scikit-learn basics

Level 4 · Lesson 4

Hook

A GM doesn’t just describe last season — they predict next season. scikit-learn is the tool. Three methods (fit, predict, score) handle almost everything; the rest is feature engineering and not lying to yourself about how well it works.

Concept

Every sklearn model has the same shape:

from sklearn.linear_model import LinearRegression

model = LinearRegression()
model.fit(X_train, y_train)
predictions = model.predict(X_test)
r2 = model.score(X_test, y_test)

• X is a 2-D array (rows = samples, columns = features).
• y is a 1-D array of targets (what you’re predicting).
• fit(X, y) learns from training data.
• predict(X) outputs predictions for new data.
• score(X, y) returns a default metric (R² for regression, accuracy for classification).

The model menu, starting points:

Problem	Default first model
Regression (predict a number)	LinearRegression, then RandomForestRegressor
Classification (predict a category)	LogisticRegression, then RandomForestClassifier
When in doubt	Try both linear and tree-based; compare

Lions example

Predict QB passing yards in a game from rushing yards. (Toy example — the real version of this needs more features.)

import pandas as pd
from sklearn.linear_model import LinearRegression
from sklearn.model_selection import train_test_split
from sqlalchemy import create_engine

eng = create_engine("postgresql+psycopg://onepride:lions@localhost:5432/onepride")

df = pd.read_sql(
    """
    SELECT player_display_name, season, week,
           passing_yards, attempts, completions
    FROM weekly_stats
    WHERE position = 'QB'
      AND season BETWEEN 2021 AND 2024
      AND season_type = 'REG'
      AND attempts >= 15
    """,
    eng,
)

X = df[['attempts', 'completions']].to_numpy()
y = df['passing_yards'].to_numpy()

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42
)

model = LinearRegression()
model.fit(X_train, y_train)
print(f"R² = {model.score(X_test, y_test):.3f}")
print(f"Coefficients: {dict(zip(['attempts', 'completions'], model.coef_))}")

You’ll get an R² in the 0.8+ range — completions and attempts predict yards well because they’re mechanically tied (every completion is some yards). That’s high R² for the wrong reason — it’s a target-leakage near-miss, since completions is observed at the same time as passing_yards. Lesson 5 covers leakage formally; for now, just notice that this number looks too clean. Real predictive problems are much harder.

Try it

Predict rushing yards per game for NFL running backs from carries, in 2024 regular season only. Use a LinearRegression. Report R² on the test set, plus the coefficient on carries. Interpret the coefficient — what does it physically mean?

Common mistakes

• Fitting and scoring on the same data. Always split train/test. A model that perfectly predicts training data tells you nothing.
• Not setting random_state. Train/test splits should be reproducible.
• R² on classification. R² is for regression. For classification, use accuracy, precision, recall, or F1.
• Forgetting feature scaling. Linear models, SVMs, and KNN are scale- sensitive — use StandardScaler before fitting. Tree-based models don’t care.

Quick check

1. What do fit, predict, and score do?
2. Why do you need separate train and test sets?
3. When is feature scaling necessary?