Neural Network: NBA Draft Combine Predictions

Brya Patterson, et al., Summer 2025, University of Washington, MS Information Management (MSIM)

Brya K. Patterson

Executive Summary

  1. The Mission: I partenered with a peer at the University of Washington - Masters of Information Management, to develop a Neural Network predictive model to evaluate the ROI of "pure athleticism" in the NBA Draft, analyzing 25 years of longitudinal athlete data (2000–2025).
  2. Impact: Identified a critical "Diminishing Return" on physical combine metrics, leading to a strategic recommendation to shift scouting resources toward multi-dimensional skill assessments rather than physical testing alone.
  3. Core Capabilities: Predictive Modeling, Resource Optimization, Data Governance, Risk Mitigation.



Remediation & Recommendations

This project isn't about just basketball; it’s about using Data Science to determine where to allocate millions of dollars in capital (draft picks) while identifying the limitations of automated models.

  • The Remediation: Addressed the "Noise" in athletic testing data. By standardizing 20 years of disparate combine measurements, I created a clean, longitudinal dataset that allows for cross-era performance comparisons.
  • The Recommendation: Move beyond "Pure Athletics." The model proved that athleticism alone is an unreliable predictor of draft rank. I recommend a Hybrid Scouting Model that integrates these Neural Network predictions with qualitative "Game Intelligence" data.
  • The Future State: Proposed the integration of "Athletic Intelligence Quotient" (AIQ) data into future iterations to capture the mental processing speeds that physical drills currently miss.

    • Strategic Resource Allocation

    Proved that physical metrics provide a diminishing baseline for draft success, recommending a shift in scouting investment toward context-aware skill sets.

    Predictive Governance

    Architected a Sequential Neural Network via TensorFlow to classify "Drafted vs. Undrafted" outcomes, managing 25 years of inconsistent cross-era data.

    Model Integrity & Ethics

    Applied rigorous data preprocessing to handle legacy metrics (like the discontinued bench press), ensuring the model remained accurate across changing industry standards.

    View Full Analysis

    Overview

    This report investigates whether NBA Draft Combine performance metrics can predict draft outcomes. Utilizing a comprehensive dataset (2000–2025), we applied exploratory data analysis (EDA) and neural network modeling to evaluate the predictive weight of "pure athleticism" versus historical selection trends. While physical metrics provide a baseline, our findings suggest that the modern NBA draft process increasingly prioritizes multi-dimensional skills not captured by combine drills alone.

    Key Insights

    • Across three model families—MLP, a small Keras neural network, and a tree ensemble—the ceiling stayed around 56–60% accuracy on combine-only features. That convergence points to a data limitation, not a tuning issue.
    • Out-of-time testing in 2025 confirmed the story: accuracy ≈ 0.53, precision ≈ 0.59, recall ≈ 0.80, AUC ≈ 0.50. The model is better at finding most drafted players than ruling out non-draftees, so it’s more useful for ranking than for hard yes/no calls.
    • Permutation Feature Importance shows lane agility drill is the only consistently influential metric; most other drills contribute little. Adding Position didn’t help and sometimes hurt, suggesting limited stable signal or minor overfit.
    • The pipeline choices (fit preprocessing on 2000–2024, hold out 2025, early stopping, dropout) gave a clean generalization check and interpretable diagnostics (PFI, PDP, confusion matrix, ROC).

    Conclusion: Can NBA Draft Combine performance metrics alone predict whether a player will be drafted?

    The answer is No. Across three very different model families (MLP, a Keras neural network, and a tree ensemble) our accuracy stayed in the 56–60% range on past seasons, and the true out-of-time 2025 test landed at ~0.53 accuracy with an AUC near 0.50.

    Adding a simple positional feature did not help. These results point to a data ceiling rather than a tuning issue: draft decisions depend on factors the combine does not measure. A more holistic dataset should raise the ceiling

    Methodologically, we’d frame the task as ranking (top-K recall, PR-AUC) or predicting tiers/pick ranges rather than a hard yes/no, calibrate probabilities, tune thresholds to the intended use, and validate with rolling year-by-year splits. With that broader feature set and evaluation framing, a neural network or calibrated gradient-boosting model should produce a more accurate and actionable draft projection—something the combine alone simply can’t deliver. With a richer, context-aware dataset and calibrated modeling, we’re confident the next iteration will move from “interesting” to genuinely predictive—an exciting step toward a reliable draft forecasting tool.

    Resources