How We Built an NFL Draft Model for the Minnesota Vikings in 2015

The story of how a scouting intern, a data scientist and a draft board full of blue stickers helped bring machine learning into an NFL war room

By Mike Band and Jared Lander

In December of 2014, an email started making its way through the Minnesota Vikings organization.

It began with Jared. He had seen the Vikings owner, Mark Wilf, speak at an event in New York, introduced himself afterward, got his contact information, and followed up. In that email, Jared asked a simple question: what kind of analytics are the Vikings doing?

That message got forwarded from the owner to the VP of Player Personnel and then eventually to me, the scouting intern.

At the time, the honest answer was not much.

I was working in the scouting department, and while I had a strong instinct for data and a real interest in where the game was headed, I did not walk into that role knowing how to code or query a SQL database. I definitely was not building machine learning models. What I did have was access to information, a comfort level with Excel and the growing sense that we were sitting on a gold mine of draft data without a real system for turning it into insight.

We had hundreds of data points on every draft prospect. There were combine numbers, measurements, scout evaluations, background details, health reports, psych evals and all manner of other information collected throughout the process. It lived across spreadsheets, PDF reports and various internal sources. The more I worked with it, the more one question kept bothering me: how do we actually use all of this to predict who is going to be good?

That question led to my first real conversation with Jared.

When we got on the phone, I explained the challenge the only way I knew how, which was through football. We had all this information on player traits, but no clear framework for translating it into a usable projection. What I was really asking was something closer to: how do we predict who’s going to be good based on all of this traits data?

Jared immediately started describing the statistical side of the problem. He talked about regression, ensemble methods and the idea that you could combine different dimensions of a player into a more structured forecast. It was the first time I had heard someone take what felt like a football problem and map it so clearly into a data problem.

Then I pushed back with something I already believed from working with the data. I told him that in football, faster did not always mean better. A player did not need to be the fastest at his position. He needed to be fast enough. The same went for other traits. Bigger was not always better. Stronger was not always better. In a lot of cases, the real question was whether a player cleared a certain threshold.

That was the point where the conversation really took off.

Jared said that sounded like decision trees, and then he started walking through random forests, boosted trees and the kinds of modeling approaches that could capture those non-linear relationships. At the time, that language felt completely foreign to me. It also felt like someone had just handed me the vocabulary for ideas I had been circling around without being able to name.

That is really where the project began. Over the next few months, we built it from scratch. My job was to aggregate the data, which meant a lot of manual work in Excel. And by “a lot,” I mean the kind of work where if you closed the wrong spreadsheet at the wrong time, your whole afternoon could disappear. I pulled together information from every place I could get it, cleaned it, aligned it and shared it with Jared in spreadsheets and CSV files. Jared took those files and built out the models in R.

One of the first major insights was that this could not be one model for all players. Different positions demanded different logic. A wide receiver and a defensive tackle are evaluated through completely different lenses, so the indicators that matter for success are different too. We needed separate models by position, each one designed around the traits and outcomes that mattered for that role.

As the work progressed, we began generating outputs around the likelihood that a player would become a starter or a high-level “blue-chip” star. The challenge was that those probabilities were not naturally digestible in a draft room. Scouts and executives were not sitting around waiting for a statistical explanation in the middle of draft prep. They were trying to make football decisions, quickly, with conviction, and in a language they already trusted.

So we translated it.

We scaled the probabilities into a composite percentile score (CPS as we called it), something that felt more intuitive, more like a familiar player rating. Then we simplified it even further. On the draft board, those scores became stickers. If a player scored above the 80th percentile, he got a blue sticker. If he scored below roughly the 25th percentile, he got an orange sticker.

That was it.

Out of several hundred players on the board, about 60 ended up with a blue sticker and another 60 with an orange one. We were not trying to tell the scouts how to rank every player. We were trying to layer analytics on top of the board they had already built. The message to the general manager Rick Spielman and the rest of the staff was straightforward: scouting sets the board, and analytics can help support or challenge certain profiles within it.

That distinction mattered then, and honestly, it still matters now. Good analytics in football is not about replacing expertise. It is about sharpening it.

When the draft came around, the Vikings started with the 11th overall pick and selected cornerback Trae Waynes. As the second round, third round and later rounds unfolded, we began to see more of the blue-sticker players come into focus. By the end of the draft, nine of the team’s ten picks had blue stickers.

At the time, we were too busy living it to sit back and appreciate what that meant. We were not thinking in terms of case studies or success stories. We were just trying to get the work right and see whether the process held up once the draft actually started moving.

Over time, that class became one of the defining draft classes of that era for the organization.

Eric Kendricks developed into an All-Pro and the centerpiece of the defense. Danielle Hunter went from a third-round project to one of the most productive edge rushers in the NFL, stacking Pro Bowls and double-digit sack seasons. And Stefon Diggs, a fifth-round pick, turned into one of the most consistent and explosive receivers in football, putting together multiple 1,000-yard seasons and earning All-Pro recognition.

The class also had depth beyond those headline names. Several players became long-term starters or key contributors, and the team found additional value after the draft with undrafted free agents who went on to have real NFL careers.

As the years went on, the broader view of that draft only strengthened. What looked like a solid class in the moment became, in hindsight, one of the most impactful drafts of the decade for any team. The reason was not just the star power at the top, but the combination of star power, depth, and value relative to where those players were selected.

Retrospective analyses have consistently placed that class among the best of the 2010s. CBS Sports recently ranked the 2015 Vikings class fifth among the top NFL draft classes since 2010, while ESPN and the Star Tribune both pointed to it as one of the strongest classes from that draft cycle based on Approximate Value and long-term production.

None of the biggest hits were top-five picks. Two came in the third and fifth rounds. That is where the draft is often won or lost, and it is also where having a structured process can make the biggest difference.

Looking back from 2026, what stands out is not just that the class worked. It is that the process held up in the exact part of the draft where uncertainty is highest, where teams are trying to separate similar players with incomplete information.

At the time, using machine learning concepts and ensemble models in an NFL Draft context felt early, because it was. The tools look completely different now. The models are better, the data is richer, and AI is part of the broader conversation in nearly every industry. But the central challenge has not changed: you still have to turn information into judgment, and judgment into decisions people can actually make.

In 2015, that meant turning a complex statistical output into a blue sticker on a draft board.

Looking back, that was not a simplification of the work. It was the whole point.

Mike Band

NFL Next Gen Stats Research & Analytics

Lander Analytics Contributor

Jared P. Lander

Founder and Chief Data Scientist

Lander Analytics

---

Subscribe to our Substack and below to our monthly emails for practical AI strategies for your organization: what to build, what to avoid, and how to make systems reliable in the real world.

Work with us: If you want help identifying the right first workflow, building a permissioned knowledge base, or training your team to ship responsibly, reach out at info@landeranalytics.com.

About the author: Mike Band is the Sr. Manager of Research & Analytics at NFL Next Gen Stats and AI Researcher at Lander Analytics.

About the author: Jared P. Lander is Chief Data Scientist and founder of Lander Analytics, where he helps organizations build practical, measurable AI workflows grounded in strong data foundations.