Every Shot Is a Measurement: Possibility-Adjusted Shot Value for NBA Decision Quality
Open-source code, data, pre-registration history, and validation artifacts for the MIT Sloan Sports Analytics Conference 2027 Basketball-track submission.
Author: Robert “Bobby” Morong (DataDunkNBA)
Contact: bobby@trainingties.com
Current SSAC27 abstract: paper/SSAC27_PASV_Abstract_v4_2026-08-26.md
Historical full manuscript: paper/SSAC27_PASV_PAPER_FULL_v1.md
License: MIT
Current evidence state (Sept. 2026): PASV remains an interpretable counterfactual shot-decision framework. In the held-out per-shot study, the current public-event continuation term does not add predictive information beyond shot quality. The negative result is part of the contribution and defines the richer-state research frontier.
PASV grades a shot against the value of continuing the possession:
PASV = xPTS(shot) - V*(s*)
where xPTS(shot) is expected points from the shot taken and V*(s*) is the estimated value of declining that shot and preserving the possession's best available continuation.
The research question is not merely whether a shot was good. It is whether the shot was good relative to the alternatives it ended.
PASV is explicitly downstream of prior work on shot selection and possession value. See Related work below.
The v4 Sloan abstract uses the held-out per-shot Study 1 as the empirical spine:
- Calibration: 219,527 field-goal attempts from the 2024-25 NBA regular season.
- Held-out test: 14,377 playoff attempts.
- Within-player R² on the playoffs:
- xPTS: 0.0231
- PASV: 0.0216
- public-event Skinner comparison: 0.0125
- Adding PASV to xPTS + player fixed effects: Delta R² ≈ 0.00007, nested-test p = 0.31, cluster-robust p = 0.24, with worse AIC.
- PASV is approximately 0.97–0.98 correlated with xPTS within player at current public-event resolution.
PASV improves on the repository's coarse public-event implementation of the classical cutoff benchmark, but does not beat xPTS. The current V*(s*) varies too little shot-to-shot to create independent signal.
That means the next test is not “tune the same proxy harder.” It is to measure continuation value with richer possession state: defender positioning, help configuration, spacing, transition state, action history, and true shot-clock context.
- A shot decision can be expressed as a signed counterfactual comparison between the chosen shot and preserved continuation value.
- PASV is an interpretable framework for that comparison.
- The current public-event continuation implementation is empirically near-collinear with xPTS.
- The held-out null is informative because it localizes what additional state is required for possibility cost to become independently measurable.
- The pre-registration and failure receipts are part of the research record rather than being hidden.
- PASV does not currently outperform xPTS/shot quality.
- The current clock/public-event continuation proxy does not establish independent predictive value.
- The two 2026 preregistered series misses are not retroactively reinterpreted as wins.
- The public Skinner comparison is not a claim of superiority over a fully observed shot-clock implementation.
- Holding Math is a conditional mechanism model, not an observed universal NBA law.
- The AST%-style OPC proxy is not a validated substitute for tracking-grade option preservation.
.
├── README.md
├── LICENSE
├── requirements.txt
├── paper/
│ ├── SSAC27_PASV_Abstract_v4_2026-08-26.md # current Sloan abstract framing
│ ├── SSAC27_PASV_Abstract_v3_2026-06-29.md # historical submission version
│ ├── SSAC27_PASV_Section_5_v2_2026-06-29.md # updated empirical section
│ ├── SSAC27_PASV_PAPER_FULL_v1.md # historical full manuscript; not current empirical authority
│ └── SSAC27_PASV_Holding_Math_Theorem_v2_2026-06-13.md
├── code/ # PASV, baselines, sensitivity, DTI/continuation work
├── data/ # bundled public-data artifacts where redistribution is allowed
├── notebooks/ # reproduction notes/specifications
├── pre_registration/ # immutable pre-tip predictions + grading receipts
├── results/ # held-out Study 1 and other receipts
└── npss/ # separate playoff-fragility subproject; see its own correction history
The historical full manuscript remains in the repository for auditability, but its June framing should not be treated as the current empirical authority where it conflicts with the later Section 5 / v4 abstract.
PASV builds on, and must be read against, four prior lines of work:
- Skinner, B. (2012). The problem of shot selection in basketball. PLOS ONE 7(1): e30776. https://doi.org/10.1371/journal.pone.0030776. The optimal-stopping frame PASV uses.
- Cervone, D., D'Amour, A., Bornn, L., & Goldsberry, K. (2016). A multiresolution stochastic process model for predicting basketball possession outcomes. Journal of the American Statistical Association 111(514). arXiv:1408.0777. Expected possession value as a continuously updated quantity.
- Sandholtz, N., & Bornn, L. (2020). Markov decision processes with dynamic transition probabilities: an analysis of shooting strategies in basketball. Annals of Applied Statistics 14(3): 1122–1145. https://doi.org/10.1214/20-AOAS1348. Region-level expected points inside an MDP of shot policy; names defender-conditioned action values as future work.
- Duma, N., & Benaharon, L. (2026). Efficient frontier optimization of NBA shot selection using Markov reward processes. arXiv:2607.02933. A static shot-diet frontier finding the league near-efficient, with gains from more rim attempts; names missing personnel identification as a limitation.
The shot-value ladder in this repository (dunk 1.81, rim 1.26, corner three 1.17, above-the-break three 1.07, mid-range and paint non-rim 0.86) reproduces public values and is pipeline validation, not a contribution. The direction "more rim, less mid-range" is Duma & Benaharon's published result. PASV's open question is the state-contingent one: whether conditioning on defender identity changes the value of continuing a possession.
The May 26, 2026 filing committed PASV v0.1 to playoff-series predictions before the outcomes resolved. The subsequent grading recorded two consecutive misses. Those misses remain immutable research history.
The project treats this as methodology, not embarrassment: claims are graded against the version that existed before the result was known.
The pre-registration files are frozen under git tag prereg-pasv-v0.1, which points at commit 7019298, the first commit containing them (2026-06-08). The filing itself is dated 2026-05-26 and the grading receipts 2026-06-03; the tag proves the files have not changed since 2026-06-08, not the earlier filing date.
The repository also contains the NPSS / Schemable Big subproject. Its original held-out implementation treated every non-zero continuous hunt value as truthy and therefore inflated the earlier 100% recall claim. That claim is retired. With the corrected hunt >= 0.30 rule, held-out recall is 66.7% (4 of 6) and the held-out schemable-big decline is −0.190 AQI; the full table and raw output are in npss/results/.
The repository contains both early team-aggregate PASV work and the later per-shot validation work. When reproducing the Sloan v4 result, use the files referenced by the v4 abstract / Study 1 receipt rather than assuming the original pasv_v01.py team-aggregate quick start reproduces the held-out per-shot study.
A remaining hardening task is to make the full per-shot Study 1 input pipeline self-contained or regenerable from frozen source manifests, with all Python/parquet dependencies pinned. The regeneration path is documented in DATA.md; the possession parquets themselves are not bundled.
Until conference acceptance, cite this as a submitted/open-source research project rather than as accepted proceedings:
@misc{morong2026pasv,
title={Every Shot Is a Measurement: Possibility-Adjusted Shot Value for NBA Decision Quality},
author={Morong, Robert},
year={2026},
note={Submitted to the MIT Sloan Sports Analytics Conference 2027; open-source research repository},
url={https://github.com/GrobeStreet/pasv}
}MIT. See LICENSE.
PASV builds on Brian Skinner's The Problem of Shot Selection in Basketball (2012) and Cervone, D'Amour, Bornn, and Goldsberry's work on Expected Possession Value (2014). The framework extends those ideas by making the foreclosed continuation explicit in the shot-decision score.
Independent DataDunkNBA research. The evidence record includes positive findings, nulls, implementation corrections, and failed predictions.