Predicting Method of Victory and Round of Finish

As a logical next step in our fight prediction model, we decided to attempt to predict when and how fights will finish. While predicting a fight outcome can be very profitable, being able to accurately predict the round and method of victory would essentially be like a license to print money. Obviously, achieving this model would be incredibly difficult, and would probably require years and years of work and some luck along the way. We are here to share our preliminary results when attempting to predict the rounds in which fights end and how fights tend to finish.

To start off with, Reed Kuhn over at Fightnomics provides an excellent analysis of how fights tend to finish. The list of fights we use to generate our predictions is much smaller, and only encompasses fighters who have at least 3 fights worth of stats recorded in the fightmetric database.

Division Avg Rnd KO (%) Sub (%) Dec (%)
Heavyweight 1.96 66 15 18
Light Heavy 2.33 29 24 47
Middleweight 2.50 34 3 59
Welterweight 2.25 41 13 46
Lightweight 2.56 41 13 46
Featherweight 2.32 26 17 56
Bantamweight 2.32 26 19 55
Flyweight 2.58 26 32 42

We should note before expanding on this data that we have very few flyweight fights in our database. Therefore, even though the numbers are pretty much in line with the weight classes slightly above, we will not yet put very much stake into the output for that specific division. The division we have the most recorded fights for is the lightweight division with 44 fights. The first thing to notice, and it is probably expected, given the weight cap, is that heavyweight fights finish a third of a round faster than the next fastest division. In addition, an astounding 2/3 heavyweight fight finishes with a referee protecting one of the competitors (TKO/KO). Surprisingly, the only other division with more than 40% of fights ending via TKO/KO is the welterweight division. This may just be an anomaly over the last couple years, but memorable knockouts by Johny Hendricks, Jake Ellenberger, Tyron Woodley and newcomer Robbie Lawler all help to make the last couple years especially knockout-filled for the welterweight division.

How can we use this data to our advantage?

The obvious answer is to look for lines that seem advantageous based on the data above. A great example of this, is featherweight champ, Jose Aldo taking on Chan Sung Jung. Our model has Aldo as a very heavy favourite, as most people do according to the odds. The one thing that stands out about the odds though, is Aldo sits at 4.47 to 1 to win by decision. Less than half of the featherweight fights we have observed have finished inside the distance and Aldo has only been able to finish 2 of his last 5 opponents. We see some definite value at 4.47 to 1 in betting on Aldo to win via decision. An example of a well adjusted line would be Johnson vs Moraga going the distance sitting at 1.48 to 1. Flyweights generally lack the power to knock each other out and submissions tend to happen less frequently than knockouts (by extrapolating from other divisions), so if anything the line may even be a little low.

Modelling Finish Round

Attempting to model fight lengths has been a frustrating experience. The output of our model is almost always under 2 rounds. It does seem however, there is a correlation between very low round estimates and actual fight finishes. For example, Herman vs Gonzaga was predicted to last less than a single round, which is exactly what happened. Unfortunately, even the aforementioned Aldo vs Jung fight is expected to last half a round as well. At this point, it is a struggle to find fights that are expected to last more than one full round. Modelling round outcomes will continue to be a work in progress as we move forward and collect more data.

Modelling TKO/KO Outcomes

Predicting fights that end in TKO/KO seems to be a bit more reliable. For example, Gonzaga vs Herman at UFC 162 was predicted to end in TKO/KO 62% of the time. Pierce vs Mitchell at the same event was predicted to have someone lying on the canvas 46% of the time. Kennedy vs Gracie was expected to end in TKO/KO just 20% of the time. The preliminary results of attempting to predict fights that end via KO or TKO seems promising. When Rory MacDonald takes on Jake Ellenberger, we project the fight to end in TKO/KO just 22% of the time. Machida vs Davis at UFC 163, on the other hand is expected to end in TKO/KO over 40% of the time.

Conclusion

There seems to be little evidence that we can accurately predict the round that fights end in so far. It will probably take a lot more work and a lot more data to accurately be able to come up with the rounds in which fights end. TKO/KO finishes on the other hand seem a bit easier to predict. The difficult part will be finding patterns in the data that are not easily seen with a brief look at the fighters stats. For example, predicting an Anderson Silva TKO is not nearly as impressive as would be predicting a Frankie Edgar TKO/KO. From now on, we will include our method of victory along with our fight outcome predictions in our blog posts. Stay tuned, UFC on Fox 8 and UFC 163 are both happening in the next couple weeks, leading into an incredibly busy month of August for The Fight Predictor.

Don’t forget to like us on Facebook and follow us on Twitter

Fight Predictor Model Update, May, 2013

In terms of model accuracy, April was a decent month for The Fight Predictor. Over the 22 fights predicted, our prediction record was 13-9 and we made a profit at 3/4 events. Unfortunately, UFC on Fuel TV: Mousasi vs Latifi went terribly as we were wrong on all 4 fights and all 4 bets. We spent the rest of the month trying to scrape back the money we had lost, but came out just negative on the month. 13 out of 22 fights won puts us at an accuracy of around 60%. Our tests on historical data show we should be hitting close to 80% accuracy. Therefore, since we have added significant amounts of data since our last model update (Prior to UFC 157 in February) we decided it’s time for another update.

One thing that we felt was important to consider is the differences between weight classes. For example, we have seen that takedown accuracy is an extremely important variable in determining who is going to win a fight. By splitting it up into weight classes and holding all other variables constant, our results show that in heavier weight classes (above lightweight), each 1% increase in takedown accuracy increases a fighters chance of winning by 0.1%. In the lighter weight classes, each 1% increase in takedown accuracy increases a fighters chance of winning by 0.4%. Our analysis shows other differences between the lighter and heavier weight classes as well as specific relationships in certain weight classes. For example, both heavyweight and middleweight have unique height explanatory variables.

Another couple variables that have become significant in this model, but weren’t before are “is the fighter a southpaw?” and “is the fighter fighting in their home country?” What we see from these variables is that a left handed fighter fighting a right handed fighter is 5% more likely to win. Fighting in your home country, on the other hand, actually makes a fighter 8% less likely to win. This is likely due to the complacency that comes with being around friends and family. Frank Mir spoke before his fight vs Daniel Cormier about how leaving home really allowed to train to what he felt was his maximum potential.

We tested our model by going back in time and removing observations. Then we recalculate the model to see whether its predictions are accurate to what actually happened in the fight. Over 10,000 observations, here is our output:

Total Correctness: 78.71%

Heavyweight: 85%
Light-Heavyweight: 94%
Middleweight: 83%
Welterweight: 76%
Lightweight: 60%
Featherweight: 84%
Bantamweight: 76%
Flyweight: 100%

This can be compared to our calculations in our article on parlays back in February. It is worth noting however, that even though our accuracy dropped slightly overall, the additional fights in our data should help with our future accuracy.

It is also interesting to note that in our Georges St Pierre vs Anderson Silva super fight article the model had Georges St Pierre as the heavy favourite to defeat “The Spider.” With our new model however, the output is 50.07% for St Pierre. That means this fight is basically as close as it could possibly get and adds more intrigue to the potential match-up.

Don’t forget to like us on Facebook and follow us on Twitter

Next event is UFC on FX: Belfort vs Rockhold coming up in 3 weeks on May 18. We will have our early picks for that event as well as our prediction of the main event of UFC 160: Silva vs Velasquez, coming in the next week or so.