Computing New Columns: Rate Stats and Feature Engineering

The columns a dataset ships with are rarely the ones you want to analyze. A standings file gives you season totals — runs scored, runs allowed, games played — but the questions you actually ask are about rates: runs per game, run differential per game, winning percentage. So you build the columns you wish the file came with. That's feature engineering, and in pandas each new column is a single line. Four of those lines, then a dumbbell chart that only makes sense once they exist — that's the whole job here.

This builds on Pandas for Sports Data. The data is the bundled sample_standings.csv (real 2023 MLB standings, MLB Stats API, retrieved June 2026), so it runs offline.

1. Build new columns from old ones

   Assigning to a column name that doesn't exist yet creates it. Because pandas math is vectorized, dividing one column by another applies to all 30 rows at once — no loop.

   import pandas as pd
   
   df = pd.read_csv("sample_standings.csv")
   df["RS/G"] = (df["RS"] / df["G"]).round(2)
   df["RA/G"] = (df["RA"] / df["G"]).round(2)
   df["RunDiff/G"] = (df["RunDiff"] / df["G"]).round(2)
   df["WinPct"] = (df["W"] / df["G"]).round(3)
   print(df.sort_values("RunDiff/G", ascending=False)[["Team","RS/G","RA/G","RunDiff/G","WinPct"]].head(6).to_string())

   Four engineered columns, best run differential first

         Team  RS/G  RA/G  RunDiff/G  WinPct
   0   Braves  5.85  4.42       1.43   0.642
   2  Dodgers  5.59  4.31       1.28   0.617
   3     Rays  5.31  4.10       1.20   0.611
   7  Rangers  5.44  4.42       1.02   0.556
   5   Astros  5.10  4.31       0.80   0.556
   1  Orioles  4.98  4.19       0.80   0.623

   Now the table speaks in rates. The 2023 Braves scored 5.85 runs a game and allowed 4.42 — a differential of +1.43 per game, the best in baseball — on the way to a .642 winning percentage. Those numbers existed in the raw file only as totals; one division each surfaced them.

2. Why vectorized beats a loop

   You could write a for loop over rows to compute these, but you shouldn't. df["RS"] / df["G"] hands the whole operation to pandas, which runs it in fast compiled code and returns a new column aligned to the index. It's shorter, faster, and far less error-prone. The same idea extends to any arithmetic: add, subtract, multiply, or combine several columns in one expression.

3. Chart two engineered columns at once

   A dumbbell chart draws two values per team — here runs scored and allowed per game — as two dots joined by a line, so the gap between them is the run differential. It's a compact way to show offense and defense together.

   import matplotlib.pyplot as plt
   
   top = df.sort_values("RunDiff/G", ascending=False).head(12).sort_values("RunDiff/G").reset_index(drop=True)
   y = range(len(top))
   fig, ax = plt.subplots(figsize=(8, 7))
   ax.hlines(y=list(y), xmin=top["RA/G"], xmax=top["RS/G"], color="#C2B7A1", lw=2)
   ax.scatter(top["RS/G"], list(y), label="runs scored / game")
   ax.scatter(top["RA/G"], list(y), label="runs allowed / game")
   ax.set_yticks(list(y)); ax.set_yticklabels(top["Abbr"]); ax.legend()
   fig.savefig("rate_dumbbell.png", dpi=144, bbox_inches="tight")

   

   For every team here the "scored" dot sits to the right of the "allowed" dot — the visual signature of a positive run differential — and the length of each connecting line is exactly the per-game margin we engineered. A chart like this is impossible from the raw totals; it needed the rate columns first.

Troubleshooting