Which NHL Teams Are Good At Drafting?

Which NHL Teams Are Good At Drafting?George McKeownBlockedUnblockFollowFollowingFeb 1Lots has been written about the value of draft picks in the NHL.

David Wilson even wrote his Master’s thesis on the topic.

I will quickly summarize the standard findings before getting into how well each team has performed.

One way to measure the value each draft pick is to determine the proportion of players drafted at that position to hit a certain milestone.

David Wilson chose 160 GP because that’s enough to earn a pension, so we will use that.

All that is really required here is an arbitrary metric that qualifies players who “made it”.

The question of how many drafts to use is tricky.

Recent drafts provide only a tiny sample, but historical data may no longer be applicable to the current league.

Using drafts since 1990 should provide enough data for a reliable signal, while being recent enough and having a similar number of teams in the league.

For simplicity, I will refer to players with over 160 career GP as “pension players”.

The 2012 draft is the most recent one that appears to have reached equilibrium, so the data set will be 1990–2012.

The following is two charts which have been made a lot of times before.

We will be using this data to approximate the expected number of pension players that each team should have drafted.

The following players were all selected 156 overall: Connor Brown, Jared Spurgeon, Ryan Reaves, James Wisniewski, Brian Campbell, Patrick Lalime, and Peter BondraSame plot with a log scaleAside from the fact that 15 overall has the same success rate over 22 drafts as 156 overall, the graph does what we expect.

The first few picks are basically guaranteed to be NHLers; this probability drops sharply after pick 5 and plateaus after the third round.

This is fine so far, but number of players with 160 GP doesn’t tell us anything about quality.

For example, if you are given a choice of two players drafted in 2010 and you know nothing else besides draft position and the fact that they have each played at least 160 games, would you rather have the player taken 13th overall or 131st?To answer this question, all we need to do is take the set of pension players by draft position and add a performance metric.

I will be using per-82 game point shares.

It’s a complicated statistic to calculate, but the number of point shares a player has in a season is supposed to approximate how many fewer standings points his team would have had without him.

This stat is more effective than points because forwards and defence of comparable talent generate similar values.

Elite goalies tend to have crazy high point shares, so they are excluded from this next data set.

The table below has some names to provide context of what counts as a good number:Every player with 160 GP is plotted on the graphs below.

Data is clustered to the left because more first round picks make it to 160 GP.

Same plot with log scaleThe PS per 82 games can be averaged for each overall pick to give a cleaner plot.

Because there are much fewer NHLers per pick in the later rounds, overall picks past 30 are grouped in 10s.

Picks from 5 to 30 are grouped in 3s.

Same plot with log scaleThere are a few conclusions we can draw from this:First and second overall picks always become NHLers who are elite on average.

Picks 3–5 are very likely to become NHLers who are on average slightly better than NHLers drafted later.

Beyond 5th overall, there is virtually no correlation between draft position and player quality, given that the player makes it to 160 GP.

Now that we have established both the probability that a player becomes a pension player and the expected player performance given that a player makes it to this threshold, we can evaluate each team’s draft history.

Team Draft PerformanceTo measure team performance by the number of successful draft picks (where success is measured by 1 if the player makes it to 160 GP and 0 if not), we can assign a probability value to each pick using the yellow curve from the second figure.

We then add up all of these probabilities to find the expected number of NHLers that each team should have drafted.

This is the number on the horizontal axis.

The vertical axis is the actual number of pension players drafted, divided by expected.

This is our “quantity” metric.

ARI includes WIN/PHX, CAR includes HFD, COL includes QUE, DAL includes MNS, WPG includes ATL.

Picks where a goalie was selected are omitted, because goalies have a much lower chance of playing 160 GP.

On the far left are the newest expansion teams who have not participated in as many drafts.

Teams farthest to the right have high expected values because they have consistently been awarded top draft picks for finishing poorly in the standings.

Unsurprisingly, EDM has the highest number of expected players drafted.

Once again, this data does not tell us anything about quality.

Ottawa has been by far the best in terms of drafting NHL caliber players, but how do we know they aren’t favouring players with low upside but a better chance of being a regular player?.Looking at this graph, I suspect that TBL and WSH aren’t terrible at drafting, but are more likely to pick risky but high-upside players.

If we once again limit the dataset to players who have made it to 160 GP, and assign “expected PS per 82 games” values using the yellow point shares curve, we can find out if there are differences in quality between each team’s successful draft picks.

This is the result.

The horizontal axis measures how many draft picks play 160 games (divided by the number expected), the vertical axis measures the difference in player quality (minus the expected quality).

Teams with a low proportion of 160 GP players selected and poor 160 GP talent are in the lower left.

Teams that draft a lot of good NHL players are on the top right.

Does any of this matter?It’s one thing to determine that EDM is bad at drafting and OTT is good, but does any of this translate into team success?.We can measure team success by total points percentage (P%) over a given time period.

Because we are measuring the 1990–2012 drafts, and we have determined from the first plot that it takes about 6 years for all drafted NHLers to be full-time players, we will use team wins between the 95–96 and 17–18 seasons.

We can compare both the quantity and quality metrics to P% over these seasons:Quality seems to be much more correlated to wins than quantity.

Finally, we can bring back the chart from earlier but colour it by P%.

Blue is the lowest, red is the highest.

Yellow is higher than green.

While each metric is not correlated perfectly, take a look at the quadrants.

The teams in the top right have done very well over this time period, the bottom left has done much worse.

The top left and bottom right are mostly intermediate, though top left is better than bottom right.

Capitalizing at the draft is integral to team success over the long run.

SummaryDrafting high quality players seems to be more important than maximizing the fraction of drafted players who become NHLers.

This should answer the question of whether the ideal draft strategy is to select for high upside or select for high probability of making it.

The answer is that teams should always aim high and not be afraid of draft busts.

In the next instalment, I will look at which types of players the good and bad teams draft.

For a team-by-team representation, click here.

NotesA more in-depth version of this would be to group drafts by GM, not by team.

I might do this later.

Point shares per 82 games is calculated by dividing total point shares by total games played.

For example, if player 1 has 200 GP with 2 PS and player 2 has 1800 GP with 38 PS, the average per 82 is 82*(2+38)/(1800+200)=1.

64Goalies are excluded from all data because a) they have a much lower rate of reaching 160 GP and b) their point shares are much higher than skaters.

Drafting goalies will be explored separately.

Draft data is taken from Hockey Reference combined with NHL.


Points percentage by team is taken from NHL.


Plots and fits are done with Mathematica 11.


. More details

Leave a Reply