Demand Forecasting in Python — Core Concepts

Demand forecasting feeds into nearly every supply chain decision: how much to order, when to manufacture, how many staff to schedule, and how much warehouse space to lease. Getting it right reduces both waste and lost sales.

Time series decomposition

Sales data over time typically contains three components:

• Trend — the long-term direction. Are sales growing 5 percent per year or declining?
• Seasonality — repeating patterns at fixed intervals. Ice cream sales peak every summer.
• Residual — the random noise left over after removing trend and seasonality.

Python’s statsmodels can decompose a series in one call, revealing which component dominates. If seasonality is strong, your model must capture it. If the trend dominates, simpler methods work fine.

Forecasting methods ranked by complexity

Moving average

Average the last N periods. A 4-week moving average smooths out weekly noise. It reacts slowly to sudden changes but requires zero configuration.

Exponential smoothing

Recent observations get more weight than older ones. Triple exponential smoothing (Holt-Winters) captures both trend and seasonality. Python’s statsmodels.ExponentialSmoothing fits this with automatic parameter selection.

ARIMA

AutoRegressive Integrated Moving Average models the relationship between an observation and lagged values. ARIMA(p, d, q) has three parameters: autoregressive order, differencing degree, and moving average order. Seasonal ARIMA (SARIMA) adds seasonal equivalents.

Choosing p, d, q manually is tedious. The pmdarima library automates it with auto_arima, which tests combinations and selects the best fit by information criteria (AIC).

Prophet

Facebook’s Prophet handles daily data with multiple seasonalities (weekly, yearly), holidays, and trend changepoints. It requires minimal tuning and works well for business data with known calendar effects.

Machine learning approaches

Gradient boosting (XGBoost, LightGBM) treats forecasting as regression. Features include lagged values, day of week, month, promotional flags, and weather. These models capture non-linear relationships that ARIMA misses but need more data and careful feature engineering to avoid overfitting.

Evaluation metrics

• MAE (Mean Absolute Error) — average magnitude of errors. Easy to interpret: “off by 12 units on average.”
• MAPE (Mean Absolute Percentage Error) — error as a percentage of actual values. Useful for comparing across products with different scales. Breaks down when actuals are near zero.
• RMSE (Root Mean Squared Error) — penalizes large errors more heavily. Good when big misses are costlier than small ones.
• Weighted MAPE — weights errors by revenue contribution. A 20 percent miss on a high-value SKU matters more than a 30 percent miss on a low-value one.

Always evaluate on a held-out test period, not on the training data. Walk-forward validation — train on data up to month T, forecast month T+1, slide forward — gives the most honest assessment.

Practical considerations

Data quality matters more than model sophistication. A simple exponential smoothing model on clean data beats a complex neural network on garbage data. Check for missing values, duplicated entries, and returns logged as negative sales.

Granularity tradeoffs exist. Forecasting at the SKU-store-day level is noisy. Aggregating to SKU-region-week smooths the signal but loses local patterns. Start at a higher level and disaggregate later.

External signals improve accuracy. Weather, promotions, economic indicators, and competitor actions all influence demand. Adding a promotional flag alone can cut forecast error by 10-20 percent for retail.

A common misconception

People expect forecasts to be right. They should not. The goal is to be less wrong than the alternative (gut feeling, last year’s number). A forecast with a 15 percent error that consistently beats manual planning by 30 percent is a success.

The one thing to remember: Demand forecasting combines time series decomposition with methods ranging from moving averages to machine learning, and Python’s ecosystem makes it practical to test multiple approaches and pick the one that minimizes error for your specific data.