College Bowl Season 2012-13 Recap: Our Bowl Pick’em Advice Hits Again

posted in College Football, College Football Bowls, College Football Pick'ems

Now that the 2012-2013 college bowl season has ended in a perhaps unsurprising fashion — and I don’t mean Alabama winning, I mean the entire sports world up in arms over something Brent Musberger said — it’s time to look back and grade the performance of our bowl season predictions and contest advice.

For those new to the site, as part of our premium services we offer picks for college bowl pick’em contests, which seem to be getting quite a bit more popular over the years. We also offer computer-determined betting predictions (point spread, over/under, and money line value picks) for every bowl game.

In summary, we had our third straight strong year with our bowl pick’em advice, while our betting pick performance was mixed.

2012-2013 College Bowl Pick’em Results

We published a total of 12 bowl pick’em pick sets this year, 9 for competing in game winner based bowl pick’ems with confidence points, and three for game winner contests without confidence points. As a group, they performed very well, including a few outstanding results.

  • 11 of our 12 published pick sets finished in the top 3.5% of ESPN and the top 6% of Yahoo. That generally should be good enough to come in top five or six in a 100-person pool, top two or three in a 50-person pool, and first or second place in a 25-person pool.
  • Our non-confidence pick sets did especially well this year on ESPN, with 2 of 3 finishing in the 99.6th percentile. On Yahoo! they were 97th percentile.
  • Our worst pick set out of all twelve still beat 85% of the nation, and that was a pick set targeted for a large pool, which uses more of a boom-or-bust strategy.

(Note: Why does the exact same pick set end up with a different percentile on Yahoo! compared to ESPN? We can’t say exactly. It could be that in general, better sports pickers compete on Yahoo!, or there could be differences between the two sites in how they compute percentiles, or both. But over the years, our pick sets consistently seem to do about 2% worse on Yahoo! than ESPN, although both sites usually get at least several hundred thousand users in their contests, and on the order of 5 million entries into bracket contests. Go figure.)

We’ve already heard from several subscribers relating tales of first and second place finishes in small to midsize bowl pick’em pools using our advice, and from some contestants in larger pools who finished just out of the money but very close to cashing.

We’re very happy with this performance, especially because it was our first time using some brand new technology we’ve been developing since the summer to optimize picking strategies for bowl pick’em contests. We still have several improvements we plan to make to the overall optimization logic, but given these initial results we are very excited about the long-term promise of what we have built so far.

Detailed Bowl Pick’em Pick Set Results

Confidence Point Pick Sets
NameESPN PercentileYahoo! Percentile
Small Pool 197.095
Small Pool 297.796
Small Pool Mike97.395
Midsize Pool 198.497
Midsize Pool 297.595
Midsize Pool Mike96.694
Large Pool 198.797
Large Pool 287.1n/a
Large Pool Mike97.896
Non-Confidence Pick Sets
NameESPN PercentileYahoo! Percentile
Huge Pool99.697

As mentioned above, we published 12 pick sets in total for bowl pick’em contests. We highlighted our primary pick sets, but also included alternative pick sets for users who want to play in multiple contests with varying picks, or who want to enter the same contest with several different pick sheets. Although 9 of our 12 published sets of bowl pick’em picks were 100% computer determined, we also published three pick sets from original TeamRankings founder and contest strategy guru Mike Greenfield, both for nostalgia’s sake and to pit man vs. machine in a brutal bowl picking death match. (However, calling Mike’s picks “man” is pushing it; he’s obviously very analytical in his approach.) The machines won, although it was close.

Since contest strategy depends largely on how big your pool is, we targeted each of our pick sets to one of three pool size ranges: Small, Midsize, Large. The exact definitions of each size range were slightly different for confidence vs. non-confidence point pools. We entered all our pick sets into both ESPN’s and Yahoo!’s bowl pick’em contests before the first bowl of the season kicked off, to see how they fared against the nation as a whole.

College Bowl Betting Picks Results 2012-13

In contrast to the bowl pick’em picks, the performance of our bowl betting picks was mixed this year, and down overall. Our over/under picks ended up doing well again, but after starting out strong, our bowl point spread picks went on a losing streak that even a hot finish couldn’t totally erase.

Overall, our playable (2- and 3-star) spread and totals picks went a combined 20-21 (48.8%), two wins short of profitability; our playable bowl money line value picks went 4-5 for a -2.1 unit loss. Clearly those are not the results we hoped for, but at the same time, there’s no way we’re going to have winning bowl seasons every year. Let’s break it down in more detail.

Bowl Point Over/Under Picks

First the good news: our playable over/under picks for the 2012-13 college bowl season went 10-8 (55.6%), and all totals picks went 20-15 (57.1%, +3.2 units) overall. Playable bowl over/under picks have been solid performers over the past seven years (see our bowl picks performance page for the numbers), and now we get to add an eighth year to that list.

The one thing that raised our eyebrows early on was our models favoring so many “under” plays in a bowl season that was projected by Vegas to be one of the highest scoring in recent memory. No less than 13 of our top 15 rated over/under picks were “under” plays, and overall that ended up being the right side to be on for those games.

Bowl Point Spread Picks

The performance of our bowl ATS picks this year exemplified the impact of randomness and luck (both bad and good) over small sample sizes. We had a couple highs (starting off the season 4-1 ATS, finishing it off 8-4 ATS including going 4-1 ATS in the five BCS bowls) and one serious low, a losing streak in the middle of bowl season that included a highly improbable 8 losses in a row at one point. We clearly would have preferred that losing streak to have taken place during, oh, any other freakin’ time of the year besides bowl season maybe?

When the dust settled after these up and down streaks, our playable (2- and 3-star) bowl spread picks ended up 10-13 for a losing year. Combined with a losing record for our 1-star picks, that performance also meant that anyone playing our ATS picks in a spread-based bowl pick’em that included all 35 games probably had no shot at winning. So we’re unhappy with that, while at the same time being relieved that at least our luck bounced back in the end. If you mostly cared about the BCS bowls or happened to be in a spread-based pick’em that focused on the final games, you probably did pretty well.

Bowl Money Line Value Picks

With only nine playable money line value picks for the bowl season, things could have broken any way, especially given that money line value is often the biggest on large underdogs. Neither of models’ two biggest underdog plays, Pitt and Navy, came through (there was about a 40% chance of getting both wrong); another pick, Rutgers, ended up losing in OT; and another pick, Oregon State, yielded 14 unanswered points in the fourth quarter to lose. Catch a few breaks like that in a 9-game sample, and it’s tough to stay in the black. We ended up going 4-5 for -2.1 units.

It was interesting, though, to see most of the top money line value picks being favorites this year, and that’s something we’re going to look into more in the offseason.

Some Closing Observations

  • Our Decision Tree Model did significantly better than our official TR Picks. The logic for our final TR betting picks blends the results of several underlying prediction models. Our Decision Tree model has the biggest impact on our bowl predictions, but other models like our Similar Games model do have some weight too. And unfortunately, the other models basically sucked this bowl season. How bad did they suck? Well, bowl point spread predictions from our Decision Tree model alone went 19-16 overall, a profitable 54.3% ATS, and 8-6 (57.1%) ATS on picks with a 53% or higher confidence rating. So the other models basically turned a potentially profitable point spread season into a losing one.
  • Decision Tree did better, continued…In addition, 53%+ confidence over/under picks from Decision Tree went 9-5, also several games better than TR picks, and money line value predictions from Decision Tree would have returned +12.8 units of profit. Sheesh. We do know that some of our users specifically follow the DT model’s picks. If you did for bowl season, good for you. But the takeaway for us in the offseason is to look into making the Decision Tree model have even greater weight than it currently does for our final college football predictions.
  • For the final set of bowl games, point spread predictions we made earlier in bowl season outperformed the final graded picks on the site. To those unfamiliar, our algorithmic predictions can often change as a game gets closer to kickoff, based on factors like betting line movement and other game results coming in; we “freeze” our final picks 1-2 hours before kickoff. According to this study, our point spread picks for the last 15 bowl games went a seriously good 11-4-1 (73.3%) ATS … if you had placed bets on all those games in late December, using the TR picks at that time. That’s better than our models ended up doing against near-closing lines in those same 15 games, and it’s something else we’re going to study in the offseason: how our “earlier” vs. “later” picks do.

Wrapping It All Up

While our point spread pick performance prevents us from grading the 2012-13 bowl season as a big success all around, we’re encouraged by the results of our new pick’em technology as well as the fact that our primary model actually did very well with betting picks. In retrospect, we just didn’t “listen” to it as much as we should. There’s no guarantee, of course, that the Decision Tree model will continue to do well in the future, or that our other betting models will continue to underperform. But it’s certainly a much better situation than if all our point spread models had performed poorly, and we can learn from it.

We’ll close with one final point for anyone out there who is interested in betting or winning money on sports in general. If your goal is to maximize your odds to win money on sports, you shouldn’t be betting on individual games, where your edge against the house is usually pretty slim. Rather, you should be getting in as many multi-game contests (bowl pick’ems, March Madness brackets, etc.) as you can. The ROI opportunities are much greater, and we’ll be publishing some more information related to this topic in the future.

  • Scott C.

    I used your large pool 1 and 2 this year to finish 36th and 55th out of 545 entries. Last year I finished 67th and 91st out of 455 entries. The pool pays out the top 15 to 18 places. I do not doubt that I will finish in the money one of these years as 1 or 2 games in my favor makes a huge difference because confidence points are subtracted (instead of getting a zero) if your pick is wrong. Thanks for giving me a good shot to win it! Looking forward to March Madness.

    Scott C.


    Thanks for letting us know, Scott. Interesting, I’ve never heard of a “subtract points” rule like that. I wish you had let us know before, as it definitely would have changed the strategy a bit, most likely biasing it toward being more conservative. Out of curiosity, which pick sets did you play exactly?

    You’re dead on that winning a 500+ person pool is pretty much a shotgun approach. Entering multiple pick sets is a good idea — entering three or four next may actually increase your expected ROI, btw — then just keep rolling the dice. 18 cashing spots out of 545 means you need to be top 3.5% or so to cash, so playing 2 pick sets a year, the average person would expect to cash once every 14-15 years. Hopefully using TR, it’s more of a once every 4-5 year thing for you. I actually think our 2010 picks should have done it, wish you had come across TR one year earlier!

    Thanks for being a subscriber.

  • Scott C.

    I used your “Large Confidence Pool Picks”. Within that category I used the “Large Pool” and “Large Pool 2”. Large Pool finished with 306 points and Large Pool 2 finished with 292 points. An “in the money” finish would have required 322 points, so I was not far from the cash. I was curious and figured out the Large Pool Mike would have gotten 292 points if I had played it. Next year I will let you guys know before the contest starts so I can get some advice on fine tuning my picks.

  • bill B

    I do not know if it is an anomaly…but betting on FSU in a bowl game seems to be a good bet…they have covered in 9 straight bowls…

  • David Hess

    I would caution against putting too much stock in something like that, especially as it stretches over two coaches and multiple player career cycles. Even if there *is* something to Florida State’s culture that leads them to do well in bowls … many other people betting their games are probably aware of the streak, and the market ought to correct for it some.

    That said, Florida State was one of our highest confidence ATS picks this year, and they came through. :)

  • Fred kramer

    In your rankings section for each sport you have a – Homefield Advantage number posted right below the gainer / losers for the previous week. Has this Homefield advantage number already been applied to the rankings data or should that be added to any home teams ranking number?


  • David Hess

    Fred — The rankings list teams from best to worst according to a certain rating. The HFA value tells you what the home field advantage used in the calculation was, and what HFA you should apply if you are comparing two teams.

    So, say two teams are meeting at a neutral location. You would ignore the HFA entirely. If they are playing at the home location of TeamA, you would add the HFA value to TeamA’s rating before comparing them.

    If you’re looking at a Power Ratings prediction somewhere on the site, it will already account for the HFA. If you’re just comparing two teams on your own for a hypothetical matchup, you’ll have to add in the HFA yourself.

  • kizifriv

    I feel admiration before this chart!