What if we consider the NBA draft as just (another) Moneyball model?

I recently wrote about how I've been working on turning the models I presented at the MIT Sloan Sports Analytics Conference into actionable metrics that NBA teams can use to augment their draft decision making. I received some really good feedback asking how my approach compared to human decision making (thanks Joel Shapiro and Seth Partnow!)

A simple way to benchmark any draft model is to compare it to the draft itself. So I thought I'd give that a crack.

A Simple Baseline

I treated draft position as a very coarse predictive model:

• Round 1 → Impact Player

• Round 2 → Rotation Value

• Undrafted → Replacement Risk

This is obviously crude, but it reflects how draft capital is implicitly used.

I then compared this baseline to the framework I’ve been developing that separates:

• Level (expected NBA impact)

• Delta (development trajectory / variance)

The draft is strong at identifying Impact Players, but struggles to distinguish Rotation Value from Replacement Risk. The Level/Delta model improves this middle-tier separation.

What the Draft Gets Right

The draft performs reasonably well overall:

• Accuracy: ~48%

• Strong at identifying Impact Players

This aligns with intuition:

Teams are generally good at identifying top-end talent.

Where It Breaks Down

The issue is the middle.

From the confusion matrix:

• A large number of players projected as Rotation Value end up as Replacement Risk

• The draft struggles to distinguish between these two outcomes

In other words:

The draft is much less reliable at separating useful players from non-contributors than it is at identifying stars.

What Level vs Delta Changes

The Level/Delta model improves overall accuracy by around 5-percentage points (~53%), but more importantly is:

• Better at identifying Rotation Value players

• Better at identifying Replacement Risk

• More balanced performance across all three archetypes

It doesn’t dramatically outperform the draft in headline accuracy.

Instead, it improves where the draft is weakest.

Why This Matters

Most draft value isn’t created at the very top.

It’s created by:

• Avoiding misses

• Identifying reliable contributors

• Pricing risk correctly

The draft is already strong at identifying upside.

The opportunity lies in improving decision-making in the middle of the distribution.

A Different Framing

As I outlined in my previous post, instead of asking:

How good will this player be?

This approach asks:

What kind of bet is this player?

• Level → expected return

• Delta → trajectory / variance

I’ll follow-up soon with a more detailed analysis, including a machine learning model that predicts archetype directly from draft position and how it compares to the Level vs Delta approach my colleagues and I at Synaptiq are developing.