2 Question
2.1 Approach
Depending on the context, there can be countless approaches one could take to define the relative value of each selection of an NHL Entry Draft. As one example, one could look at prior trades and then define the value of a pick to be what it has historically been traded for (the “market rate”), as was done by Isacke (2022) and Tulsky (2013). For example, this approach might say that pick \(i\) is typically traded for pick \(j\) and pick \(k\). This means that. This is not the approach we will take.
In contrast, this report will take the approach of defining the value of selection \(i\) to be the typical career contribution of players drafted at \(i^{\text{th}}\) overall (we will define what “contribution” means when we pick a model). This matches the definition chosen by Moreau, Perera, and Swartz (2025) (NHL), Cacchione (2018) (MLB), and Liner (2020) (NBA) and is preferred in this context because we wish to find opportunities for general managers to make trades that help their team make trades which allow them to draft higher quality players. Using this definition of value will help NHL General Managers make data-backed decisions as opposed to using past trades as a starting point in trade negotiations. This is preferable since previous trades could be based on what “feels” favourable instead of what is favourable. Furthermore, in context of trades which only include draft picks, one team usually gets the better end of the trade. In other words, basing trades off what “feels” favourable doesn’t make a whole lot of sense because draft pick exchanges are typically only favourable to one team, not both.
This report will estimate \(v_i\) by fitting several regression models incorporating player contribution data by the players previously selected at pick \(i\). We will consider GP and PS as our metrics for measuring player quality (and thus the quality of a pick), and for most of them we will fit a linear and non-linear regression model.
The first two methods of evaluating player quality are similar to approaches taken by Moreau, Perera, and Swartz (2025), Cacchione (2018), and Liner (2020), and will take the PS (or GP) of all players taken at pick \(i\). Next, will consider a similar metric which adjusts the PS values for active players based on what we estimate their PS will be at the time of their retirement. The final approach is from Luo (2024), and will define an indicator random variable \(Z_{ij} = 1\) if and only if the player selected at pick \(i\) of draft \(j\) became, or is on track to become, an NHL regular. We will then fit a logistic regression model, giving us an estimate of the probability that a player selected at pick \(i\) will become an NHL regular. Thus in total we have four measures of the quality which will be considered: PS, GP, adjusted PS, and probability of becoming an NHL regular.
2.3 Considerations
We will fit 7 models in the Model chapter, so we need some criteria for what a feasible model looks like. We define a feasible model to be a model which satisfies both of the following:
\(v_i > v_{i+k}\) for all \(i, k \in \mathbb Z^+\)
- This ensures later draft picks are not considered more valuable than picks earlier in the draft. This should intuitively make sense because the players available at pick \(i+k\) are a proper subset of the players available at pick \(i\), so there is no reason for a later pick to be more valuable in a trade context than an earlier pick.
\(v_i > 0\) for all \(i \in \mathbb Z^+\)
- Every pick has a positive value since there is no real negative impact to picking a bad player, other than the opportunity cost of the “wasted” draft pick. Even if a player plays in NHL games and objectively contributes negatively to their team, this report takes the stance that the player has zero value instead of a negative value.
While we will not include it in the definition of a feasible model, it is common knowledge in ice hockey circles (and confirmed by the previous work listed below) that NHL draft picks do not decrease in value linearly. In particular, \(v_i\) decreases quickly in \(i\), so the difference in value pick 1 and 30 is much greater than between pick 101 and 130 (ie \(v_1 - v_{30} >>> v_{101} - v_{130}\)). In the Model chapter we will fit this model and show that it is not appropriate, before quickly moving on to non-linear models.
Note that if picks did decrease linearly in value linearly then it would be very easy to create a model of draft pick value since we would have
\[ v_1 = v_2 + c = v_3 + 2c = ... = v_{224} + 223c \]
where \(c > 0\), meaning we would only have to find the value of \(c\).
2.4 Previous Work
Some work in this area has been done before, such as:
Valuation of NHL Draft Picks using Functional Data Analysis (Moreau, Perera, and Swartz (2025))
MLB Rule IV Draft: Valuing Draft Pick Slots (Cacchione (2018))
Determining the Value of NBA Draft Picks using Advanced Statistics (Liner (2020))
Examining the value of NHL Draft picks (Isacke (2022))
NHL draft: What does it cost to trade up? (Tulsky (2013))
This report will most closely follow the work done by the first three papers listed. As an interesting aside, Eric Tulsky, who wrote the last article listed above in 2013, was hired as General Manager of the Carolina Hurricanes in 2024.