Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/Stronauta/MLB-Performance-Analytics/llms.txt

Use this file to discover all available pages before exploring further.

This page presents the results of the three-class performance classifier trained on MLB Statcast data. Using xwOBA, hardhit_percent, barrels_total, and hits as input features, the model assigns each batter one of three performance tiers — Bajo (Low), Medio (Medium), or Alto (High) — derived from observed wOBA quantiles. The goal is to move beyond simple batting averages and instead capture the full depth of a batter’s offensive contribution using both traditional counting stats and modern expected metrics.

​
Target Variable Summary

The target label Rendimiento_labels is built directly from each batter’s woba value using equal-frequency quantile binning (pd.qcut with q=3). This guarantees a balanced class distribution regardless of the underlying wOBA spread, placing roughly 285 batters per tier across the 857-player cleaned dataset.
TierLabelDescription
Bottom third by wOBABajo (Low)Below-average offensive producers
Middle third by wOBAMedio (Medium)League-average offensive producers
Top third by wOBAAlto (High)Above-average offensive producers
Caracteristicas = ["xwoba", "hardhit_percent", "barrels_total", "hits"]

df["Rendimiento_labels"] = pd.qcut(
    df["woba"],
    q=3,
    labels=["Bajo", "Medio", "Alto"]
)

X = df[Caracteristicas]
Y = df["Rendimiento_labels"]

X_train, X_test, Y_train, Y_test = train_test_split(
    X, Y, test_size=0.2, random_state=42
)
Because pd.qcut splits by equal-sized quantile buckets, each class contains approximately the same number of players. This avoids the class imbalance problems that can distort model training when one tier is far more common than another.

​
Feature Importance Findings

After training the Random Forest ensemble (n_estimators=300, max_depth=8), feature importances were extracted and visualized as a horizontal bar chart. The result is unambiguous: xwOBA is the single dominant predictor, followed by hits as the second most important feature. 
importancias = rf.feature_importances_
indices = np.argsort(importancias)

plt.figure(figsize=(10, 6))
plt.title('Importancia de las Características - Random Forest')
plt.barh(range(len(indices)), importancias[indices], align='center')
plt.yticks(range(len(indices)), [Caracteristicas[i] for i in indices])
plt.show()
The importance ranking from highest to lowest is:
  1. xwOBA — The strongest single signal. xwOBA accounts for the quality of contact on every batted ball, stripping out luck components like defensive positioning.
  2. Hits — Traditional hit count still carries predictive value, reflecting sustained plate appearances where contact translates to reaching base.
  3. barrels_total — Barrels (exit velocity ≥ 98 mph at optimal launch angle) contribute, but are somewhat redundant with xwOBA, which already encodes contact quality.
  4. hardhit_percent — Provides marginal additional signal beyond barrels; hard-hit balls that miss barrel thresholds still contribute to overall performance.
The dominance of xwOBA over raw hits confirms that Statcast expected metrics outperform traditional counting stats in classifying batter performance tiers. If you are building a scouting model, prioritize exit-velocity-based expected metrics over box-score totals.
This finding is consistent with modern sabermetric research: expected metrics derived from exit velocity and launch angle are more stable year-to-year than observed outcomes, making them superior features for tier classification.

​
Model Performance

Two classifiers were trained and compared: a Decision Tree baseline and a Random Forest ensemble. The Decision Tree (max_depth=6, min_samples_split=20) provides interpretability, while the Random Forest (n_estimators=300, max_depth=8) leverages ensemble averaging to reduce variance and improve generalization.

​
Decision Tree Baseline Metrics

Evaluated on the 20% held-out test set using weighted averaging across the three classes:
MetricScore
Accuracy0.6337
Precision0.6594
Recall0.6337
F1 Score0.6378
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score

acc       = accuracy_score(Y_test, y_pred_dt)
precision = precision_score(Y_test, y_pred_dt, average='weighted')
recall    = recall_score(Y_test, y_pred_dt, average='weighted')
f1        = f1_score(Y_test, y_pred_dt, average='weighted')

print(f"Árbol de Decisión - Accuracy: {acc:.4f}, Precision: {precision:.4f}, "
      f"Recall: {recall:.4f}, F1 Score: {f1:.4f}")
# Output: Árbol de Decisión - Accuracy: 0.6337, Precision: 0.6594,
#         Recall: 0.6337, F1 Score: 0.6378

​
Random Forest Improvement

The Random Forest improves on every metric through the ensemble effect: by averaging predictions across 300 decision trees with bootstrapped samples and random feature subsets, it reduces overfitting and corrects for the instability of any single tree’s splits. On a three-class problem with non-linear interactions between xwOBA, barrels, and hits, ensemble averaging captures the feature interaction signal that a single tree misses at the boundary between Medio and Alto tiers.
The Decision Tree baseline precision (0.6594) is slightly higher than its accuracy (0.6337), indicating that when the tree does commit to a class, it is generally correct — but it misses a portion of true positives. The Random Forest addresses this recall gap.

​
Key Insights

​
wOBA vs xwOBA: The Luck–Skill Diagonal

The scatter plot of woba (x-axis) versus xwoba (y-axis) shows the majority of batters clustering tightly along the diagonal — players whose actual results closely matched the quality of their contact. The most analytically interesting batters are the outliers:
  • Above the diagonal (xwOBA > wOBA): Players whose contact quality exceeded their observed results. This suggests bad luck — well-struck balls that happened to find fielders, or a defense-suppressed BABIP — and these batters are candidates for positive regression.
  • Below the diagonal (wOBA > xwOBA): Players whose results exceeded their contact quality, often driven by favorable defensive positioning, high BABIP, or clutch timing. These batters carry regression risk.
Most batters fall within a narrow band around the line, confirming that over a multi-year sample, observed wOBA and expected wOBA converge.

​
Hard Hit Rate and Barrels: Correlated Metrics

The scatter of hardhit_percent vs barrels_total reveals a strong positive correlation — as expected, because a barrel is a strict subset of hard-hit balls. Every barrel is a hard-hit ball, but not every hard-hit ball meets the launch angle criteria to qualify as a barrel. This correlation means including both features in the model introduces some redundancy, which the Random Forest handles naturally through its feature subsampling mechanism at each split.

​
Top Performers by xwOBA

The top batters in the Alto tier by xwOBA in the dataset:
PlayerxwOBAwOBA
Aaron Judge0.4690.457
Shohei Ohtani0.4330.427
Juan Soto0.4330.402
Ronald Acuña Jr.0.4240.403
Yordan Alvarez0.4190.397
All five land comfortably in the Alto tier, with Aaron Judge showing the largest xwOBA in the entire dataset — driven by his historically elite exit velocity and launch angle profile.

​
Classification Rationale

​
Why wOBA Over Batting Average?

Batting average (BA) treats every hit identically — a bloop single and a 450-foot home run count the same. wOBA (Weighted On-Base Average) corrects this by assigning each offensive outcome a weight proportional to its actual run-creation value:
EventApproximate wOBA Weight
Walk (BB)~0.69
Single~0.88
Double~1.27
Triple~1.62
Home Run~2.10
This means a batter who draws many walks and hits for extra bases will have a higher wOBA than a contact hitter who accumulates singles despite a similar batting average. Using wOBA quantiles to construct the performance tiers therefore creates labels that genuinely reflect offensive value, not just contact frequency.
xwOBA (expected wOBA) extends this further by computing what a batter’s wOBA should have been based entirely on the exit velocity and launch angle of each batted ball, independent of where the ball landed or how the defense was positioned. Because xwOBA strips out luck, it is more stable across seasons than observed wOBA — making it the most powerful individual feature in this classifier.
When interpreting tier assignments, look at both woba and xwoba. A batter labeled Bajo with a high xwoba may be a value target: their observed results understate their true contact quality, and regression to the mean should push them into the Medio or Alto tier in future seasons.

Build docs developers (and LLMs) love

Get started for freeTalk to us