Plan.md

Here's a plan built around the five stage files, with the pedagogical beat each stage is meant to hit.

Stage 0 — The hook (10 min)

Run the finished script, show the forecast plot, close it. No typing. "In two hours this is yours."

Stage 1 — Load and inspect (20 min)

Goal: nothing is real until you've looked at it.

• read_csv("day.csv") — deliberately without parse_dates first. Run df.info(), discover dteday is an object. Then re-read with parse_dates=["dteday"]. This is the single most important lesson in the session and it should hurt slightly.
• head(), shape, describe(), isna().sum()
• Rename to human columns: dteday → date, cnt → rides
• set_index("date"), then check completeness: compare len(df) against pd.date_range(start, end). 731 days, no gaps — but show them the check, because most data does have gaps.
• Blanks: the parse_dates argument, the rename dict.

Stage 2 — Clean and transform (25 min)

Goal: raw columns are rarely usable columns.

• Denormalize: temp * 41, hum * 100, windspeed * 67. Concrete, verifiable, and they see why reading the data dictionary matters.
• Decode codes into labels: season, weathersit, yr → 2011/2012, weekday → day names
• Derive: is_weekend, casual_share = casual / rides
• Sanity-check the result — plausible temperature range, no negative counts
• Write stage2_clean.parquet as the checkpoint everyone can rejoin from

Stage 3 — Explore and visualize (25 min)

Goal: this is where the insights live. Budget generously.

• Time plot of rides. Immediately visible: growth from 2011 to 2012, a yearly hump, and enormous day-to-day noise.
• Add a 7-day rolling mean. Noise disappears, trend appears. The most useful two lines of code in the whole session.
• Spot the late-October 2012 collapse → Hurricane Sandy. Teaches that outliers usually have causes, and that you look before you model.
• Groupby weekday, plot casual and registered separately. The payoff moment: registered ridership peaks midweek, casual peaks on weekends. Two populations hiding in one column.
• Groupby month → the seasonal shape, cleanly.
• Scatter rides vs temp → clearly non-linear, falls off at both extremes.
• Switch to hour.csv, groupby hr, split by workingday → the commuter double-peak versus the weekend midday hump. Best single visual you'll show; save it for last.

Stage 4 — First forecast (20 min)

Goal: a forecast is only meaningful next to a baseline.

• Split: train on everything except the final 28 days, test on those 28. Emphasize why the split is chronological, not random.
• Three predictions, each one line:
  1. training mean
  2. naive — last observed value
  3. seasonal naive — the value from 7 days ago
• Compute MAE for all three. Seasonal naive wins, and they can see why from the weekday plot they made in stage 3.
• Plot all three against actuals.
• Stretch prompt: scale seasonal naive by the year-over-year growth ratio. Does MAE improve?

Stage 5 — Close (10 min)

What you deliberately skipped and what it's called — stationarity, ACF, ARIMA, prediction intervals — so they know the vocabulary to search for. Point at the FPP book. Suggested homework: rerun the whole pipeline on hour.csv, or on the Seoul dataset where the columns are different enough to be real practice.

Scope discipline

Cut these ruthlessly if you're behind: the temp scatter, the monthly groupby, the growth-ratio stretch. Protect the weekday split, the rolling mean, and the baseline comparison — those three carry the session.

Want me to turn this into the actual stage files now?