All Pages
NBA Miner | NBA Miner’s Prediction Success
NBA Miner’s NBA Games Prediction Success
prediction, statistics, nba, games, winner, forecasting, teams
1774
single,single-post,postid-1774,single-format-standard,ajax_updown_fade,page_not_loaded,smooth_scroll,boxed,wpb-js-composer js-comp-ver-4.1.2,vc_responsive

NBA Miner’s Prediction Success

On 23th December 2014, we started making daily predictions for winner of each NBA game and we published them on our prediction page.

 

Our winner prediction algorithm is based on multiple statistical forecasting methods whose major parameters are:

  • Teams’ winning percentage
  • Teams’ score margin
  • Teams’ pace
  • Teams’ recent form
  • Teams’ opponents strength in previous games
  • Teams’ fatigue based on a resting day counts and back-to-back game types

 

All these factors are calculated separately for both home and away games and evaluated accordingly.

 

And now, 2014-15 NBA regular season is over, so we are shutting down our prediction page until next NBA regular season begins. But before that, let’s look at how successful we were.

 

We published predictions for 815 NBA games; and 532 of them were correct. Our overall prediction success rate is 65.3%.

 

 

Time Based Results

 

You can see below its distribution according to time: monthly and weekly.

 

1

 

It is pretty normal that by time the predictions become more accurate, because it is a data driven statistical forecasting algorithm and number of data is crucial for its accuracy. Yet, February (week 6 and 8) results are suspicious. It is maybe due to bad luck and some games concluded with big surprise there. But starting with 1st of March, we see that our predictions were highly satisfactory with over 70% prediction success.

 

2

 

3

 

 

Risk Category Based Results

 

While making predictions, we also separated games according to risk level. We divided it 5: low, low to medium, medium, medium to high and high risk categories. Again, for risk level we only used statistics, nor bet rates neither our game difficulty perception. Here are the results according to these risk categories:

 

4

 

5

 

Majority of games (39.0%) are in the low risk category where we predicted correctly with a rate of 76.7%. But higher risk level means lower prediction success rate, thus for medium to high levels the rate drops around 55%.

 

Adding the time factor into the table, our results are:

 

6

 

Especially, 91.3% of correct prediction (42 of 46 games) in April for low risk category games is exceptional.

 

7

 

 

 

Team Based Results

 

Finally let’s look at which teams were more reliable when we had predicted it as the probable winner of that game.

 

8

 

9

 

10

 

56 times, we predicted Golden State Warriors would be the winner and in 44 of them that prediction became true. This means 78.5% success rate and it makes GSW as of our most reliable team. On the other hand, our algorithm never predicted New York Knicks as the probable winner that in fact is enough to prove how reliable are our predictions.

 

 

TAGS

5 Comments

  • Sedat Koç

    I like your site guys and I am aware you are not interested in betting, but as a long time bettor, I respectfully want to say that these prediction stats doesn’t mean anything at all without considering the spreads.

    I took a close look at your archive. I didn’t have much time so I only checked December predictions one by one and I will use that month as an example.

    First of all, there are already risk levels in sports betting and they are called odds, which are calculated with a simple formula from probabilites. I checked all pregame odds of all 66 December predictions.

    If you had bet the same amount of money (let’s say 1 unit) for every single prediction on December, you would have lost 10.85 units. 38 correct predictions would bring you only +18.85 units while 28 wrong predictions would make you lose 28 units. That’s because, the average odd of your 38 correct predictions is only 1.496.

    So your %58 hit rate for December actually means a loss around %16. (return of investment = profit/staked = -10.85/66)

    • Gürhan Günday

      You are right we’re not interested in betting. As we already stated in prediction page, these game predictions do not pursue any bet objective. And we strongly advise to not use these predictions for betting purposes. NBAMiner do not bear any responsibility in any means.

      Still, these predictions can be useful inputs in that area. You said you only checked December games; we started to publish predictions on the last week of December. Since these predictions are based on previous games data, number of data is crucial for its accuracy. For more accurate results, you should look after enough data is accumulated (let’s say after all-star break).

      So, why not doing these same simulation after our March or April data? We guess, it will please you.

  • Mike

    If the predictions are not for betting purposes, then what are they exactly for?

  • Sedat Koç

    Thanks for your response and warning but as I stated, I am a long time bettor. Any experienced bettor wouldn’t bet according to an algorithm he doesn’t know anything about. Besides I haven’t seen a successfull betting strategy so far that includes betting moneyline on favourite teams with tiny odds since it’s not worth the risk.

    I am aware the sample size is not enough for December (I don’t have hours to check every single pregame odds of all your predictions , you can find them from sportstats.com, besides it’s not my website :) ) and you might even make the best predictions of the world in the following months. But that’s not my point. My problem with this article is the way you calculate and rate your prediction success. You simply can’t treat every prediction same, they have different values. That’s why we have spreads and odds in sports. For example, 10 of 38 correct predictions had odds between 1.02 and 1.15 (Roughly, 1.02 means a probability of %98 while 1.15 means a probability of %85). If a team offers around 1.10, it means they are at least 10 points favorites. So predicting a 10 point favorite team to win the game (even with a single point) cannot be treated same as predicting an underdog team with 2.00+ odds to win (naturally, the latter is a much much more valuable prediction) and you can’t rate your success with your hit rate. My modest advice for you to add the pregame odds to your predictions to rate each prediction’s value and have a healthier evaluation of your work.

    In addition, if it has nothing to do with betting, what’s the point of predicting the winner of the games then or what’s the objective of this work? You might not be interested in betting but the most advanced outcome prediction algorithms or simulators are used for betting industry. Only a few bookmakers use their own and the rest of the world just copies their spreads and total lines. That’s why I was intrigued by this piece and your work in the first place.

    The most and the least reliable teams of this year, according to their performance against the spread.
    http://www.covers.com/pageLoader/pageLoader.aspx?page=/data/nba/statistics/2014-2015/ats_regular.html

    I also recommend some online pieces about Haralabos Voulgaris who is a well respected handicapper and spent a lot of time to build a complex prediction model to beat the bookies.

    Have a good day.

    • Gürhan Günday

      We are a statistical based basketball website. We gather data, all kind (team or player) of basketball data about NBA, then we try to use them as much as possible, ty to consolidate them and produce some significant results with them. In this way, we’re publishing daily stats and analytical researches.

      The ultimate aim of a basketball game is to win it. So, we wonder if we get enough/qualified data, is it really enough to predict the winner of the game by just looking them? Our real purpose of developing a prediction algorithm is a challenge to measure the power of basketball metrics. It is simply about estimating the future of a game by using past knowledge.

      For betting issues, we’re aware of betting parameters/strategies and moneyline odds. If it was our point, we would surely use/add them. In fact we made a simulation according to the odds and put 10 units of money for each of our prediction after 1st March. The result is 345 units of profit at the end. And for some of games, our prediction favored the underdog not the bet favorite. Finally remember also we are categorized games by risk levels. Even for current version, we do not consider each game equal. You can see the prediction success rates according to the risk levels as well.

      But again betting is not the concern of this site.
      Nevertheless, this kind of comments make us happy and we give values of our readers’ comments. Next season, we will develop our algorithm further by adding injured/not played players’ effect (what about the prediction if a star player will get DNP before tip-off).
      Thanks.

Post A Comment