Course: DTSC 2302 — Project 2
Author: Hampton Abbott
Date: April 2026
Repository: github.com/hamptonabbott/predicting-the-madness
Companion piece: Predicting the Madness case study →

1. Problem Framing & Research Question

Research question: Can a machine learning model trained on historical NCAA regular season statistics accurately predict tournament game outcomes — and which features are most predictive of tournament success?

The NCAA Men’s Basketball Tournament presents a well-defined binary classification problem: given two teams, predict which one wins. The label is unambiguous (one team wins, one loses), historical data is plentiful (2003–2025), and the domain has a rich set of quantifiable team performance metrics.

The practical motivation: most bracket-picking approaches rely on seeding and intuition. This project tests whether a data-driven, efficiency-based approach can outperform that baseline on held-out tournament data.

2. Dataset

Source: Kaggle — March Machine Learning Mania 2026 URL: https://www.kaggle.com/competitions/march-machine-learning-mania-2026

Key Files Used

File Description Rows
MRegularSeasonDetailedResults.csv Game-level box scores, 2003–2026 124,529
MNCAATourneyDetailedResults.csv Tournament game results, 2003–2025 1,449
MNCAATourneySeeds.csv Seed assignments per team per season 2,694
MTeams.csv Team ID to name mapping 381
MMasseyOrdinals.csv Third-party rankings (KenPom, Sagarin, etc.) 5,865,001

Dataset Characteristics

  • Seasons covered: 2003–2025 (training/validation/test), 2026 (live predictions)
  • Tournament games per season: 63
  • Total matchup rows after symmetric expansion: 2,898 (each game duplicated with flipped features — see Section 4)
  • Target variable: Binary (1 = Team A wins, 0 = Team A loses)

3. Exploratory Data Analysis

EDA was conducted in Section 2 of the notebook to inform feature engineering decisions.

Key Findings

Seed win rates: 1-seeds win ~85% of first-round games. Win rates decline steadily through seed 8, then stabilize — seeds 9–16 all win between 15–35% of their games. This confirmed seed as a meaningful but non-deterministic feature.

Upset frequency: The seed matchup heatmap (notebook Section 2) shows that 5 vs. 12 and 6 vs. 11 matchups produce disproportionately frequent upsets relative to seed gap. This is consistent with the well-known “12-over-5” phenomenon in bracket folklore.

Win margins: The median win margin is 10 points; mean is 11.7 points. 28.9% of games are decided by 5 or fewer points — confirming significant randomness at the margins that no model should be expected to capture.

Score correlation: Winner and loser scores are weakly correlated (high-scoring games tend to have higher loser scores too), suggesting pace of play is a confound that efficiency metrics correct for.

4. Feature Engineering

All features are derived from regular season data only — no tournament game statistics are used to predict tournament outcomes (no data leakage).

Step 1: Per-Team Season Aggregates

For each (Season, TeamID), the following statistics were computed from MRegularSeasonDetailedResults.csv:

Feature Formula Rationale
WinPct Wins / Games Baseline team quality measure
OffEff Pts / Possessions × 100 Points per 100 possessions (offense)
DefEff OppPts / Possessions × 100 Points allowed per 100 possessions (defense)
NetEff OffEff − DefEff Net efficiency margin
eFGPct (FGM + 0.5×FGM3) / FGA Effective field goal %, accounts for 3-pt value
TOVRate TO / (FGA + 0.44×FTA + TO) Turnover rate (Dean Oliver formula)
ORBRate OR / (OR + OppDR) Offensive rebound rate
FTRate FTA / FGA Free throw attempt rate
Def_eFGPct OppFGM / OppFGA Opponent effective FG% (defensive quality)
Def_TOVRate OppTO / (OppFGA + OppTO) Opponent forced turnover rate

Possessions estimated as: FGA + 0.44×FTA − OR + TO

Step 2: Strength of Schedule

For each team-season, average opponent win percentage was computed across all regular season games. This captures whether a team’s record came against strong or weak competition.

Step 3: KenPom Rankings

From MMasseyOrdinals.csv, the most recent KenPom rank (system code KP) before day 133 (approximate Selection Sunday) was extracted per team-season. KenPom coverage was 0% prior to the years it appears in the dataset, so missing values were imputed with the training set median (see Section 5).

Step 4: Matchup Feature Construction

For each tournament game, the feature vector is constructed as the difference between Team A’s stats and Team B’s stats:

feature_diff = team_A_stat - team_B_stat

This difference-based encoding has two advantages:

  1. The model learns relative team strength, not absolute values
  2. The same game can be represented from both perspectives (symmetry)

Symmetric expansion: Each game is added twice — once with Team A as the winner (label=1) and once from the loser’s perspective (all features negated, label=0). This doubles training data and ensures the model sees both winning and losing feature patterns.

Final feature set (13 features): WinPct_diff, OffEff_diff, DefEff_diff, NetEff_diff, eFGPct_diff, TOVRate_diff, ORBRate_diff, FTRate_diff, Def_eFGPct_diff, Def_TOVRate_diff, SOS_diff, KenPomRank_diff, SeedDiff

5. Data Preprocessing & Assumptions

Missing Value Handling

  • KenPomRank has 0% coverage in early seasons — the Massey Ordinals dataset does not include KP rankings for all years
  • Strategy: impute with training set median for all features where median is non-null, then fall back to 0 for columns where the median itself is NaN
  • Rationale: median imputation avoids mean distortion from outliers; using only the training set median prevents data leakage into validation/test splits

Feature Scaling

  • StandardScaler (zero mean, unit variance) applied to features for Logistic Regression and KNN only
  • Tree-based models (Random Forest, XGBoost) do not require scaling — they are invariant to monotonic feature transformations
  • Scaler is fit on training data only and applied to validation/test sets

Key Assumptions

  1. Regular season stats are predictive of tournament performance — this is the core modeling assumption. It holds reasonably well but ignores in-season injuries, momentum, and bracket position
  2. Stationarity — the relationship between team stats and tournament outcomes is assumed stable across 2003–2025. Rule changes (e.g., shot clock modifications) could introduce drift
  3. Symmetric matchups — both teams in a matchup are treated as interchangeable; home court advantage is irrelevant in tournament play (all neutral sites)

6. Train / Validation / Test Split

A strict temporal split was used to prevent any future data from influencing training:

Split Seasons Games (rows after expansion) Purpose
Train 2003–2023 2,630 Model fitting
Validation 2024 134 Hyperparameter selection, model comparison
Test 2025 134 Final held-out evaluation
Predict 2026 — Live bracket predictions

Rationale for temporal split over random split: Random splitting would allow the model to train on 2025 data and validate on 2010 data, which would overstate generalization performance. Tournament prediction is inherently a forward-looking task — we always train on the past and predict the future.

7. Model Experiments

Four models were selected to represent a range of complexity and interpretability:

Model Why Selected
Logistic Regression Linear baseline; highly interpretable via coefficients; fast to train
K-Nearest Neighbors Non-parametric; tests whether similar teams (in feature space) have similar outcomes
Random Forest Ensemble method; handles non-linearity and feature interactions; provides feature importances
XGBoost Gradient-boosted trees; typically strong performer on tabular data; provides feature importances

Hyperparameters

LogisticRegression(C=1.0, max_iter=1000, random_state=42)
KNeighborsClassifier(n_neighbors=11)
RandomForestClassifier(n_estimators=300, max_depth=6, random_state=42)
XGBClassifier(n_estimators=300, learning_rate=0.05, max_depth=4,
              subsample=0.8, eval_metric='logloss', random_state=42)

KNN with n_neighbors=11 was chosen to avoid ties (odd number) while smoothing over local noise. Random Forest max_depth=6 prevents overfitting on a small dataset. XGBoost learning_rate=0.05 with 300 estimators follows standard practice for small tabular datasets.

8. Model Evaluation & Interpretation

Validation Set Results (2024 Tournament)

Model Accuracy Log Loss ROC-AUC
Logistic Regression 0.672 0.569 0.764
K-Nearest Neighbors 0.567 0.631 0.655
Random Forest 0.694 0.565 0.769
XGBoost 0.642 0.631 0.720

Best model selected: Random Forest (lowest log loss on validation set)

Why log loss over accuracy: Log loss penalizes confident wrong predictions — a model that says “70% chance Team A wins” and Team A loses is penalized more than one that says “52%.” For bracket prediction, calibrated probabilities matter, not just binary picks.

Test Set Results (2025 Tournament — Held Out)

Model Accuracy Log Loss ROC-AUC
Logistic Regression 0.761 0.470 0.869
K-Nearest Neighbors 0.851 0.441 0.881
Random Forest 0.739 0.468 0.883
XGBoost 0.754 0.477 0.848

All models significantly outperform the 50% random baseline on the 2025 tournament. The improvement from validation to test accuracy is notable — 2025 may have been a year with fewer major upsets than 2024.

Interpretation

The Random Forest was best on validation log loss but KNN outperformed it on the test set. This suggests the optimal model may vary year-to-year — an ensemble or stacking approach could provide more stable performance across seasons.

All models share the characteristic of predicting expected winners well but struggling with upsets. This is expected: upsets are, by definition, low-probability events that the model correctly assigns low probability to — but they happen anyway.

9. Feature Importance

Feature importance was extracted from Random Forest (feature_importances_), XGBoost (gain-based), and Logistic Regression (coefficient magnitudes). See Section 6 of the notebook for full charts.

Consistent top features across models:

  1. SeedDiff — seed gap between teams; strongest individual predictor by a wide margin
  2. SOS_diff — strength of schedule difference; second most important feature
  3. NetEff_diff — net efficiency difference; strongest efficiency-based predictor
  4. OffEff_diff — offensive efficiency gap
  5. WinPct_diff — win percentage gap

Notable finding: TOVRate_diff and FTRate_diff ranked consistently low across all models, suggesting that turnovers and free throw attempts are less predictive of tournament outcomes than efficiency and overall record. This aligns with basketball analytics literature suggesting that efficiency (not volume) is the key discriminator.

Logistic Regression coefficients: Positive coefficients (features that increase win probability when Team A is better) were highest for NetEff_diff and SeedDiff. Negative coefficients appeared for DefEff_diff — a higher defensive efficiency number is worse (more points allowed per possession).

10. 2026 Predictions & Results

2026 regular season stats were computed using the same build_team_stats() pipeline applied to 2003–2025 data. The 2026 KenPom rankings and strength of schedule were computed from 2026 regular season games only.

The 2026 seeds are available in the Kaggle dataset (68 teams). Win probabilities for each matchup are generated by predict_game(), which constructs the difference feature vector and passes it through the trained Random Forest model.

Final Results (Complete 2026 Tournament)

All game outcomes were manually sourced from NCAA.com and entered in Section 8 of the notebook, as the Kaggle dataset had not published 2026 tournament results by the project deadline.

Round Correct Total Accuracy
Round of 64 23 28 82.1%
Round of 32 12 16 75.0%
Sweet 16 5 8 62.5%
Elite Eight 3 4 75.0%
Final Four 2 2 100.0%
Championship 1 1 100.0%
Main Bracket Total 46 59 78.0%

Note: The First Four (play-in games) are excluded from main bracket accuracy — those games match teams of identical seed, so the model has no meaningful differential signal and coin-flip accuracy is expected.

Interpretation

The model performed strongest in the Round of 64, where seed gaps and efficiency differentials are typically largest. Performance declined in the Sweet 16, where remaining teams are closely matched and genuine randomness dominates. The model recovered strongly in the Final Four and Championship, correctly predicting all three games including the national champion.

2026 Champion: Michigan Wolverines — defeated UConn 69–63 on April 6. The model correctly identified Michigan as the winner of both the Final Four and the Championship game.

11. Limitations, Ethics & Reflection

Limitations

  1. Small sample size: 63 tournament games per year × 22 years = ~1,400 total games. This is a small dataset for machine learning — the model cannot learn rare upset patterns reliably.

  2. No in-season dynamics: Features are season-level averages. A team that peaked in February and declined in March looks the same as one that’s on a hot streak. Recency-weighted features could improve this.

  3. Missing KenPom coverage: KenPom rankings are absent for early seasons in the Massey Ordinals file, requiring imputation. This weakens the model for those years.

  4. Bracket path independence: The model predicts individual matchups independently. It does not account for fatigue, injury, or the fact that a team’s opponent in round 3 depends on who won rounds 1 and 2.

  5. Outcome variance: March Madness has genuine randomness. Games decided by 1–3 possessions are not reliably predictable from season-level statistics. A model that achieves 75–85% accuracy is likely near the ceiling for this approach.

Ethical Considerations

  • The model uses publicly available game data with no personally identifiable information
  • Tournament predictions could theoretically inform gambling decisions — this project is purely academic and not intended for that purpose
  • Seeds and rankings reflect human committee decisions that may embed biases (e.g., conference prestige, geography). These are used as features here without attempting to correct for them

Reflection

The most surprising finding was that KNN — the simplest non-parametric model — outperformed XGBoost on the 2025 test set. This suggests the feature space may have a relatively smooth structure where “similar teams” (in terms of stats) genuinely have similar outcomes. It also highlights that more complex models don’t always win on small datasets.

The project confirms that data-driven efficiency metrics are genuinely predictive of tournament success, but the gap between good predictions and perfect predictions is largely explained by the inherent randomness of close games — not model inadequacy.

12. Code Quality & Reproducibility

The full analysis is contained in a single Jupyter notebook (notebooks/march_madness.ipynb) with the following structure:

  1. Setup & Data Loading
  2. Exploratory Data Analysis
  3. Feature Engineering
  4. Train / Validation / Test Split
  5. Model Training & Comparison
  6. Feature Importance
  7. 2026 Tournament Predictions
  8. Results vs. Actuals

To reproduce:

# 1. Clone the repo
git clone https://github.com/hamptonabbott/predicting-the-madness.git

# 2. Install dependencies
pip install -r requirements.txt

# 3. Download data (requires Kaggle API token)
kaggle competitions download -c march-machine-learning-mania-2026 -p data/raw
unzip data/raw/march-machine-learning-mania-2026.zip -d data/raw

# 4. Run notebook
jupyter notebook notebooks/march_madness.ipynb

References & AI Transparency

Dataset: Kaggle — March Machine Learning Mania 2026. https://www.kaggle.com/competitions/march-machine-learning-mania-2026

Libraries:

  • pandas, numpy — data manipulation
  • scikit-learn — Logistic Regression, KNN, Random Forest, StandardScaler, metrics
  • xgboost — XGBoost classifier
  • matplotlib, seaborn — visualizations

Basketball analytics reference: Oliver, D. (2004). Basketball on Paper. Brassey’s Inc. — source for the Four Factors framework (eFG%, TOV rate, ORB rate, FT rate) and possession estimation formula.

AI tools used: Claude (Anthropic) was used to help structure and debug the notebook and assist with feature engineering. All modeling decisions, interpretations, and written analysis reflect my own understanding and judgment.