. Our upset bonus logic was based on bonus points being awarded based on the difference in seed between a bracket winner and its opponent; that framework seems to be the most popular for upset pools, although many different ones exist.

All of these brackets were available to customers of our BracketBrains premium service, but since the tournament is over, we’ve opened up access to all of them. You can use our new Bracket Finder tool to check all of them out.

Overview Of How Our Brackets Did

Tracking how we did in a given year has become substantially more complicated now that we are publishing brackets for multiple types of pool sizes and scoring systems, especially when the big contests like ESPN Tournament Challenge only report on national standings for the traditional (1-2-4-8-16-32 points per round, no upset bonus points) scoring system.

As a result, we can only judge our performance on a national scale for the traditional scoring system brackets. For the rest of our brackets, we do the best we can to estimate, plus, our customers tell us how they did. So we’ll start from the traditional brackets and go from there.

In short, our brackets again did very well. we had a several “hits” this year, and came one or two Florida wins away from some fantastic results:

One of our Best Brackets for traditional scoring, Jordan, finished in the top 5-7% of national contests, good enough for an expected victory in pools up to 15-20 people (exactly what it was designed for)
Our Best Brackets for midsized pools with traditional scoring had a great shot this year, but Florida’s loss to Michigan (the eventual 2nd place finisher) derailed them
Our Best Brackets for small and midsize pools with upset bonus points appear to have done extremely well, with multiple customers reporting winning pools of 100+ people

How The Tournament Played Out

In terms of putting the success of our bracket strategies in 2013 into context, we need to first look at how the 2013 NCAA tournament unfolded. There are really just two key points to note:

The most popular NCAA champion pick in brackets nationwide, Louisville, ended up winning the tournament. Over 20% of the public picked Louisville to win it all.
None of the next nine most popular national champion picks (Indiana, Miami, Kansas, Duke, Ohio St., Gonzaga, Georgetown, Florida, Michigan St.) even made the Final Four

Given that the traditional 1-2-4-8-16-32 scoring system is so heavily biased toward getting the last few games of the tournament correct, the above results had major implications:

Even in small pools, if you didn’t have Louisville winning it all, you likely had no chance at first place
In larger pools, even if you did have Louisville winning it all, you probably also needed to pick Michigan in the final game in order to have a shot at winning, since about 1 in every 50 bracket pickers had a Louisville vs. Michigan final
If you put literally zero thought into your bracket picks this year and picked all “chalk” (seed favorites) to advance in every round — that is, you had all four #1 seeds in the Final Four, with Louisville, the tourney’s overall #1 seed, as your champion — you would have outperformed over 90% of brackets nationwide; that’s still not good enough to expect to win a 20+ person pool, but certainly better than usual for this “blind” strategy

If that last point seems crazy to you, well, it just shows you how important picking the NCAA champion correctly is in 1-2-4-8-16-32 scoring. But there are a couple other underlying reasons why picking “chalk” did very well this year.

First, although Louisville was the public favorite, they were still a less popular champion pick than, for example, Kentucky was last year. So getting Louisville right came with greater rewards. Secondly, very few people got more than one Final Four pick right this year, which added even more weight to the Louisville pick.

Results Of Our 1-2-4-8-16-32 Best Brackets

We’ve already written a ton for our BracketBrains subscribers about our overall bracket strategy and its application to 2013. But to summarize, we did think Louisville was the most likely team to win the NCAA tournament this year, unlike several other noted prediction systems out there such as Ken Pomeroy (whose ratings favored Florida to win it all), LRMC (Florida), John Ezekowitz’s Survival model (Florida), Accuscore (Kansas) and Prediction Machine (Indiana). The AP poll, for whatever it’s worth, had Gonzaga as #1 heading into the tourney.

However, because Louisville was a very popular pick and their odds to win were being overrated by the public, bracket contest simulations showed that the Cardinals were not the best pick to win in many pool situations. It’s easy to say that Louisville was an obvious championship pick in hindsight, but that’s hogwash. Anyone who looked at the data on this year’s field objectively on March 20 recognized that it was a relatively wide open tournament. Let’s not forget so easily that the team that won it all was losing by 10+ points in the second half to a #9 seed, and the team that finished in second place needed a miracle last-minute comeback just to make the Elite Eight.

We knew full well going into this year’s tourney that if Louisville did end up winning it all, most of our brackets for larger pool sizes would not do particularly well. That was fine, though, because especially in bigger pools, picking Louisville to win didn’t even come close to assuring you a top finish, given their popularity. The goal with bracket pools is to maximize your odds to win the pool, not to maximize your odds to get your champion pick correct. Especially in larger pools, those two goals can quickly diverge. If it’s not immediately apparent to you why, you need to read our posts about bracket strategy, but some more explanation is below.

In that context, this year’s Best Bracket results were unsurprising:

Jordan, our most conservative Best Bracket, designed for less than 15-person pools, had Louisville as champion and finished in the top 5% of all Yahoo! brackets and the top 7% of all ESPN brackets. If you used this bracket for its intended purpose, you had a great shot of winning a 10-20 person pool, and customer emails we’ve received have confirmed multiple top finishes in this size range.
The rest of our Best Brackets for traditional scoring all landed in the top 30-50% nationally, but did not perform well enough to contend to win bigger than 15-person pools.

While it would have been nice to have at least one Best Bracket “hit” in the midsize pool range to complement the performance of Jordan, picking Louisville to win simply didn’t give us the best odds to win a midsize pool.

For proof of that, you can check out two of our Alternate Brackets, Perkins and Ellis. Each of these brackets got its champion pick right and landed in the top 10% of ESPN. That’s good, but still not quite good enough to win a 30-70 person pool.

The odds were better to pick an undervalued team to make a run. Going in, we saw Florida, Michigan, and Syracuse all as undervalued potential sleepers. The proof isn’t hard to find; for example, we had a Best Bracket with Michigan in the championship game, and at least one Alternate Bracket with Syracuse in the Final Four.

However, while we liked both Michigan and Syracuse, Florida still looked like the best combination of risk and reward, and our simulations bore that out. So our biggest bets in midsize pools were on the Gators. In short, we came close to a great year. If Florida had beaten Michigan, our Best Brackets for traditional scoring could have done extremely well. In fact, all of our “big bets” on value picks in larger pool sizes look much better in retrospect:

Pitt lost in the first round, but to a team that made the Final Four and came within four points of the eventual tournament champion
Ditto for Gonzaga; it turned out to be just bad luck that two of our big value plays happened to play a red hot Wichita State in the early rounds
Florida made it one round further than most people predicted, then lost to a team that came within 6 points of winning the national championship

If Florida and Michigan hadn’t been in the same region, it would have been really interesting to see if our pick optimization algorithms would have produced a lot more brackets with Michigan in the Final Four.

Results Of Our Upset Bonus Brackets

It looks like the real superstars of this year, though, were our brackets for up to 150-person pool sizes and upset bonus scoring like Maravich, Bird, and Abdul-Jabbar. We’ve had multiple reports from customers who won 100+ person pools with these brackets.

In general, these brackets combined a conservative Louisville-as-champion pick with very aggressive upset picking in the early rounds, including calls such as #12 Oregon winning two games, #13 La Salle winning two games, #14 Harvard beating #3 New Mexico, and both #12 Ole Miss and #12 California winning their opening games.

Results Of Our 1-2-3-4-5-6 Scoring Brackets

We don’t have a great sense of how our brackets for 1-2-3-4-5-6 scoring systems did outside of customer feedback, but given that the scoring system dictates a more conservative approach, they should have done pretty well for most people. Although getting the champion pick means much less in this system, our 1-2-3-4-5-6 Best Brackets with no upset bonus had Louisville as champion up through the 75-person pool size, and all pool sizes for 1-2-3-4-5-6 upset bonus scoring had Louisville as champion.

However, if you happened to play Riker, you are most likely now a living legend in your pool. That bracket, a 1-2-3-4-5-6 Alternate Bracket, would have ranked around #1,000 on ESPN this year if used in a 1-2-4-8-16-32 scoring system, out of over 8 million brackets entered! More importantly, we’ve already gotten confirmation of a customer finishing in first place in a 500+ person 1-2-3-4-5-6 pool using it.

Conclusion

With another exciting tournament now over, we’re looking forward to the future. We plan to continue to refine and improve our NCAA bracket prediction algorithms, and also to apply our latest techniques to other popular contests like NFL survivor pools and weekly football office pools. We should have some great new stuff to announce along those lines before the fall.

As for our bracket performance in 2013, overall, we’re happy with it. Putting a major new technology into practice, we had several big hits (the Jordan bracket for small pools, and our small and midsize upset brackets), and came within one break — Florida beating Michigan in the Elite Eight — of a potentially epic year.

The post NCAA Tournament 2013 Recap: How Our Brackets & Betting Picks Did appeared first on Notes from the Sports Nerds.

Jordan Sperber Wins $2,000 Grand Prize In Stat Geek Idol 2

2013-04-09T02:32:42Z

See more at TeamRankings.com

First, we asked the amateur stat geeks of the world to submit us up to four pages of insightful, unique analysis about college hoops. (Here’s the invite.) We staked a $2,000 cash prize for the entry that exhibited the best combination of a compelling topic, rigorous analysis, and refined, persuasive presentation. The entries we received came from people of all backgrounds, from students to professors to professionals.

Then, we whittled the number of submissions down to five finalists, which we sent to an esteemed judging panel stocked with consumers and practitioners of basketball analytics. In the end, eight judges read every finalist’s entry, and weighed in with their rankings and feedback: Mark Cuban, Dean Oliver, Ken Pomeroy, John Gasaway, Ben Alamar, Toby Moskowitz, John Stasko, and Jeff Haley. For judge bios, see the original post.

Now, the judges have spoken, and we are happy to crown our champion of Stat Geek Idol 2: Jordan Sperber.

About The Champion: Jordan Sperber (@hoopvision68)

First, read Jordan’s winning entry here: Do Matchups Matter In College Basketball?

Jordan’s submission takes a straightforward yet well reasoned look at important questions for which unsubstantiated assumptions are the norm in today’s media narratives. When a great offensive rebounding team plays a very poor defensive rebounding team, for example, what should one expect to happen? If a team is great at offensive rebounding, is it better to play a team that’s terrible at defensive rebounding, or one that’s average?

We’ll leave it to you to read Jordan’s winning entry, and comments are welcome. We’re just excited to see where Jordan’s passion for college basketball analysis takes him in the years ahead. He’s currently a sophomore at Villanova. Besides winning Stat Geek Idol 2, he’s also landed himself a sought-after internship this summer with ESPN’s Stats and Information group.

Jordan’s no flash in the pan, either. As a freshman last year, he also made the finals of Stat Geek Idol 1, and his in-depth submission “A Video Charter’s Guide To The Final Four” was in contention for the $1,000 grand prize.

Complete List Of Finalists

This year’s group of Stat Geek Idol finalists showed a lot of promise, from exploring novel research questions to creating interactive data visualizations. They are listed below in order of their placement in the final judging results.

We’re also calling out the second place finisher specifically, as his entry received several high marks from the judging panel, as well as the second highest average rank by a decent margin. (As it turns out, Greg also made the final round of Stat Geek Idol 1 last year, but the rest of the group were first time finalists.)

First Place:
Jordan Sperber (@hoopvision68) for Do Matchups Matter In College Basketball?

Second Place:
Greg Matthews (@StatsInTheWild) for Expected Points And Efficiency

Finalists:
Stephen Shea for Measuring Offensive Balance in College Basketball
Ryan Silvis for Finding Northwestern A New Coach
Joshua Riddell (@Joshua_Riddell) for Charting Screens In College Basketball

Judging Breakdown: So Much For Consensus!

What was most interesting about the judging was the wide range of rankings for almost all entries. The table below presents the distribution of the judges’ rankings of all five finalists, with each judge made anonymous as a letter A through H (just click on the table to enlarge it):

Some specific observations of note:

Jordan’s winning entry was clearly a favorite, with five of eight judges ranking it as best, yet Judge A saw Jordan’s winning entry as his least favorite
However, it should be noticed that based on correlation to the judging group as a whole, Judge A’s rankings were pretty extreme outliers
Greg’s second place finish could be described similarly, as five judges thought it was second-best, while two judges rated it as their least favorite
Ryan’s third place finish was the most polar: the favorite of two judges, the least favorite of three judges. Go figure!
Four of the five finalists received at least one first-place rank
All five finalists received at least one fifth-place rank

Overall, we like it when a bunch of smart people (the judges) don’t agree. We do feel like Jordan is a worthy champion, and that the judging process ended up working well. However, sports analysis is a complex world and there’s rarely ever a cut and dry “right answer” or universally compelling topic. As the judges’ feedback shows, when it comes to judging Stat Geek Idol, opinions and assessments vary widely. We’re perfectly OK with that.

So What’s Next?

In closing, we wanted to lay out some thoughts about where we think the Stat Geek Idol competition could go from here. We’ve now run two iterations of this contest, and really enjoy the fact that people are generating interesting and novel college hoops analysis for it.

We’ve been lucky enough to turn our passion for sports data and prediction into a successful business, and we view Stat Geek Idol as our way of giving back to people who are in a similar position as we were once. Way back in 1999 you could find our original founder, Mike Greenfield, sequestered away in his dorm room on a Saturday night, writing Perl scripts to scrape basketball stats off of Yahoo! and designing his own power ratings system.

It seems to us like the next steps in improving the contest could involve the following:

More time? Set deadlines months, if not a year, in advance, and give contestants a ton of time to prepare their submissions. If needed, they would have time to explore a few different topics in relative depth, and settle on the most promising one for their SGI submission. This would also give us more time to spread the word about the contest.
More data? Work to make a robust historical data set for college basketball available to all contestants. Complete, reliable historical college basketball data isn’t the easiest to gather, and not providing data likely biases the contest toward contestants with access to it, either because they have connections or because they happen to have decent computer programming skills. (Although if you want to go far in sports data analytics, you better get some computer programming skills!)
More guidance and iteration? A lot of SGI entries we read elicit a somewhat common response internally: “Wow, this seems like a really good idea, but it could use some more [polish / focus / historical testing / some other tweak].” Instead of just having contestants submit an entry once, we could have an initial weed-out stage, then provide feedback and mentorship on how to improve a finalist entry for its final submission…and, of course, more time for a finalist to iterate on his/her entry.
More distributed prize structure? We’re not entirely convinced that a winner-take-all prize structure is the most effective approach. Maybe some contestants who are good, and would have a reasonable shot to win it, could get discouraged thinking that their odds to take home the grand prize are too low. Maybe cash isn’t even the most effective motivator we could offer.
More sports? Of course, we could always run Stat Geek Idol for more than just college basketball.

If you’re got any thoughts or opinions on how we can improve the contest, please let us know in the comments below.

Congratulations Jordan, and thanks to all of the enterprising stat geeks who submitted an entry to Stat Geek Idol 2, as well as to all of the very busy judges who volunteered their time! As John Gasaway put it best on Twitter:

The post Jordan Sperber Wins $2,000 Grand Prize In Stat Geek Idol 2 appeared first on Notes from the Sports Nerds.

Measuring Offensive Balance In Basketball (Stat Geek Idol 2 Finalist)

2013-04-08T16:50:52Z

See more at TeamRankings.com

This post is one of the five finalists in our second Stat Geek Idol contest. It was conceived of and written by Stephen Shea.

In 2006-07, the champion Florida Gators employed a balanced offensive attack, with five players averaging between 10.3 and 13.3 points per game. In contrast, the 2010-11 UConn Huskies relied heavily on the shots of Kemba Walker. The amount of balance in an offense can vary greatly between teams, and the game has seen champions at both ends of the spectrum. We quantify offensive balance (OB). We observe a surprisingly high correlation between OB and rank among the AP’s top 25 teams.

Higher OB corresponds to lower rank with one extreme outlier. In 2011, 11-seeded VCU made headlines with their run to the final four. VCU features an aggressive full court press, a system that requires a great deal of effort from its players on the defensive side of the ball. Perhaps reflecting Shaka Smart’s atypical system, the Rams are the one anomaly in our statistical analysis.

Defining Offensive Balance

How does one quantify the degree of balance in an offense? Looking at the season statistics (as of March 4) for Gonzaga and Indiana in Table 1, it is not obvious which offense is more balanced.

Table 1: Gonzaga and Indiana Statistics through March 4, 2013

In Economics, one might be interested in the concentration in a particular market. How many companies are selling the same product, and what percentage of the market does each company have? Concentration affects competition and ultimately price and profits. Economists sometimes aggregate the concentration using information theoretic entropy formulas (from [4]). See [1] for more details. Motivated by the work of the economists, we will use entropy formulas in the calculation of OB. Instead of evaluating the concentrations of market shares, we will be aggregating the concentrations of minutes played and field goals made by players on a given team. Let our team T have players {1,2,…,N}. Let m_i denote player i’s minutes played, and let

Let M_i=m_i/m_T, the percentage of minutes for player i among all minutes played. Similarly, define FGM_i to be the percentage of field goals made by player i among all players on team T. Define the distribution of minutes (DM) and the distribution of field goals made (DFGM) for this team to be

Intuitively, the more players to play and the more balanced the minutes, the larger DM will be. Define OB=DFGM/DM. OB generally falls between .90 and 1, with higher numbers representing more balance. Kemba Walker’s UCONN Huskies of 2010-11 had an OB of .911. The idol of all gunners, Jimmer Fredette, ball-hogged BYU to a .890 that same year. The 2006-07 Florida Gators had an OB of .989.

Over the course of a season, several factors can inflate DM, such as injuries and back-up players getting extra minutes in a blow out. DM drives DFGM, as demonstrated in Chart 1. Neither DFGM nor DM correlates highly with rank for the AP top 25. By considering the ratio of DFGM to DM we are asking the degree which all 5 players on the court share the offensive load, regardless of how many players have seen significant playing time. Surprisingly, Chart 2 shows a correlation between OB and rank for this year’s top teams.

Chart 1: DFGM to DM for AP Top 25 as of March 4, 2013

AP Top 25 as of March 4, 2013

We believe the high correlation demonstrated in Chart 2 is atypical. We do not mean to imply that college basketball teams should strive to have a low OB. Although, there have been some impressive basketball teams with unbalanced offenses. For example, the 1995-96 (72 win) Chicago Bulls had an OB of .919. We present the correlation in Chart 2 as a characterization of this year’s best teams. None of this year’s top 10 teams use a balanced offensive attack comparable to the 06-07 Florida Gators.

Chart 2: OB to Rank for AP Top 25 as of March 4, 2013

Table 2 contains the OB for each of the last 10 championship teams. While the range of outputs nearly covers the spectrum of possibilities, there has not been a champion recently with as unbalanced an offense as Fredette’s BYU team of 10-11.

A low OB and high rank typically means that the team has an elite offensive player or two. The top teams in this year’s rankings reflect this phenomenon. Gonzaga has Harris and Olynyk. Indiana has Zeller and Oladipo. Duke has Plumlee and Curry. The correlation in Chart 2 shows that the higher ranked the team, the more that team’s offense leans on its couple stars. The one outlier in the data, marked by the red dot, is VCU. If we were to remove VCU, the value of R² goes to .707.

Table 2: OB for Last 10 Tournament Champions

The VCU Anomaly

In Chart 2, VCU stands out. They have the lowest OB of the group, but are ranked 21st. As discussed above, a high rank and low OB generally means that the team has an elite player or two. Unlike Gonzaga, Indiana, and Duke, VCU does not have any players typically considered to be Wooden Award contenders [3]. Instead, VCU’s position out of position on Chart 2 is a statistical representation of the team’s unique system.

According to [2] “Smart coaches one of the most distinctive styles in NCAA ball.” The article backs up the claim with some statistics. VCU allows only .92 points per possession while scoring 1.13 points per possession. VCU also forces the highest turnover rate in the country. Table 3 contains the minute and field goal data for VCU.

Table 3: VCU Statistics as of March 4, 2013

Future Directions

Next, we would like to understand how OB varies depending on time and score in the game. We suspect that teams generally will have specific players that they look to in clutch time and that OB will drop in these situations. We would also like to study OB in conjunction with other statistics that may help explain why teams are generating high or low OB. Does OB have predictive power? Are certain teams more successful against teams with unbalanced offenses? Finally, we would like to observe OB over more teams across more years, and even in different eras. We would like to know if OB varies greatly depending on the league and level of competition. How would the great Wooden led UCLA teams compare? Does OB generally follow coaches, or is it more dependent on the specific players?

References

[1] Bicker, Jacob and Haff, Katharina. “Measures of Competition and Concentration in the Banking Industry: A Review of the Literature,” Economic and Financial Modeling 2002, Vol. 9, pp. 53-98.

[2] Haley, Jeff. “NCAA Bracket Predictions: Three Final Four Sleeper Picks for 2013.” TEAMRANKINGS.com, March 8, 2013.

[3] King, Jason. “Wooden Award Watch List Announced.” ESPN.com, January 19, 2013.

[4] Shannon, Claude. A mathematical theory of communication. Bell Systems Technical Journal, 27: 379-423, 623-656, 1948; Republished, University of Illinois Press Urbana, IL, 1963.

The post Measuring Offensive Balance In Basketball (Stat Geek Idol 2 Finalist) appeared first on Notes from the Sports Nerds.

Charting Screens In College Basketball (Stat Geek Idol 2 Finalist)

2013-04-08T15:57:59Z

See more at TeamRankings.com

This post is one of the five finalists in our second Stat Geek Idol contest. It was conceived of and written by Joshua Riddell (@Joshua_Riddell).

Teams have the ability to measure or quantify nearly every aspect of a basketball game in today’s game. With the use of Synergy Sports, they can easily pin down how they succeed on offense against man and zone defenses, how strong they are on the pick and roll, how their offense performs in transition and countless other scenarios. Many teams chart other aspects of the game themselves, including deflections, most famously done by Louisville and Indiana and recently profiled in this NPR article.

This begs the question; why don’t teams chart and quantify screens set by their players?

Synergy Sports can measure the effectiveness of players using screens but the action of setting the screens to free a teammate seems to be an undervalued commodity. Brett Koremenos first proposed this idea on Hoopspeak and gave some excellent reasons as to why teams need to begin measuring this crucial basketball action. The argument could be made that the primary role of a coach in the preseason is to establish roles for each team. If a player’s offensive role is established as a screener, he has little way of knowing how well he has done, nor does he have any attainable goals on how many screens to set in each game, even though his role could be a critical part of a successful offense. Screening is a skill and measuring it will encourage players to set their mind to improving this portion of their game.

If teams do chart and count screens set, they do not publicize this fact like some teams do with deflections, charges taken or other similar metrics. Coaching staffs see deflections as an integral part of a fierce defense and provide their players with concrete goals for this statistic and follow up by counting each deflection. While it could vary based on the offensive scheme, the argument could be made that screens are just as essential to an offense.

So why isn’t this aspect of the game given as much attention as other aspects? And, if teams do chart screens, could this exercise provide any value to help the team’s success on the offensive side?

Charting Screens

I wanted to see if charting screens could provide any usable information for a coaching staff. To do so, I charted three games for the top two offenses in 2013, as measured by Ken Pomeroy’s offensive efficiency, Michigan and Indiana.

To chart screens, you have to accept that there is a large measure of subjectivity in determining what is involved in a countable screen. What one person sees as an effective screen may not be the same as someone else. Here are a few points I tried to adhere to when charting screens:

All screens that I deemed to be effective were counted, even if it did not result in an immediate basket, shot or pass to the player using the screen. If the player was open after using the screen, I counted it, even if nothing happened as a result of the screen.
I counted screens that gave the teammate space, even if the screener did not make contact with the defender. If the screen was effective enough to get his teammate open for a possible drive, shot or received pass, I gave the screener credit as I determined that the screener did his job.
I generally did not count screens on baseline or sideline out of bounds plays, unless the play led directly to a basket.
All possessions were charted in this exercise, which includes backcourt turnovers and fast breaks.

Michigan Screens

Michigan is an interesting first case for this exercise, as their offense does not rely on setting screens to get players open regularly. Instead, they rely more on cutting and individual dribble penetration by their guards to give their players good looks at the basket. The majority of their screens come from the big men on pick and roll screens, as their guards rarely set screens.

In charting three games against South Dakota State, VCU and Kansas, I counted the following screens, with the minutes played added for reference. (Just click the table to enlarge it.)

McGary clearly deserves a ton of credit for his screening. He had two noticeable big screens, one in the VCU game and one in the Kansas game, but he was a workhorse throughout the two games and his contributions on both ends of the floor are a big reason Michigan has advanced to the Elite Eight.

As expected, there is no noticeable benefit from setting more screens per possession, as offensive efficiency does not increase as the number of screens increase in this small sample. Even over a larger sample, I would assume there would be no linear relationship between screens set and points per possession.

Michigan would benefit most from charting by tailoring this exercise to track which player sets the most effective ball screens and utilize this player more often in that role. By giving their big men an attainable goal to reach of effective screens set, they would focus more on setting screens and freeing their teammates. This would mean their dynamic guards would be more dangerous by improving the ball screens they use to create off the dribble.

Indiana Screens

Indiana’s offense is built more on screening than Michigan’s and we see that both in the number of screens set and the players who set the screens. While Michigan’s guards rarely set screens, Indiana’s guards screen on a regular basis as shown by the data from their three tournament games. Indiana’s big men still set the most screens on the team, but the guards get in on the screening action as well. (Just click the table to enlarge it.)

The Syracuse zone played a major role in skewing this data downward, as they set only 33 screens against the zone. Just looking at the first two games, Indiana set 1.1 screens per possessions in these games, which decreases to 0.9 screens per possessions when incorporating the Syracuse game.

The points per possessions slowly trends upward for Indiana in this sample but I would expect the data to look similar to Michigan’s over a larger sample size. Even with the different styles of offense, I don’t foresee a trend to appear showing that offenses perform better based on the number of screens set per possession.

Is The Data Useful?

After charting these six games, there are definitely some uses for this statistic but also some words of caution. This is a small sample of games, so looking at how the offense performs on a per possession basis based on screens set should already be taken with a grain of salt and I think it should be taken lightly, if emphasized at all, on a larger scale. Each team’s sample is further skewed by opponents who forced them to change their offense; Syracuse by playing their zone defense and VCU by forcing an up-tempo transition game. The data shows that there is little correlation between screens set and points scored and I expect that lack of correlation to continue as the sample increases.

Providing the offense with this information runs the risk of turning them into robots by screening just to screen and counting the screens before they look for a shot. Coaches need to be wary when citing this information or they run the risk of bogging down their offense as players focus too much on the number of screens set.

Where this information will be useful is on the individual level and counting the number of screens set by each player. Players will be able to take pride in their screening ability, which in turn, should make offenses run more fluidly as offensive players will be more open after using better screens. Bonus points could be given to the screener for points scored directly off a screen or ‘pancake screens’, as demonstrated by McGary in a screen that took over my Twitter feed in the VCU game.

The Future Of Charting Screens

It’s been shown that players will always base their performance on measurable statistics. Players have been known to not take long-range shots at the end of halves (or NBA quarters) to not harm their field goal percentage. By charting and counting screens and turning it into a counting statistic, you will give players an incentive to become better screeners, by giving them a measurement by which they can base their success in this part of their role in each game. This should make offenses more fluid, leading to easier baskets as players begin to take more pride in their screens.

Teams could modify the charting system based on their own preferences and needs for their offense. They could only chart screens that lead directly to baskets to focus their players on setting good screens and not just setting half-hearted screens to boost their numbers. If they want to dive into team data, they should consider charting only half-court possessions, as charting all possessions tend to skew this number away from the goal of this project. They need to ensure that they chart only what they deem to be acceptable screens, so players do not breeze through screens just to boost numbers but focus on setting a strong screen to free their teammate.

There will likely never be a perfect way to measure screens nor will there ever be a uniform way to measure screens, allowing comparing of players and teams across the country. Teams’ mileage will vary on the usefulness gained from this procedure but the main benefit will certainly be from measuring individual player’s screens and not from measuring the screens per possession and any resulting offensive efficiencies. The hope would be that measuring screens gives the screener a quantitative measurement that players can reach, causing them to set better screens for their teammates. It’s unlikely that this solves the scoring problem, but hopefully it could be one step in the process.

The post Charting Screens In College Basketball (Stat Geek Idol 2 Finalist) appeared first on Notes from the Sports Nerds.

Expected Points And Efficiency (Stat Geek Idol 2 Finalist)

2013-04-08T15:45:49Z

See more at TeamRankings.com

This post is one of the five finalists in our second Stat Geek Idol contest. It was conceived of and written by Greg Matthews (@statsinthewild).

Let me start with an obvious statement: In basketball, players who attempt more shots score more points on average. Other individual stats are also associated with more points, such as rebounds and blocks. Similarly, some statistics, like turnovers, are associated with fewer points scored. In this way, we can compute an expected number of points a player should score based on their other statistics such as shots attempted, rebounds, turnovers, etc. Then we can compare the expected number of points a player should have scored to the actual number of points scored and evaluate players based on their tendency to be above or below their expected number of points. I’m going to call this player efficiency.

In the same fashion, we can do the same thing with teams. Based on the number of shots a team attempts, team blocks, team turnovers, etc. we can calculate an expected number of points a team should score. Additionally, for the team analysis, we further consider the opponents’ statistics since the more shots an opponent attempts the more points on average a team will score. So, based on statistics from a team and its opponent the expected number of points a team should score can be calculated. Then, in the same way as with players, we can examine if teams are consistently scoring above or below their expectation and calculate a team’s offensive efficiency. For teams, a defensive efficiency can also be calculated measuring whether a team’s opponents are consistently scoring above or below their expected number of points. This will be referred to as defensive efficiency. Alright, let’s calculate!

We start by taking box scores for each player from every game and regressing each player’s points against attempted number of field goals (FGA), three point attempts (3PA), free throw attempts (FTA), number of defensive rebounds (DRB), blocks (BLK), and turnovers (TOV), and we find, for instance, that each field goal attempt is worth, on average about 0.91 points. This means the average college basketball player will average about 9.1 points per 10 shots when all other things are held constant. Likewise, a 3PA is worth about 1.1 points on average, though there is considerably more variability involved in shooting a three pointer. (Note: This doesn’t mean teams should only shoot threes since they are worth more on average, it just means that the kind of players who are attempting three pointers make these shots more valuable on average than all other regular field goals.)

Likewise, a free throw attempt is worth about 0.76 points, and defensive rebounds, blocks, and turnovers are worth on average about 0.06, 0.14, and -0.11 points, respectively. Using this, we can take a player’s box score and, based on their FGA, 3PA, FTA, DRB, BLK, and TOV, evaluate how many points they were expected to produce based on their other stats. A good player will score more points than expected, whereas poor players will score less than the expected number of points.

For example, on November 23, 2012 Belmont’s Ian Clark played 26 minutes in a game against Northeastern and shot 13 FGA and 11 T3PA with no FTA. He further had 3 DRB, 0 BLK, and 1 TOV. With these stats a player is expected to score about 14 points. However, Clark actually scored 29 points on 10 of 13 shooting, including 9 of 11 on three point tries. While this was certainly an impressive game, the best game of the whole season by this standard was played by Christian Williams of Jackson State on February 23 of this year. In 33 minutes, Williams took 16 shots, 8 of which were threes and attempted 11 free throws. He added 1 DRB, no blocks, and only had one turnover. This stat line should result in a little over 24 points; however, Williams actually scored 44 points on 13 of 16 shooting including 7 three pointers, while shooting a perfect 11 for 11 from the free throw line.

So, this works for individual games, but what if we want to evaluate the expected number of points for a player over the course of many games? In order to do this, we’d need to have data on the expected minutes a player will play, along with the expected number of shots a player takes, etc. In order to do this we can build a model for each separate component in the prediction: A model for minutes, a model for attempted shots, a model for three point attempts, and model for free throw attempts, etc. We need one model for each of the six factors mentioned in the previous paragraph (FGA, T3PA, FTA, DRB, BLK, TOV). Additionally, a model for the number of minutes played will also be estimated. Once we have all of this we can calculate the expected stat line for a player and based on this we can estimate the expected number of points for a player when they produce their average number of minutes, shots, rebounds, etc.

However, the ways in which players score their points can be very different. Scoring 20 points on 10 shots is much different than scoring 20 points on 20 shots. The first scenario is much more efficient than the second. One way to assess this is to look at the difference between expected points for a game and actual points in a game across all games over the course of the season, and to try to explain some of the excess variability by adding in the effect of a player. This can be accomplished by adding a random effect for each player to the model for predicting points. A large positive estimate will indicate that a player is consistently scoring above their expected points based on their game statistics, and a negative estimate shows they are scoring fewer points than expected. The resulting random effects estimates are what I’m referring to as effectiveness.

The plot below shows the relationship between expected points and effectiveness. That is the y-axis represents the expected number of points a player will score based on their expected stat line and the x-axis is the offensive estimate for each player. The players with the highest expected points this year are Erick Green and Brandon Likins, followed by Lamont Jones, Travis Bader, and Doug McDermott. But you can see that Doug McDermott is scoring points more effectively than Travis Bader and much more effectively that Brandon Likins. Other highly efficient scorers this year include Victor Oladipo, Ian Clark, Marshall Bjorklund, T.J. Warren, and Kelly Olynyk. The full list of players with their expected points and effectiveness scores can be found in this google doc. And if you’re interested in finding a specific player on the plot below, I’ve created an interactive version of the plot here using the shiny package in R.

The same type of analysis can be applied to teams. Each stat from before (FGA, 3PA, FTA, DRB, BLK, and TOV) can be assigned an average number of points to a team based on regression analysis. But here, not only will each team’s stats be considered, but also each team’s opponent’s stats. For instance, every field goal that a team attempts is worth on average about 1.19 points, and every shot a team’s opponent attempts is worth about 0.63 for the team (since more shots by your opponent means more shots for you which means more points, on average).

Likewise, a FTA is worth about 0.65 points for your team, and every free throw that your opponent attempts is worth about 0.29 points on average. Incorporating all of these different stats, we can predict the expected number of points for a team if we knew how many shots they attempted, how many shots their opponent attempted, etc. Since we don’t know this exactly, we can estimate all of these and simulate an expected stat line which we can then use to predict the number of points a team is expected to score with an average stat line.

In order to calculate team effectiveness, we need to look at the actual stat lines rather than a simulated one, and the process is similar to the case with players. Below is a plot of offensive effectiveness versus expected team points. You can see that Indiana and Iona had the highest expected points this season, but they weren’t nearly the most efficient team. The most efficient team this year was Creighton, thanks in large part to Doug McDermott being the most efficient player in the country. The rest of the top five most efficient teams this year includes Richmond, Belmont, San Francisco, and Denver followed by Duke, Kansas, North Carolina State, Pittsburgh, and Kentucky. The least efficient teams this year were Howard, North Texas, and Prairie View and the team with the lowest expected points this year was Grambling.

But offense is only half the story. We can further apply this same type of analysis to each team’s defense. Now instead of just estimating only an effect for offense, a defensive effect is added to the model. In this way, we can derive both offensive and defensive effectiveness ratings for each team in the NCAA. Below is a plot of offensive versus defensive effectiveness for all teams in the NCAA for the year 2013. I’ve highlighted some teams in the plot below, including all of the number 1 seeds (Indiana, Gonzaga, Kansas, and Lousiville), as well as a few other teams that I found interesting, like Kentucky, Ohio State, and Florida.One can see that of all of the number 1 seeds, Gonzaga is the least efficient on offense, but they are the most efficient on defense. Similarly, Indiana is nearly the same as Gonzaga in defensive efficiency, but slightly more efficient on offense. Other teams, like Louisville and Kentucky, are nearly average in defensive efficiency, but well above average in offensive efficiency. (It’s interesting that Kentucky and Louisville fall so close to each other on this plot; one team missed the tournament entirely and the other was a 1 seed.) Finally, there is a team like Florida, who many picked to win the entire tournament, who are about as efficient on offense as teams like Ohio State and Louisville, but are very nearly the most defensively efficient team in the whole country (only Alcorn State and Mississippi were more defensively efficient this year).

So, a big question is, does offensive or defensive efficiency actually mean anything in terms of winning tournament games? In order to address this, I looked at last year’s Sweet Sixteen and Final Four teams. In the plot below, Sweet Sixteen teams are indicated by light green dots and Final Four Teams are represented by their team colors. You can see that nearly all of the Sweet Sixteen teams last year had positive offensive efficiency and negative defensive efficiency, meaning they were above average in both measures.

However, this is likely due to the fact that teams with above average offensive and defensive efficiency are the only types of teams in the tournament to begin with. So, a much more interesting example occurs in looking at the Final Four teams. One can see that all of the Final Four teams from last year had and offensive efficiency above 1, and two of the teams were above 2. Further, the champion last year, Kentucky, can be found in the lower right corner of the graph near Ohio State, indicating both above average offensive and defensive efficiency. The only team close to this area of the graph this year is Florida.

For a full list of teams’ expected points and efficiencies, you can go to this google doc. Finally, an interactive version of offensive effectiveness versus expected points can be found here, and a plot of offensive versus defensive effectiveness can be found here for both 2012 and 2013.

The post Expected Points And Efficiency (Stat Geek Idol 2 Finalist) appeared first on Notes from the Sports Nerds.

Finding Northwestern A New Coach (Stat Geek Idol 2 Finalist)

2013-04-08T16:00:01Z

See more at TeamRankings.com

This post is one of the five finalists in our second Stat Geek Idol contest. It was conceived of and written by Ryan Silvis.

In 1978, a then promising young coach at Army, Mike Krzyzewski, interviewed for the vacant Northwestern coaching job. Northwestern offered the position to a different coach and Coach K went back to Army only to be hired by Duke in 1980. Since, Coach K has led Duke to 29 NCAA Tournaments, while Northwestern is still looking for their first. Northwestern is again looking for a new coach and hopefully Athletic Director Jim Phillips is on the Stat Geek Idol judging panel, because I will tell him who he should hire.

First things first, let’s dismiss what seems to be the popular choice of Duke assistant Chris Collins. Since 2003 only four first time head coaches were hired into a power conference. As you can see in Table 1 below, none of the coaches were able to improve the team three years after they took over. In fact, on average they made the program worse by 18%.

Table 1: Conference Win % from first time head coaches in a power conference

In addition, according to data gathered from the Office of Postsecondary Education, Northwestern spends the least amount of money on its basketball operations in the Big 10. In fact, from 2003-2011, Northwestern has spent $11M less than the average Big 10 school and $40M less than Michigan State. It is foolish for Mr. Phillips to think that any assistant coach can turn a program around with the type of budget.

If Northwestern truly does want to turn the program around, they should hire an established head coach. Since I don’t see any power conference coaches willing to take a pay cut to lead the Wildcats, they will have to look at the mid-majors. The question is, how do you know what coach can turn Northwestern around? To answer that question, I developed a coach ranking system based on two factors, talent accumulation and in-game coaching.

The first component in my coaching rankings is talent accumulation and is broken down into two parts. The first part is how well the coach gets talented players into his program and is calculated based on the productivity of each player in their first season with the coach (including freshman and transfers). The second part is player development and is measured by the increase in productivity over each player’s career. My chosen measure of player productivity is Win Shares and I’m only using conference games from 2003-2013 where the players have played more than 100 conference minutes.

Table 2: Top 10 Coaches Talent Accumulation (Normalized)

Table 2 above shows the top 10 in talent accumulation and includes five mid major coaches and 5 power conference coaches. The beauty of only using conference games is that it measures each coach based on similar competition. All of the mid-major names on this list recruit well for their systems and have accumulated talent that is far superior to the rest of their conference.

Graph 1: Talent Accumulation vs Conference Win %

Graph 1 above shows the strong relationship between talent accumulation and career conference win % with an R² of .6737. Former Northwestern Coach Bill Carmody falls near the bottom 10% of all coaches in this measurement.

Now that we have established how much talent each coach has on their roster we need to find out how well the coach used the talent on his roster in games. To determine an in-game coaching ranking, I took the difference between the Vegas spread and the final outcome for each game. If Team A is favored by five points, Vegas views the talent of Team A as five points better than Team B (also reflecting other factors such as home court advantage). Let’s say Team A wins by nine points. Team A would receive a +4, while their opponent would receive a -4. The problem is that Vegas doesn’t produce a line for the Tuesday night Robert Morris vs St Francis game. To keep everything consistent, I used adjusted efficiencies similar to kenpom to come up with an implied line. The implied line has an R² of .82 and does a good job of approximating team A’s skill vs team B.

Table 3: Top 10 in-game coaches

Table 3 above shows John Calipari is top dog by averaging +3.9 points above the implied line. Bill Self again finds himself in the top 10 and is the only power conference coach to appear in the top 10 of both categories. Other coaches that appear in the top 10 of both lists are John Pastner, Fran Dunphy, and Rick Byrd.

Graph 2: In-Game Coaching vs Conference Win %

Graph 2 above shows in-game coaching correlates a little less to conference win percentage as talent accumulation, but it still has a significant relationship with an R² of .5614. Again, former Northwestern Coach Bill Carmody falls near the bottom of In-Game Coaching.

Table 4: Top 15 Overall Coach Rating

The top 15 coaches are listed in table 4 above. Bill Self has established himself as the man to beat with a wide lead over second place. Five of the top 15 are from power conference schools.

Graph 3: Overall Coach Rating vs Conference Win %

On the right side in Graph 3 above is the Overall Coach Rating vs Conference Win %. There is a strong relationship between the two with an R² of .7328. Some of the outliers highlighted on the right side of the graph (Few, Stevens, and Krzyzenwski) indicate that there is some other factor not considered in my model that causes these coaches to win. I suspect this may be a combination of team prestige and much larger operating expenses compared to other members of their conference.

Now that we have a ranking system, we can use it to help Northwestern pick a new coach who can lead them to the NCAA tournament. Over the past three seasons, Northwestern’s conference winning % has been 34%. To get into the tournament, they are going to have to get to a minimum of 55-60%. That would represent a 20-25% increase in winning percentage. On top of that, Northwestern’s basketball operations have had a budget that is almost 25% less than the average team in the Big Ten. Since 2003, four coaching changes have occurred in a power conference (plus the A10) where the new coach increased winning percentage by more than 20% with a budget of less than 75% of the conference average. On average, the new coach had a talent accumulation rating of 53 better than the old coach. In addition, the new coach’s had an average in-game coaching rating of 69 better than the old coaches. Table 5 shows a summary of this information. (Click the table to view a bigger version.)

Table 5: New Coaches with more than 20% improvement in win %
with less than 75% of conference spending average

Using Northwestern’s previous coach Bill Carmody’s ratings of 24 and 28, Northwestern should hire a new coach with ratings of 77 for talent accumulation and 97 for in-game coaching. Graph 4 below shows the improvement made to the coaching changes that resulted in a 20% winning percentage increase.

Graph 4: Change in coaching ratings that resulted in a win % increase of more than 20%

Graph 5 below depicts current Big Ten coaches and where Northwestern needs to be in order to increase their winning percentage by more than 20%.

Graph 5: Current Big Ten coaches and the improvement Northwestern needs from their next coach

Of all the current active coaches only three have greater than a 77 for talent accumulation and 97 for in-game coaching: Memphis’ Josh Pastner, Temple’s Fran Dunphy, and Davidson’s Bob McKillop. Josh Pastner and his $1.7 Million salary is too high for Northwestern. Bob McKillop appears to be staying put at Davidson.

That leaves us with Fran Dunphy. Before last season, Dunphy’s peers named him the most underrated coach in America. When Dunphy took over the Temple program, it has not been to the NCAA tournament in five years but he has since took them to the tournament six out of the last seven years. Dating back to when Dunphy was at Penn, he has made the NCAA tournament 15 times. Dunphy would be hard for Northwestern to land, but he currently makes $700,000 a year which is quite a bit less than the estimated $1.1 Million Bill Carmody made. And who knows, maybe Dunphy doesn’t want to be the most underrated coach in America and wants to be known as the guy who took Northwestern to the NCAA tournament?

The post Finding Northwestern A New Coach (Stat Geek Idol 2 Finalist) appeared first on Notes from the Sports Nerds.

Do Matchups Matter In College Basketball? (Stat Geek Idol 2 Finalist)

2013-04-08T14:26:55Z

See more at TeamRankings.com

This post is one of the five finalists in our second Stat Geek Idol contest. It was conceived of and written by Jordan Sperber (@hoopvision68).

Introduction

There are 347 teams in Division I college basketball. The nature of the sport allows for all different kinds of styles of play. Every team has varying player personnel and coaching philosophy. College basketball analysts are given the tough task of forecasting the end result of games featuring contradicting styles. It seems undeniable that, in some cases, certain teams can be bad matchups for other teams. Still, quite frequently analysts just say what sounds good. To illustrate this point, let’s look at a first round matchup from this year’s NCAA tournament:

The four factors (shooting, rebounding, taking care of the ball, and drawing fouls) are a very good way to assess style of play. The Minnesota-UCLA matchup featured the best offensive rebounding team in the country (Minnesota) and the 263rd defensive rebounding team in the country (UCLA). A smart analyst would point this compatible Minnesota strength and UCLA weakness out, but what does it really mean for the expected outcome of a game? On one hand, Minnesota should kill UCLA on the offensive boards, possibly creating a huge advantage for Minnesota. On the other hand, Minnesota kills just about everyone on the offensive boards. UCLA wouldn’t be able to stop the lethal Minnesota rebounding attack regardless, so maybe this is a waste of an opponent weakness for Minnesota.

Essentially, the question I am asking here is simply: If you are really good at one of the four factors, would you rather play a team that is normally good at defending that factor (strength on strength) or really bad at defending that factor (strength on weakness)? At first thought, strength on weakness feels like the right choice. The goal of the following analysis is to try to answer this question.

Part 1: The effects of opponent on each four factor

To begin this study, I compiled a sample size of every single Division I college basketball game from 2009, 2010, 2011, and 2012. Games from 2013 (up until around the first week of March) were also included. I wound up with exactly 26,000 games to draw conclusions from.

In order to look at what happens when a {good/bad} offensive {eFG/TO/OR/FTR} team played a {good/bad} defensive {eFG/TO/OR/FTR} team, I had to define what exactly good or bad means. I decided that any team in the 90th percentile or better of a given four factor was “good” at that skill and any team in the 10th percentile or worse of a given four factor was “bad” at that skill.

The next step was to use these definitions of good and bad to find instances of strengths meeting strengths, weaknesses meeting weaknesses, and so on in the 26,000 game sample. First, let’s take a look at what happens when a good shooting team plays a good defensive shooting team:

The above table shows that there were 409 games where a good offensive eFG% team played a good defensive eFG% team. The offense averaged an eFG% of 54.5% on the season. However, when they played a good eFG% defense, that number decreased to 49.8%.

I did this same analysis for all types of matchups and all the four factors. The results are below (click the image to enlarge the tables):

There is a lot going on here, but the two biggest takeaways are:

1. Bad vs. bad brings out more good than good vs. good brings out bad. Basically, when two bad teams at one factor play each other, the offense improves a lot. When two good teams at one factor play each other, the offense does not diminish quite as much.

2. The defense controls FTR the most and eFG% the least. If you look at the percent change column, an offenses ability to get to the foul line changed a lot depending on the defense. On the other hand, an offenses ability to make shots did not change nearly as much. This is consistent with past research on similar topics.

Part 2: The effects of style on efficiency

Part 1 showed exactly what happens to the individual four factor based on opponent, but that is only so helpful in determining if there is evidence for good/bad matchups. The more important thing to look at is the effects of style on points per possession. Let’s go back to UCLA-Minnesota. Say UCLA decided that they needed to make an extra effort to keep Minnesota off the offensive glass. This decision might come at the risk of a different four factor. Maybe UCLA focusing on defensive rebounding diminishes their ability to create turnovers. This idea wouldn’t show up in the part 1 results, but it would show up in points per possession.

To look at the effects of efficiency, I first calculated an expected points per possession using simply the ORtg (adjusted for schedule) of the offense and the DRtg (adjusted for schedule) of the defense. This expected PPP was made without looking at matchups or style of play at all. Then, the expected PPP could be compared to the actual PPP. If the two numbers significantly differ, that means that mismatches in four factors can give us more information on which team will most likely win the game. (Again, click the image to enlarge.)

As you can see, matchups had virtually no effect on the actual points per possession of the game. I was able to predict PPP by simply using the offensive and defensive averages extremely effectively. Here are the final key takeaways from the tables above.

1. Four factor matchups don’t increase prediction accuracy. If we once again go back to Minnesota-UCLA, this means that we shouldn’t have looked too far into the offensive rebounding advantage. Simply looking at which team is better efficiency wise is adequate.

2. FTRate had very little effect on the points per possession of an offense. If you look at the Actual PPP column, there is not much change in general. This particular study indicates the eFG% is the most important four factor, followed by OR%, TO%, and finally FTR.

Conclusion

It would be foolish to say that specific matchups have no effect on the outcome of a basketball game. It doesn’t mean that matchups can’t possibly matter just because this study shows no evidence for it. However, the study does indicate that it may not be wise to focus too much on the compatibility of the strengths and weaknesses of opponents. Trying to breakdown strength and weaknesses may be a futile activity. Simply put, the best way to predict the winner of a game appears to be just picking the better of the two teams.

The post Do Matchups Matter In College Basketball? (Stat Geek Idol 2 Finalist) appeared first on Notes from the Sports Nerds.

2013 MLB Predictions: Projected Standings, Most Likely World Series, Preseason Ratings

2013-04-01T16:41:26Z

See more at TeamRankings.com

The 2013 MLB season starts in only two days, as the Rangers and Texans kick things off with a Lone Star State Showdown. With opening day fast approaching, it’s time to release our preseason ratings and projections.

The main purpose of these ratings is to provide a data-driven starting point for our MLB projected standings. Just like last season, we’ll have fully automated win-loss predictions, playoff chances, and World Series win odds, and all the info will be updated every single day of the season to reflect the latest results and the most up to date MLB power ratings.

A word of caution — while our preseason projections for other sports have proven to be useful indicators of where values may lie among the various full season futures bets, we’re not nearly as confident in our MLB preseason ratings. We’re publishing these in the interest of full disclosure, so that you know what the initial rating in our projection system was for each team. We’re most definitely not recommending that you use these ratings and forecasts to go place preseason bets.

You may look at the projections below and think that they aren’t extreme enough. In a way, you’d be right — the best team in the league will almost certainly win more than the 90 games we’ve forecast for the Yankees, and the worst will likely lose more than 95. However, picking which teams will wildly exceed expectations is rather tricky, and on average these conservative predictions should provide a less biased starting point than more aggressive ones.

However, if you’d like to see our best case and worst case scenarios for each team, check out their team projections page. Here are the New York Yankees projections as an example. Follow that link and you’ll find a chart showing the projected odds of the Yankees winning any specific number of games, as well as a list of their toughest & easiest games, and a table showing how their chances of winning the World Series change depending on what seed they get in the playoffs.

Quick Predictions For The 2013 MLB Season

Let’s go over a few of the insights these projections provide, before laying all the details out below.

The most likely World Series result is the New York Yankees beating the St. Louis Cardinals. Of course, that exact outcome only has about a 1 in 100 chance of occurring.
The NL East appears to be the most balanced division — it’s the only division where no single team has a 30% or better chance to win.
The Detroit Tigers in the AL Central appear to be the biggest lock for a division title. However, our projections still don’t see them as a sure thing, putting their chances of a division crown at only 40%.
The race for the two NL Wild Cards is wide open. Only Pittsburgh, Miami, Colorado, and Chicago have less than a 10% shot at snagging a Wild Card berth, and favored Washington has only an 18% chance.
The top 5 teams are all in the AL East and AL West. Yet at least one of the Yankees, Rays, Rangers, Angels, and A’s is going to be staying home for the playoffs.

Here is how the playoffs would play out, if these projections end up being spot on:

Wild Card Round: Rays over Angels; Braves over Phillies
Division Round: Yankees over Rays, Rangers over Tigers; Cardinals over Braves, Nationals over Giants
League Championship Series: Yankees over Rangers; Cardinals over Nationals
World Series: Yankees over Cardinals

Full 2013 MLB Preseason Projections

2013 MLB Preseason Projections
AL East	W	L	TR Rank	Playoffs	Wild Card	Win Div	Top Seed	WS Champs
NY Yankees	90	72	1	60.2%	22.8%	37.4%	18.0%	8.9%
Tampa Bay	88	74	2	56.1%	23.1%	33.0%	14.4%	7.5%
Baltimore	81	81	13	25.9%	14.8%	11.1%	4.0%	2.3%
Boston	79	83	18	22.7%	13.1%	9.6%	3.3%	1.9%
Toronto	78	84	19	21.3%	12.4%	8.9%	2.8%	1.7%

AL Central	W	L	TR Rank	Playoffs	Wild Card	Win Div	Top Seed	WS Champs
Detroit	86	76	6	50.6%	10.9%	39.7%	10.1%	5.5%
Chi Sox	83	79	16	39.7%	10.4%	29.3%	6.2%	3.6%
Kansas City	78	84	21	23.1%	8.1%	15.0%	2.4%	1.6%
Minnesota	74	88	27	12.4%	4.4%	8.0%	0.9%	0.8%
Cleveland	74	88	28	12.0%	4.0%	8.0%	0.8%	0.7%

AL West	W	L	TR Rank	Playoffs	Wild Card	Win Div	Top Seed	WS Champs
Texas	88	74	3	54.3%	21.0%	33.3%	13.6%	6.8%
LA Angels	86	76	4	46.7%	19.6%	27.1%	10.4%	5.3%
Oakland	86	76	5	46.4%	19.8%	26.6%	9.7%	5.1%
Seattle	79	83	20	23.1%	11.4%	11.7%	3.5%	1.9%
Houston	67	95	30	3.1%	1.8%	1.3%	0.1%	0.1%

National League
NL East	W	L	TR Rank	Playoffs	Wild Card	Win Div	Top Seed	WS Champs
Washington	86	76	8	47.6%	18.3%	29.3%	11.4%	5.3%
Atlanta	85	77	9	47.4%	17.9%	29.5%	11.7%	5.2%
Philadelphia	84	78	11	41.2%	18.2%	23.0%	8.2%	4.0%
NY Mets	78	84	23	23.1%	11.7%	11.4%	3.4%	1.7%
Miami	75	87	25	14.7%	7.8%	6.9%	1.8%	0.9%

NL Central	W	L	TR Rank	Playoffs	Wild Card	Win Div	Top Seed	WS Champs
St. Louis	86	76	7	52.5%	17.1%	35.4%	13.5%	5.9%
Cincinnati	84	78	14	42.1%	16.0%	26.1%	9.1%	4.1%
Milwaukee	83	79	17	39.9%	15.5%	24.4%	7.6%	3.8%
Pittsburgh	77	85	24	18.8%	9.5%	9.3%	2.4%	1.3%
Chi Cubs	73	89	29	10.0%	5.2%	4.8%	0.7%	0.5%

NL West	W	L	TR Rank	Playoffs	Wild Card	Win Div	Top Seed	WS Champs
SF Giants	85	77	10	45.6%	14.5%	31.1%	10.9%	5.0%
Arizona	83	79	12	39.7%	14.7%	25.0%	7.9%	4.0%
LA Dodgers	83	79	15	38.0%	14.1%	23.9%	7.1%	3.8%
San Diego	78	84	22	23.6%	10.7%	12.9%	3.3%	1.8%
Colorado	75	87	26	13.7%	6.5%	7.2%	1.2%	0.9%

The post 2013 MLB Predictions: Projected Standings, Most Likely World Series, Preseason Ratings appeared first on Notes from the Sports Nerds.

Attention Stat Heads: Win $2,000 In Stat Geek Idol 2 — Due March 31

2013-03-28T15:33:11Z

See more at TeamRankings.com

After a very successful inaugural year, TeamRankings’ Stat Geek Idol competition is back!

Last March, Jeff Haley captured the crown with his analysis of play-by-play data showing the impacts of pace, drawing rave reviews for his work from SGI final round judges including Mark Cuban, Dean Oliver, Ken Pomeroy, and Jeff Ma.

Who’s going to impress the judges and bring home the greenbacks this year? If you’re an armchair stat geek, this is your big chance to get your work noticed by some of the biggest names in basketball analytics and media!

Grand Prize

$2,000 Cash

Since last year’s contest was so successful, we’re doubling the prize for the 2013 winner to two grand. We were inspired by last year’s response and want to raise the bar even higher.

Finalist Judges

Mark Cuban, Owner, Dallas Mavericks – Shark #1, Shark Tank
Dean Oliver, Director of Production Analytics, ESPN – Author, Basketball On Paper
Ken Pomeroy, Owner, KenPom.com – College basketball team consultant
Jeff Ma, Founder, TenXer & Citizen Sports – Former member, MIT Blackjack team
Ben Alamar, Professor, Menlo College – Author, Sports Analytics: A Guide for Coaches, Managers, and Other Decision Makers
Luke Winn, Senior college basketball writer, Sports Illustrated
John Gasaway, College basketball analyst, ESPN Insider
John Stasko, Professor & Associate Chair, Georgia Tech – Faculty, CS 4801 SA, Sports Analytics
Tobias Moskowitz, Professor, University of Chicago – Author, Scorecasting: The Hidden Influences Behind How Sports Are Played and Games Are Won
Jeff Haley, Founder, Hoop-Math.com - Stat Geek Idol champion 2012

Submission Deadline

Sunday, March 31 @ 11:59 pm Pacific time

See the bottom of this post for how to enter.

Submission Format

This year’s submission format will be different than SGI1. In terms of a deliverable, this year we are giving you a 374 square inches of blank canvas (that’s four 8.5” x 11” pages) to blow us away with something analytical about college basketball. We are intentionally defining the format in a vague way because we don’t want to constrain creativity or skill sets. Just send us four pages of work.

A four page article, all text, works. Four pages of data visualizations with hardly any text? Yup, that also works. Or anything else that fits. Just keep in mind if you focus on graphics, that whatever you send us, we’re going to scale it to 374 square inches, print it out, and judge what we see. If it’s not easy to read and understand at that size, that’s bad for you.

Since large data sets often sit behind interesting analysis, we don’t expect you to try to cram all of the raw data that powers your SGI entry into four pages. However, if we have questions or concerns during the judging process, we may ask to see your underlying data set and/or learn more about its sources.

Contest Format

Last year, we structured SGI as a multi-round “tournament” to mimic March Madness, with three rounds of submissions. We’re simplifying for SGI2. It’s just one single submission, due on March 31.

After a review process by Team Rankings staff and some external quantitatively-minded peers, a short list of SGI finalists will be announced within several days of the submission deadline. These finalists will then be further judged via either a two or three dimensional process:

The rankings of a panel of “celebrity” judges (to be announced)
The reviews of an analytical panel set up by Team Rankings
Possibly: An audience voting component (if we have time to do it)

We plan to announce the $2,000 winner within a day or so of the 2013 NCAA tournament championship game on Monday, April 8.

Topic

You can analyze anything that relates to college basketball in an important and meaningful way. It doesn’t have to be specifically about this season, either.

Judging Criteria

The three things we’re looking for are straightforward:

Quality/rigor of analysis
Uniqueness/creativity of analysis
Presentation of analysis

A great SGI submission presents ironclad quantitative reasoning that either sheds light on something previously unknown yet important about college basketball, or confirms or dispels conventional wisdom about the game. Either the topic is completely fresh, or it’s a completely new and unique way of exploring an existing topic/debate. Finally, it is delivered in a way that is easy — and ideally entertaining — for a decently intelligent college basketball fan to follow.

You should assume that your audience for SGI are college hoops fans who respect and enjoy data-driven analysis, but who aren’t practitioners or math nerds themselves. If you drop a bunch of heavy mathematical or statistical terms without simple and clear explanation, you are going to lose them.

On the other hand, you should assume they do already have a grounding in the more foundational basketball research that’s been put out. They understand “the myth of the hot hand” and why per-game stats are silly, for example, so no need to regurgitate the basics.

It’s also OK if your submission has already been published elsewhere within the last year. Our goal is to reward great analysis, not just great analysis done on demand.

For reference, here’s the Final Four from 64 total entries competing in 2012:

Jeff Haley: Hurry Up Offense: How Pushing The Pace Affects Shooting & Rebounding
Gregory Matthews: 4 Infographics To Prep You For The Final Four
Jordan Sperber: A Video Charter’s Guide To The Final Four
Nathan Walker: Coaches Love Assists … And Turnovers?

How To Enter

Register for the contest if you haven’t yet.
(If you registered last year, you don’t need to do it again.)
Email your submission to idol@teamrankings.com before the submission deadline

Any questions, just drop a comment below, and please help us spread the word!

The post Attention Stat Heads: Win $2,000 In Stat Geek Idol 2 — Due March 31 appeared first on Notes from the Sports Nerds.