The long awaited sequel to How to Train Your Player (viewtopic.php?f=7&t=337) is finally here! For those of you that weren’t around last time, back in 2015 (IRL) I recorded every GMs training camp inputs (there were no nodes in those days) and performed statistical analysis on the training camp outcomes of every player in the league.
This time I’m did things a little differently. One problem with scientific studies is that it is not possible to observe exact counterfactuals in real life (e.g. if a patient receives the placebo pill it is impossible to see if their health would have been better or worse if they would have received the real medication or if Duke plays zone against Carolina one game, it is impossible to go back and see exactly what would have happened if they would have played man). In How to Train Your Player 1 (HTTYP1), we had a nice sample size because it was the whole league but the fact is comparing Kevin Garnett trained with recommended settings to LaMarcus Aldridge trained with all 4’s and a 5 will never be the same as comparing recommended KG to alternate universe 4’s and 5 KG. A useful tool for achieve these theoretical comparisons is simulation (hey, that’s convenient).
Methodology
The beauty of simulation is that you can get an idea of what the world would look like under various different scenarios. I simulated training camp 50 times using 5 different training schemes and recorded the results for 5 of my players. Why 5? Because I manually recorded them so it took a long time. Also because this started out as simply an exercise to try to figure out how to train these 5 players. The five players are Andrew Bogut, Ed Davis, Kyle Singler, Jayson Tatum, and Lonzo Ball. The 5 training schemes are all zeros, 3’s for all skill attributes and a 10 for athleticism attributes, 4’s for all skill attributes and a 5 for athleticism attributes, recommended, and my own special sauce. In HTTYP1 I found evidence that 5 was the optimal athleticism training level for players under 31 so all 4’s and a 5 is a common scheme I’ve used over the years for players that had a broad array of areas with potential. I also found evidence that players over 30 should have their athleticism trained at a 10 so all 3’s and a 10 is a common scheme I’ve used over the years for players on the wrong side of 30. All 0’s serves as an interesting control and a follow up from a crackpot PM conversation I once had with 42PhD. Recommended is the ultimate control group because that is probably the most commonly used training scheme. My special sauce really isn’t anything too special. I start with a 5 in athleticism training. I then add 5’s to the areas where the player has the most room for growth or areas where I really want them to develop. They get 3’s for the areas I care least about and 4’s for everything in between.
General Observations
You can review the complete data here (https://docs.google.com/spreadsheets/d/ ... sp=sharing). There are some interesting observations that aren’t going to shine through the statistics so I want to start with those.
1. Kyle Singler dies. Every. Single. Time. There was nothing I could do to save him. He was a fringe green player because of his terrible white guy athleticism and the game was going to destroy him no matter what. He actually did best with all zeros. This guy is weird.
2. All 3’s and a 10 (henceforth 310) produced “leaps” the most often. I didn’t track potential in my study (you’d probably need a very large sample size to get enough potential boosts to make it valuable) but I noticed things. Lonzo went green/green twice under this scheme and Tatum got jumped to B level 3 point shooting, offensive rebounding, and post defense twice using this scheme. To be fair though, I used this scheme for our league training camp for both of them and Tatum had to use his insurance to be saved :’(
3. Bogut was a lock for staying green/green with 310. He only dropped to y/y once in 10 tries and even then his decreases wasn’t even as big as some of the simulations in which he stayed g/g.
4. Most attribute outcomes do not vary at all with the scheme. Boguts rebounding attributes dropped to 66 every single simulation. Singler’s 3 point shooting dropped to 82 every time. Tatum’s jump never changed from 98. Even among those that did change, most didn’t change that much. Tatum’s jump shot increased to between 41 and 44. That is a pretty small range. This is good news for all the fatalists in the league that think it doesn’t really matter what they do for training. The TC gods will do what the TC gods will do. To a large extent this is true. The game has already decided what is going to happen to most attributes of most players, regardless of the randomization that pans out when Wig hits the sim button. On the other hand, the rest of this article will try to convince you that what training scheme you choose actually matters.
Overall Comparisons
We’ll start with just comparing averages across the different training schemes. The way these graphs work is they compare each training scheme to recommended. The dot represents the average score. The line is the confidence interval. If the confidence interval overlaps with the red line at zero, then the average for that training schedule is not statistically different from the average of the recommended results (i.e. the averages are different but they’re not different enough that we are confident that what we’re seeing isn’t just a random result). If the line doesn’t overlap, we are free to interpret that average difference as a real difference. Another way of looking at is that 95 times out of 100 we’re going to get a result somewhere in this range.Using the simple differences, there is no difference.
This points out the problem I started out talking about. It really isn’t fair to compare Tatum’s average results to Bogut’s. The following graph controls for that so we are only comparing Tatum to Tatum and so forth.
Here we see that not training players is bad. This is a great sanity check. This suggest that our sample size is big enough that we can get meaningful results. At the same time these seems to suggest that what training scheme you use doesn’t really matter. Do thoroughly answer that question, we need to use a valuable piece of knowledge from HTTYP1.
Split Group Analysis
In HTTYP1 we learned that players older than 30 respond to training camp very differently from players younger than 30. So we’ll run our analyses on a group made up of Davis, Bogut, and Singler and a group of Ball and Tatum. We’ll still control for player so don’t worry about that.
Young players
Here we see that on average, 310 did about 10 points better than recommended (statistically significant) and that, while no other scheme did significantly better, 45 and my special sauce also had better averages than recommended. This suggests that using recommended on young players with lots of potential is probably suboptimal. Now we’ll break it down by skills and athleticism.
Skills
Athleticism
From these we can see that the difference in total means is mostly driven by differences in athleticism as both 45 and 310 outperform recommended. This goes back to my observation that there is very little variation in most of the skill outcomes from sim to sim and from scheme to scheme. However, the recommended schemes for Tatum and Lonzo recommend conditioning scores of 10 each so why is 310 so much better?
Old players
Here we see that 45 is bad. This isn’t particularly surprising considering this is a scheme recommended for young players from HTTYP1.
Skills
Athleticism
Here again we see that the difference is driven by athleticism, which is where the action is in general. The weird part about these results is evident from looking at the recommended settings. Bogut’s recommended conditioning setting is 1, Singler’s is 5, and Davis’s is 6 and yet the results are not different than 310. This doesn’t make any sense to me. The overall message here is that recommended will work as well as anything else for old guys.
Takeaways
My takeaways from this are as follow:
Many training results are predetermined by the game and cannot be altered
It appears that recommended is bad for young players but it isn’t evident to me why
It appears that recommended is as good as anything else when it comes to old players
