How Many Finals MVPs Would Bill Russell Have Won?
Five.
Tip of the hat to basketball-reference.com, an indispensible resource for all things basketball statistics.
If we want to discuss who should have won an award, we’re discussing how best to measure the player contributions that award is meant to recognize. That would be fun too, but it’s not what we’re doing here. Instead, we’re looking at the Finals from 1969 to 1979 (the year before the three point line was introduced) and discussing how best to model the voting results for Finals Most Valuable Player, since we know which players got them in all of those years. Once we decide on a model we can apply it to the Finals from 1957 to 1968 and find out how many Bill Russell would have won.
While turnovers, blocks, steals, and offensive and defensive rebounds are available in part of our testing span, they are not available throughout the span, so let’s ignore them and hope they go away. Don’t look. Are they looking? Don’t look! This leaves us with games, minutes, field goals made and attempted, free throws made and attempted, rebounds, assists, fouls, and points from the box score. We’ll also include age, winning the regular season MVP (keeping in mind that players voted on MVP in all seasons prior to 1981), making All-NBA or All-Star, and of course the player’s team winning the Finals. The resulting model looks like this…
…where each orange dot represents a player-season, plotted at one if the player won the Finals MVP and plotted at zero if they did not. Eagle eyed readers will notice that while there are the correct number of winners (eleven) there are not nearly enough losers present. Quite right! The graph extends out of screen to the left like (to use a technical term) a really long extending thing: according to the model none of the 194 players not pictured had even a one in three hundred chance of winning Finals MVP. Absolute worst goes to 1975 Washington Bullet Truck Robinson, whom the model gives slightly more than a one in one thousand million trillion chance.
The blue line is the modeled percentage chance, so for any given orange dot take a line straight up or down until it hits the curve, and that is how likely the model thinks that player was to win in their year. Overall the model’s highest probability in any given year actually won the MVP nine of the eleven years. The only misses are 1969 Jerry West (narrowly behind John Havlicek) and 1979 Dennis Johnson (quite a bit behind teammates John Johnson and Gus Williams). Although there are other losing seasons ahead of winning seasons, such as 1978 Bob Dandridge ahead of 1969 Jerry West, these aren’t misses since Jerry West did not receive the 1978 Finals MVP. That went to Wes Unseld, whose modeled 46% chance was higher than Dandridge’s 44%.
.
But how did I get there?
My first choice was to remove games from the analysis because of the 1970 Finals, where Willis Reed played only six of the seven games and was awarded the Finals MVP. I neither condone nor condemn this decision, I just want to avoid teaching the model to reward a player for playing fewer games in general. For example, in 1964 Johnny McCarthy played one of five games for the eventual champion Celtics — a model that penalized games played would see him as a relatively strong choice that we know he was not.
My second choice was how specifically to use the box score inputs. Is it best to use raw figures? per game? per 36 minutes played? As it turns out the best match is a player’s share of the Finals total — for example, both 1979 Gus Williams and 1978 Bob Dandridge scored 143 points in their respective Finals, but the model would take as inputs figures of 143÷986 = .145 and 143÷1428 = .100 respectively. I’m not suggesting Bob Ryan and Frank Deford were out there doing long division before handing in their MVP ballots, just that this information was in their human brains somewhere and turns out to best match their eventual decisions.
My third choice was the use of field goal percentage and free throw percentage. Our friend Johnny McCarthy happened to go precisely one for one from the field in his only game, and thus shot 100% from the field. Again we know that no one would use this as an argument in his favor unless he did so over an arbitrarily high number of attempts. We also happen to know that the NBA has always had eligibility requirements for its rate statistic leaderboards, so I took a rough average and ended up with requirements of 300 field goal attempts and 125 free throws made in an 82 game season, prorated to the number of games actually played in each Finals.
I made these choices because it best matched our testing set. There are many, many, many other possible choices but they frequently result in adding misses, usually Willis Reed in 1973 (but intriguingly not 1970) and Wes Unseld in 1978.
Using the truly beautiful software package r each parameter is assigned a coefficient using logistic regression, the values are summed, and the result put into an equation of the form…
…which is to say that as x increases the result gets closer and closer to 1 without ever exceeding it, which is to say a 100% chance of winning the Finals MVP and no more.
.
Now that we have our model, we plug in all the box scores from 1957–1968 and get the results that started this post:
Overall I’d say these jive pretty well with our preconceptions. The three most striking are surely 1964, 1961, and 1958.
Sam Jones was never All-NBA First Team (although there’s no shame finishing behind Jerry West and Oscar Robertson) and was still a year away from his first Second Team in 1964, but leading the winning team in scoring while shooting 55% from the field in a Finals where everyone else averaged a stiff 40% (including Guy Rodgers’ unconscionable 11 points per game on 26%) is clearly very appealing to the model. Given that San Francisco’s Wilt Chamberlain scored 29 per game on his season average of 52% from the field it’s hard to insist upon Bill Russell’s unmeasured defense being enough to overwhelm Jones’ measured advantage.
While 1961 seems very late for Bob Cousy to still be getting awards over teammate Bill Russell it was Cousy’s last year as All-NBA First Team, and the model does give Russell a 45% chance so it’s not a “slam” dunk. Unlike 1964 in this case it is plausible that Russell’s dominant defense could have tipped the scales in the mind of actual voters when the two were statistically so close, but plausibility is a can of worms we should avoid where possible.
Like the other seven extant original NBA teams the Hawks have won a title, and managed theirs in dramatic fashion behind the heroics of Original MVP Bob Pettit, who scored 48 points and 16 of St. Louis’ last 19 in the deciding game six before tipping in a Slater Martin miss to put the Hawks up three with seconds remaining. It therefore comes as quite a surprise that the model not only favors Cliff Hagan but resoundingly so. Could there be an underlying error?
In life we talk almost universally about end of series moments — Michael Jordan’s score strip score, LeBron James’ block, and so on. But what about LeBron making both free throws with seven seconds left to seal game two of the 2012 Finals? Does it matter that Michael Jordan scored 38 points on 59 true shooting percentage in a two point win over the Jazz when it happened in an indecisive game five? To a model which looks only at total stats, a point is a point is a point, whether it comes in game three garbage time or game seven crunch time. Let’s take a peek at the total stats leading into game six of our 1958 Finals:
Yikes! It’s easy to see how even Pettit’s game six heroics (plus Hagan playing poorly and fouling out) weren’t enough to change the model’s mind although they almost certainly would a human’s. If the reader would like to give this one to the Bombardier from Baton Rouge that’s alright, because while Pettit is somewhat underrated in discussions of all time greats the main thrust of this exercise was to determine how many Finals MVPs Bill Russell would have won.
The answer of five is particularly compelling since only Michael Jordan currently has more than three. Throw in Russell’s five regular season MVPs and it’s hard to make an argument for anyone else having the second greatest personal trophy haul in NBA history, but that’s a topic for another day.
.
.
If you’re a real glutton for punishment, it’s also worth considering the statistical significance of each parameter in our model. One way of looking at this is with a z score, or dividing each parameter’s standard deviation into itself. A z score of two or higher is the commonly accepted standard, but let’s go one at a time. Our parameters look like this…
…so our least likely to be significant is minutes played with a paltry .1 z score. Continuing one by one our succeeding models look like…
…at which point we have removed our last parameter with z score lower than one (field goal attempts), and all survivors have a z score of at least 1.276. This model still correctly predicts the same nine of eleven Finals MVPs in the 1969–1979 span but does a slightly worse job with Dennis Johnson, now finding him narrowly behind Jack Sikma and Fred Brown in addition to the two previous names. For the 1957–1968 span we get:
The only outright result to change is 1961 (narrowly) going to Bill Russell instead of Bob Cousy. The model becomes somewhat less confident in predictions in general, most notably with 1959’s Bob Cousy dropping below a majority, but Russell fans shouldn’t get too excited since the second highest probability that year goes to Thomas William Heinsohn’s 37%. Since the models are so close to each other it’s probably not a good use of time to argue over which is superior. I chose the first model because it did the best job at predicting the 1969–1979 span, but I don’t begrudge you preferring the second.
.
Just for fun (this is what humans call fun, isn’t it?) let’s take our model forward in time and space and see how well it does.
That’s a brisk 43% modeled correctly with a lot of glaring mistakes. If we instead use our second model, however…
…we’re up to a real nice 67%, not bad considering the model has no idea what a three point shot is. Of course, neither did most players in the 1900s, so it’s no surprise that both models are much more accurate through Jordan (63% and 79% respectively) and fall off a cliff in the new millennium (25% and 55%).
I’m really not sure why the models hate Shaq so much, his scoring was efficient overall and he was a dominant rebounder in the 00–02 Finals. My feeling is the model overrates the increased efficiency of players who take a lot of threes, hence the proliferation of names like Curry, Battier, Allen, Horry, Miller. Throw in Ron Harper leading the 2000 Lakers in assists and you can kind of see why the model likes him, but when 38 and 17 a game on 61 FG% not only doesn’t win but is a distant fourth(!) we should pretty clearly not apply the model to contemporary basketball without a serious overhaul, as expected.
.
.
That’s it! Thanks for reading. :)
