Hawk-Eye, Second Spectrum, and the 29-Point Question

The NBA’s shift from Second Spectrum’s 2D dot tracking to Hawk-Eye’s 14-camera, 29-point 3D pose system unlocks a qualitatively different analytics layer, and the gap between franchises building on top of that layer and franchises still consuming yesterday’s derived metrics is now the single biggest competitive asymmetry in NBA analytics procurement.

The technical shift

The Second Spectrum system, which held the league-wide tracking contract from 2017, produced one coordinate per player per frame, updated at roughly 25 frames per second. It was a revolutionary dataset. It was also flat: a dot on a floor, no information about what the player’s body was doing.

The Hawk-Eye system, rolled out league-wide starting in 2023-24, uses 14 cameras producing 29 keypoints per player in full 3D. That is arms, legs, hands, torso orientation, ball position relative to body. The output is not a dot. It is a rigged skeleton.

The difference sounds incremental. It is not. Moving from 2D dots to 3D pose is the difference between knowing where a player is and knowing what a player is doing. Every metric derived from the former has a higher-fidelity analog derivable from the latter. Most have not been derived yet.

Three metrics that are cheap with pose data and expensive without

Biomechanical load estimation. Cumulative mechanical stress on a player’s body, decomposed by joint, across a season. Two-dimensional dot data cannot produce this reliably. Pose data can. Injury prediction becomes a tractable problem when the input includes joint-specific load histories rather than just minutes played.

Contested-shot classification by geometry. How contested is a jump shot? The traditional answer is “defender within four feet.” The pose-data answer is “defender’s closest hand is within X inches of the shooter’s release point, with Y degrees of arm extension.” The pose-data version is substantially more predictive of shot make probability, and it stays stable across defensive schemes in a way the distance-based version does not.

Off-ball screen geometry. Screen quality has been one of the whitest whales in basketball analytics. With pose data, screen contact, screener orientation, and defender navigation path become directly measurable. Without it, the best public work is shot-chart-adjacent proxies that throw away most of the signal.

The procurement map. Twenty commonly-cited NBA metrics, sorted by the minimum data tier you need to compute them. A check marks the lowest tier that gets you a credible version. Read it one way: a vendor pitching you anything from the first two columns is selling you something you can build.

Six metrics live in column one. Nine live in column two and converge a little more every offseason. Five live in column three, and they are the five that will separate analytics departments over the next three years. The asymmetry is not subtle once you see it laid out.

The public API gap

The public NBA Stats API exposes a thin slice of what tracking data captures. Shot locations are public. Touch statistics are public. Defensive assignments, in aggregate, are semi-public. The raw player tracking feed, the pose data itself, and most derived advanced metrics are not.

This creates a three-tier reality for any analytics vendor.

Tier one: box score derived metrics. Available to everyone, fully commoditized.

Tier two: public-API derived metrics. Available to anyone who can write the extraction code. Not commoditized yet, but converging.

Tier three: pose-data derived metrics. Available only to teams and to vendors with licensing deals. The metrics that will define competitive analytics over the next three years live here.

A vendor building exclusively on tier one or tier two is not competing at the frontier. A franchise buying a vendor that cannot demonstrate tier three capability is buying a consolation product.

The FTN and licensing calculus

Several vendors sell charted or derived data that approximates tier-three signal without direct pose-data access. Defensive matchups. Possession-level tracking extracts. Shot-quality estimates. Franchises have to decide what to buy, what to build, and what to ignore.

The question to ask about any licensed data source is: does this produce a signal I cannot produce internally from the public API, and is the incremental signal worth the cost and the dependency?

For defensive assignment data, historically, the answer has been yes. Public data on who is guarding whom at any given possession is thin. Charted vendors provide it reliably. That has been worth paying for.

For shot quality estimates, the answer is increasingly no. Enough public-API signal exists to build a reasonable shot-quality model internally. Paying a vendor to do it duplicates work.

For biomechanical load and pose-derived metrics, the answer depends on whether the vendor has pose-data access, which most don’t. Most “advanced” metrics sold to teams are tier two products marketed as tier three.

The vendor resilience thesis, applied

This matters procurement-wise because single-vendor dependency on tier-three data is a strategic risk. PFF’s acquisition by Teamworks made that concrete for the NFL side. The NBA version is less consolidated today but trending the same direction.

A franchise should be building its analytics stack with three principles.

First, own the public-data derived layer internally. Do not outsource it. The features are the product.

Second, license charted or tier-three data with contracts that allow export and internal re-derivation. Avoid opaque feeds that create downstream lock-in.

Third, invest in pose-data capability, whether through direct Hawk-Eye licensing, partnership with teams who have access, or vendors with demonstrable pose-data pipelines.

Five questions to ask every analytics vendor in your next RFP

1. What tier of data underpins each metric you sell?
2. For tier-three metrics, do you have direct pose-data access or are you inferring from lower-tier inputs?
3. If you lose data access for any reason, which metrics degrade and by how much?
4. Can we export your raw outputs for internal re-derivation, or are we locked into your aggregated feed?
5. What is your retraining cadence, and how do you detect drift when the league changes?

A vendor that can answer all five directly is rare. That is the point.

What we owe Second Spectrum

A final note. The Hawk-Eye transition is not a repudiation of Second Spectrum’s contribution. Rajiv Maheswaran and his team built the foundation that every modern NBA analytics product rests on, including the products built on Hawk-Eye output. The 2D dot era was not wrong. It was the maximum of what was computationally feasible at the time. The field moved forward because that era existed. Crediting that history is not optional for anyone serious about the space.

The pose-data era inherits the Second Spectrum framework and extends it. The vendors, franchises, and researchers who treat it that way are the ones who will ship the next generation of metrics. The ones who pretend the prior era did not happen will repeat its lessons poorly.

If you are scoping pose-data integration and want a technical read on the build-versus-buy calculus, contact us at @xVictoryarx. We have opinions and the architecture to back them.