Every Friday, we will now be reporting our algorithm’s NBA win/loss outcome predictions for the following week. These can be accessed by clicking the “Weekly NBA predictions” link in the header section of our blog. This week, we make predictions for a total of 51 games, training both on last year’s results, as well as the 70 games that have already been played this season.
We’ve spent the last couple of evenings training some preliminary algorithms on the NBA 2013-14, regular season data, which we grabbed from basketball-reference.com. Each of the 30 NBA teams play 82 times a season, summing to 1230 games total — a sizable number that we can comfortably attempt to model. Here, we cover our first pass at the prediction problem, what we’ve learned so far, and challenges we’re looking forward to tackling soon. (more…)
Tonight is the opening night of the 2014-15 NBA season. This year, we will be running a machine learning algorithm aimed at estimating underlying features characterizing each team. With these features, we hope to identify interesting match-ups (including potential upsets), similar team-playing-style categories, and win-loss probabilities for future games. As of now, the only source data that we intend to feed our system will be win-loss results of completed games. As the season progresses, our algorithm will thus have more and more data informing it — It will be interesting to see if it can begin to provide accurate predictions by the end of the season. Stay tuned for periodic updates on this experiment!