wenmin-wu/tabular-season-phase-labeling

Season Phase Labeling

Overview

Sports, retail, and event-driven datasets exhibit distinct behavioral regimes tied to calendar phases. Mapping each date to its season phase (offseason, preseason, regular season, postseason) creates a powerful categorical feature that captures regime-dependent patterns. Using np.select with ordered conditions and boundary dates handles complex multi-phase calendars cleanly.

Quick Start

import numpy as np

conditions = [
    df["date"] < df["preseason_start"],
    df["date"] < df["regular_start"],
    df["date"] <= df["regular_end"],
    df["date"] < df["postseason_start"],
    df["date"] <= df["postseason_end"],
    df["date"] > df["postseason_end"],
]
labels = ["offseason", "preseason", "regular", "postseason_gap", "postseason", "offseason"]

df["season_phase"] = np.select(conditions, labels, default="unknown")

Workflow

1. Obtain boundary dates for each phase (from a schedule table or hardcoded per year)
2. Define ordered conditions — np.select evaluates top-to-bottom, first match wins
3. Assign phase labels as a new categorical column
4. Optionally split further (regular season 1st half / 2nd half, all-star break)
5. Use as a categorical feature in GBDT models or for phase-specific model training

Key Decisions

• Condition order: must be chronological; overlapping conditions resolved by first match
• Granularity: 4-6 phases is typical; finer splits (weekly sub-phases) risk overfitting
• Multi-year: join a season calendar table by year to handle varying boundary dates
• Interaction features: combine phase with day-of-week for richer patterns
• Phase-specific models: train separate models per phase when behavior differs drastically

References

• EDA of MLB for starter