Reverse-Engineering an NHL GM Search

On July 15, 2026, the Detroit Red Wings announced that Steve Yzerman was stepping down as their GM and that they would begin a search for his replacement. This news brought much speculation about whom the Wings would hire. None of this conjecture has been borne out, since no one has been selected yet. However, reports did emerge on August 5, 2026 that Detroit hired a search firm to assist them. With that type of organization typically holding its cards close to the vest and the team being guarded, no solid leads have emerged. In spite of this, a person can make some sensible guesses. Without direct information, a systematic approach can make reasonable predictions, narrowing a larger field of potential candidates to a shortlist of viable ones.

Foundational Ideas

For this approach to make rational predictions, it must mirror the logic of a consultancy. Those organizations blend the requirements of the team with their criteria. A search firm uses them to generate a shortlist. If a process does not reflect that reasoning, the list it creates will not overlap very much with the one the consultancy builds. Yet, the search firm’s requirements, in this case, are not just generic filters. Detroit’s modernization goal means specific criteria must be taken into account. For a process to make sensible guesses, it must integrate other factors into the standard logic of a consultancy: fluency in technology and data, capability at blending technology, data, and scouting, the ability to build systems, and age.

A search firm includes them, even when a team doesn’t tell it to do so. Consultancies understand industry trends deeply. They are aware of common failure modes, and they have experience in hiring generally. This expertise is used by them to add requirements to the team’s criteria. They will not select any candidate who does not meet that blended set.

Fluency in analytics and technology is included in that collection, since the NHL is trending towards those elements being more central. If the Red Wings hire someone who doesn’t know how to use them, they will fall further behind the most modern teams, making fluency a necessity.

The organizations at the leading edge have more than just knowledge; they actually use technology, data, and scouting as an integrated unit, the way Carolina does. If Detroit doesn’t do the same, it will keep losing ground to them.

Those teams also have systems that integrate those elements. Detroit doesn’t have them and needs to build them to catch up. Constructing them poorly would undercut the whole modernization push.

GMs who can accomplish that goal have a specific age range that must be taken into account, since most of them are 35 to 45 when they are hired (also true in baseball). Very few General Managers are chosen, in all sports, before they turn 35. In the history of the NHL, there are four (examples here and here). Baseball has only three (examples here), while the NFL has two (examples here and here). In the modern hockey, numbers-forward GMs are in their late 30s or early 40s, much younger than the average of 53/54. Consequently, a candidate younger than 35 faces added scrutiny he’s unlikely to pass — the rarity of such hires suggests he needs an exceptional resume to overcome his age. A candidate older than 45 faces a different problem: since modern, analytics-driven GMs cluster around 40, the further a candidate’s age drifts from that center toward the broader GM average of 53/54, the more likely he fits the traditional mold rather than the profile Detroit is looking for.

Process Description and Justification

To create a process that matches what search firms do, two separate filtering steps are required. The complexity of what a consultancy does cannot be implemented in a single stage, without being too strict or lenient. The age window from the previous section applied too rigidly leads to the exclusion of candidates who could be hired, while that range administered too flexibly results in the inability to narrow the list. The need to make such a fine distinction is avoided by using two stages. The first feeds into the second, with the inputs for the former being different from the latter. Those distinct ingredients decorrelate the risk between the two, since each item is re-evaluated using different factors in the second step. With this setup, the first stage can include some low-probability candidates, as long it pushes through most of the high probability ones. The second phase determines which one an applicant is closer to and selects only the candidates whose fit meets a basic standard.

Both steps and the overall process can only hope to get their logic correct, not to capture every possibility. No model can cover every eventuality, but it can be directionally right, according to Aaron Brown’s Red-Blooded Risk. Recommendations can meet a goal more frequently than rejections do, with particular exclusions being superior to certain suggestions in some cases. This method and its steps only try to be directionally correct.

The execution of them focuses on reducing variability in decisions. That element contributes as much to errors in choice as flaws in judgement do, according to Noise by Daniel Kahneman, Oliver Sibony, and Cass Sunstein. Different conclusions can be drawn, at separate instances, by the same person, even when the information at hand is identical.

The steps and procedure are also focused on avoiding overfitting, which occurs when a model over-optimizes for a specific sample. This issue can make a scheme — as Nate Silver’s The Signal and the Noise notes — more correct for that sample and less so outside it.

Part of the reasoning behind this process is to use scores and thresholds. Each step produces a grade that is compared to its threshold. Candidates whose calculations meet or exceed that barrier move on to the next phase (or selection after the final one). Each score is derived differently using distinct factors and reflects the goal of its stage — justifiability for the first one and fitting requirements for the second.

Justification Step

An algorithm generates the score. It delivers the same result from the same information every time, unlike humans, as the work of Daniel Kahneman notes.

The equation encodes the logic of search firms. It weighs general elements as well as intangibles, abilities, and work history to create the score. The basic requirements of the consultancy (e.g., financial feasibility of hiring, thick skin, team success, etc.) are blended, in those categories, with ones specific to Detroit’s situation (e.g., age, mindset oriented towards where the game is headed, capability to build systems, etc.). The algorithm embodies what the foundational ideas section describes.

Each component in an area is evaluated on a 0 to 5 scale, since ranges of 5 to 7 points create a relatively low amount of noise across the same evaluator, as Kahneman’s research shows. All pieces must be on the spectrum or converted to it for the algorithm to combine them.

The equation weighs equally each category and each section’s elements to mitigate the risk of overtraining. This approach eliminates the danger of over-optimizing for a sample, even if one existed, and over-extrapolating from guesses about relative importance, as Noise: A Flaw in Human Judgement argues.

Using equal weights frequently delivers a performance near that of an optimized regression model, as the book Noise asserts, making this approach efficient.

That equally weighted algorithm is compared to a threshold of three, which is above average. The equation’s 0 to 5 scale sets two as below average and three as above, forcing a person to move a rating above or below the mean. A threshold set above that reasonably ensures a candidate is defensible, if the algorithm is roughly right.

Requirements and Fit Step

A person selects the score on a scale of 0 to 5. That grade reflects a mixture of how analytically driven a candidate is and how likely he is to replace Yzerman. A scale can handle that blend, making an algorithm unnecessary, since ranges are quite effective at sorting items on simple factors, as Kahneman’s research shows.

Candidates’ scores are compared to a threshold of 2.5, a grade that means an applicant fits Detroit’s targeted profile. Any rating below doesn’t fit, and any above matches and is likely to be hired. As long as choices fulfil the targeted profile, this step is in the ballpark.

Process Execution

A list of potential candidates is created to act like a search firm’s database. It sticks to people whom a search firm absolutely will have on its evaluation list: those currently employed in an NHL team’s front office at a high level but not as a general manager, such as assistant GMs, directors of personnel, pro scouting, amateur scouting, analytics, and hockey operations. Every individual who currently has one of those roles (or something equivalent or any other high-level position) is added to the list, unless a team’s front office info wasn’t available. The resulting “database” has 135 names.

None of those candidates is run through step two, until every applicant finishes phase one.

The justification stage evaluates each option on every area’s elements, with the exception of age. That aspect is determined by collecting a choice’s date of birth (or the closest approximation available), calculating its age, and converting that number to the 0 to 5 scale. The other components are assessed by Claude who acts as a stand-in for the experts a search firm has. Claude does not have expertise that a person working at a search firm has, but it is capable of being roughly right, when rating on a defined scale. It is given a different 0 to 5 range for every aspect, anchoring its output on that range’s definition. Under these conditions, Claude can approximate experts.

Once those specialists evaluate every candidate, each region’s traits are averaged, and each area’s average is combined to produce a score. That grade is compared to the phase’s threshold. The candidates who meet or exceed that barrier proceed to the requirements and fit stage.

Table 1: Justification Step Score

Table 2: General Score

Table 3: Intangibles Score

Table 4: Abilities Score

Table 5: Resume Score

Table 5: Resume Score

Eighteen of the 135 applicants were pushed through. Claude evaluated them on the second step’s scale, with it being handed a meaning for that range. The resulting scores were compared to the threshold for the requirements and fit phase, with the candidates who met or exceeded it being selected.

Table 6: Requirements and Fit Step Score

Only seven options passed both phases.

Outputs

The candidates who passed both steps are listed below. Each one has a description about his fit.

Darren Yorke: Yorke is 41, is fluent in analytics, is deeply experienced at integration, and was central to constructing the systems that propelled the Hurricanes to a Stanley Cup and four finals in eight years.

Tom Poraszka: With expertise in data and significant experience synthesizing multiple elements, Poraszka built procedures that were vital to Vegas winning a Stanley Cup and making three finals in nine years. He is 39 years old.

Sam Ventura: Buffalo’s methods were created by Ventura from scratch. He was crucial to turning the Sabres fortunes around. He is 39, well versed in numbers, and capable of mixing together numerous pieces.

Sean Tierney: The structures that were essential to Ottawa’s transformation were developed from the ground up by Tierney. He blends several approaches, using his expertise in statistics. Tierney is 41.

Tyler Dellow: Dellow was an integral part of the Hurricanes being one of the most successful teams of the last eight years. Carolina’s processes were not laid out by him, but they were extended. He integrated multiple parts, using his deep understanding of math. He is 47, putting him a bit outside the age window.

Alexandra Mandrycky: Seattle’s systems were created from scratch by Mandrycky, using her mastery of analytics and ability to synthesize. She is 35, putting her on the edge of the age window. Her team, the Kraken, have been more successful than some expansion teams (the 90s Senators were bad), but they have not been as historically good as Vegas.

Arik Parnass: Parnass is 33 and outside the age window, but he is deeply conversant in data, capable of integration, and experienced in building from scratch the methods that do that. His team, Colorado, is a perennial Cup contender, with the components he built being central to that success.

Takeaways

The candidates selected by this approach are justifiable bets, even if none of them is hired. A wager based on a rational process can be incorrect, and one built on a wholly irrational method can be correct. Yet, the former will be, in the aggregate, right more than the latter. Darren Yorke checks off every box the Red Wings’ search is looking for, yet he might not be hired. Given that possibility, Yorke is still a good bet to make, since someone of his archetype is more likely to be hired than baseless guesses are. Flimsy hunches are built on an unsystematic procedure, which only hits its mark when luck blows it in the right direction. Systematic methods are much less dependent on good fortune. For that reason, bets based on them are justifiable. Since the method outlined above is rational, the candidates picked by it, such as Darren Yorke, are defensible wagers, even if none of them hit.