A naive forecast gets you 80% of the way to a complex model

🏢 The situation

The operations director is in a panic: the warehouse is stuffed with winter goods, while popular items run out 3 days before restocking. Purchasing is planned "same as last month".

Lena:

"Classic demand forecasting. We need a 4-week-ahead sales forecast for the top 2000 SKUs. The peculiarity of time series: the order of the data is information, and all the usual validation habits work differently here. Start with the naive baselines — in forecasting they are harder to beat than you think."

🎯 Your task

1. Understand the series components: trend, seasonality, noise.
2. Build features: lags, rolling windows, the calendar.
3. Validate honestly (no peeking into the future) and beat the naive baseline.

---

📚 Theory

Time series components

• Trend — the long-term direction (the store is growing).
• Seasonality — repeating cycles: weekly (weekend peaks), yearly (Christmas trees in December).
• Events — holidays, promos, price changes.
• Noise — the unexplained part.

Naive baselines — respect them

1. Naive: tomorrow = today.
2. Seasonal naive: next Wednesday = last Wednesday.
3. Moving average over 4 weeks.

In retail, seasonal naive often delivers 80% of a complex model's quality. A model that can't beat it is wasted time.

Two approaches to forecasting

1. Classic series models (ARIMA, ETS, Prophet): good for a single series with clear structure.
2. ML on tabular features (boosting): turn the series into a table of "features → next period's value". For thousands of SKUs with one model — the industry standard (this is how the M5 competition was won).

Features for the ML approach

• Lags: sales 1, 7, 14, 28 days ago.
• Rolling statistics: mean/median/std over 7/28 days (⚠️ the window ends before the forecast day).
• Calendar: day of week, month, holidays, days to/after a holiday.
• SKU attributes and events: category, price, promo flag.

Validation: forward only

Random K-fold is forbidden: the model would see the future. Use a rolling/expanding window: train up to T → forecast [T, T+4 weeks] → shift → repeat. This mimics real usage.

Metrics: MAE/WAPE (error as a share of volume) — clear to purchasing; MAPE is dangerous with zero sales.

---