65. How to Forecast Sales with Time Series in Python (pandas and sklearn)

65.6 Train/Test Split by Time (Never Shuffle Dates)

By Ravindra Bagale, trainer · LinkedIn · Last updated: 31 October 2026

Why. Chapter 31 taught train_test_split. For forecasting, random shuffle leaks the future. October can train a model that later “predicts” September.

What. Cut by date: all train months < all test months. No shuffle.

How.

Worked example A — Correct time split.

import pandas as pd

df = pd.read_csv("mauli_monthly_sales.csv", parse_dates=["month"]).sort_values("month")
split = pd.Timestamp("2026-05-01")
train = df[df["month"] < split]
test = df[df["month"] >= split]
print("train", train["month"].min().date(), "→", train["month"].max().date(), "n=", len(train))
print("test ", test["month"].min().date(), "→", test["month"].max().date(), "n=", len(test))

Output you should see.

train 2023-11-01 → 2026-04-01 n= 30
test  2026-05-01 → 2026-10-01 n= 6

Worked example B — Shuffle vs time MAE (same lag model).

On this lab, time-based test MAE ≈ ₹7,258. A shuffled holdout gave MAE ≈ ₹13,241. Numbers move when the split is dishonest — managers need the time score.

# idea only: compare MAE from time split vs train_test_split(..., shuffle=True)
# Always report the time-based MAE for forecasts.

Worked example C — Interview line.

“I never shuffle a forecasting split. I train on the past and test on the future so the score matches production.”