Trading Strategies for the NHL Transaction Network
This is a conclusion of research conducted over the summer where I examine the National Hockey League as a transaction network and try to identify successful trading strategies.
Parts 1–3 can be read on my personal page.
The National Hockey League as a Transaction Network: Part 1
The National Hockey League as a Transaction Network: Part 2
The National Hockey League as a Transaction Network: Part 3
Those articles construct the league as a transaction network where each node is a team and each transaction between two teams is an edge in the network. I made a small interactive version of the networks for the first ten seasons, here. A screenshot is below.
Unfortunately, there were not many patterns to be found among the different ways to measure networks. Some of these measurements include betweenness centrality, PageRank, and hub/authority values. There were some relationships to success but always through an intermediary measurement such as average age or power play percentage. I did not find any direct relationships between network measurements and points percentage, rank, or championships. A correlation matrix is below. The network measurements are in the center of the axes and you can see how they mostly just correlate with each other. In particular they correlate very poorly with modern sports statistics like Fenwick and Corsi percentages.
The strongest direct relationship was between PageRank and league rank at the end of the regular season. There’s a scatter plot below where you can see this relationship. You can see a slight positive relationship between PageRank and rank. Because it’s a ranking in the league, it’s actually better to be lower. So what this chart is saying is that lower PageRanks are better. PageRank is a number that describes a nodes role in the network. If you think of a system of roads as a network, very busy intersections will have high PageRanks and quiet intersections will have low PageRanks. This is maybe a counterintuitive explanation given that many teams hoping to make a run at the Stanley Cup will participate very actively in the transaction network right before the trade deadline. Their idea is to trade away prospects in exchange for developed players who can contribute to their effort immediately. The conclusion here might be that it’s better to stay quiet or at least minimize the number of transactions while maximizing their impact on your roster.
The most interesting result came from examining different trading strategies. I came up with 12 different possible strategies using the geographic strategies, “National”, “Regional”, “Local”, or “Balanced” and the trading strategies, “Buyer”, “Seller”, or “Balanced”. A regional trader is a team that prefers to trade outside their division but within their conference. A national trader is a team that prefers to trade with the opposing conference. A local trader is one that prefers to trade within their own division.
Below is an example of a radar chart of the New York Rangers in the 2018–2019 season. In this season, they were tagged as having the strategy “National Buyer”.
Below is a swarmplot that plots each strategy against regular season points percentage. You can see that National Buyer and National Seller are the most common strategies. It’s not surprising that few teams prefer to trade within their own division. I determined the geographic preference by finding the number of each kind of trade (same division, same conference different division, and different conference) and calculating the standard deviation of the set. If the standard deviation was less than 1.5, they are balanced. Otherwise, their preference is for the maximum of the set.
The most interesting observation is found in the visualization below. Here, I calculate the number of expected championships if the strategies of championship teams mirrored the proportions of the league and compare that to the actual number of championship teams using each strategy.
You can see that some strategies mirror the league perfectly, like National Balanced. Other strategies, however, show a discrepancy. The biggest discrepancies are with National Buyer and National Seller. These are the two most common strategies in the league. You can see that National Seller is outperforming and National Buyer is underperforming. I think that this is result is very counterintuitive and might be worth considering further. Unbiased Seller outperforms as well. So the narrative that often surrounds the trade deadline. Cup contenders will often try to “buy”talent to make their run.
This result could possibly be explained by the fact that Cup contenders loading up on talent will often be trading away more assets than they receive which would categorize them as a Seller, not a Buyer. For example, when Ray Bourque was traded to Colorado, the Bruins sent two assets and received four. In this transaction, Colorado is the seller and Boston is the buyer. Though this is often not how we think of these types of teams. A team that is rebuilding will “sell” it’s talented players in exchange for prospects or draft picks. Even though they are a net receiver of assets, they will be said to be selling at the deadline.
Regardless of what confounding variables there might be, it’s clear that all strategies are not equal. I’m still not convinced that there’s no quantifiable relationships between network characteristics and success. Perhaps I need to be more granular and examine individual match ups and the histories of players on each teams.
