clock menu more-arrow no yes

Filed under:

The (in)exact science of predicting players' careers

New, comments

Has Nate Silver of FiveThirty Eight created the algorithm to end all algorithms?

Jaime Valdez-USA TODAY Sports

Well, has he? Silver purports to have constructed a brand-spanking new model christened CARMELO that predicts (with a confidence interval of 80%) an NBA player's WAR in the coming years. It's worth your time to follow that link over to FiveThirtyEight, and spend the next ten or so minutes perusing Silver's breakdown of his methodologies. The watered down, fast-pace (consumer friendly!) version:

  1. Start with a big pot. Heat to 350 degrees fahrenheit. Throw in the vitals: height, weight, draft position. Add in the seasoning: TS%, USG%, FTr, AST%, TO%, TRB%, BLK%, STL%, and a viscous mixture of BPM and RPM. Bring to a simmer.
  2. Pour that pudding of statistical and biographical information into a pan and bake for three hours. Once cooked, compare your resulting player information pie with historical players' information pies. The similarities and differences between these two are indexed as similarity scores. A positive similarity score is more similar than not (100 being perfectly matched), a 0 indicates an equilibrium of similarities and differences, and a negative score signifies more differences than similarities.
  3. Abandoning our pie analogy, and forging forward with actual pertinent information, the CARMELO algorithm collects the 10 most similar seasons for this hypothetical guinea pig player. From there, the algorithm dissects each of the 10 players' 7 seasons subsequent to the season tabulated in the similarity scores, and averages out all 10 players' WARs each year.
    Each average is weighted, however, towards more similar players. For instance, a player with a similarity score of 60 will have double the effect on the WAR average than a player with a similarity score of 30.
  4. Finally, CARMELO spits out a pretty scatter plot, complete with a range of 80% of the probable WAR totals each of the next 7 years for the given player.
...So, all in all, what has Silver created for us? Is it flawed? Of course. Everything is. In particular, I cringed when he threw in draft position as a factor that weighed into similarity scores. I mean, it's by no means a deal breaker--the myriad other factors will override the disparity in Draymond Green's CARMELO. It just seemed like an idiosyncratic tweak in a formula that otherwise looks at the most stripped-down, raw statistical impact of a player.

As CARMELO ultimately is just a WAR projection system, a lot of its value comes from how much you value Wins Above Replacement. There are holes and errors inherent in all metrics, so this is far from a be-all, end-all prediction system.

Holistically? Well, it's a nice continuation of Basketball-Reference's similarity score indexes--now, instead of just citing noted players with similarity scores in the approximate vicinity of Harrison Barnes, we can cite what we can expect from Barnes due to the historical similarity between them (by the way, CARMELO says: league average starter).

Ultimately, it's a bit of diminishing returns, though. Silver ends his introduction of the new algorithm by stating it has approximately 10% more predictive accuracy than FiveThirtyEight's purely-RPM based predictions from a year ago. 10 extra steps and 5 extra pages of formula to shave mere ticks off the margin of error. It's not a bad thing, of course--but as humankind progressively invests more and more brain power into such frivolities like basketball analytics, we will inevitably approach our limiting quotient.

As X approaches infinity, it gets closer and closer to 1--but it never reaches it. Achilles and the tortoise. There isn't a mathematical algorithm out there that will predict the future; even something as simple as a basketball player's WAR for a finite number of seasons.