Every operations manager knows the two ways a forecast hurts. Too high, and cash sits in a warehouse as unsold stock, later cleared at a discount. Too low, and the shelf is empty on the one weekend everyone wanted the product. Time-series forecasting for sales and demand won't eliminate either – nobody can predict the future – but a good forecasting process makes both rarer and, just as important, tells you how uncertain each number is.
This guide covers the main approaches – classical models like ARIMA, Prophet-style decomposition models and machine learning with gradient boosting – along with the evaluation habits and pitfalls that matter more than the choice of model.
What makes demand forecasting different
A time series is a sequence of values ordered in time: daily sales of one product in one store, weekly orders for a spare part, hourly calls to a support centre. Three properties make it different from ordinary prediction:
- Order matters. You can't shuffle the rows. Yesterday's value is informative about today's.
- Patterns repeat at several scales. Day of week, month, holidays, paydays, seasons.
- The future can't leak into the past. Any feature or test that accidentally uses information from after the forecast date makes results look far better than they will be in reality.
Before modelling, decompose the series mentally into trend (long-run direction), seasonality (repeating cycles), events (promotions, holidays, stock-outs) and noise. Most forecasting methods are different ways of estimating those pieces.
Baselines: the forecasts to beat
Always start with naive forecasts. They are embarrassingly hard to beat on some series.
- Naive: tomorrow equals today.
- Seasonal naive: next Monday equals last Monday; next December equals last December.
- Moving average: the mean of the last few periods.
If a sophisticated model can't beat seasonal naive on your data, it isn't adding value, no matter how impressive it sounds.
Classical models: ARIMA and exponential smoothing
ARIMA (AutoRegressive Integrated Moving Average) models a value as a function of its own past values and past forecast errors, after differencing the series to remove trend. Seasonal ARIMA (SARIMA) adds seasonal terms. Exponential smoothing (including Holt-Winters) forecasts with weighted averages that give more weight to recent observations, with separate components for level, trend and seasonality.
Strengths: well-understood statistics, prediction intervals out of the box, good on single, stable series with clear patterns. Weaknesses: one model per series, which gets unwieldy with thousands of products; limited ability to use external drivers like price and promotions; and they struggle with multiple overlapping seasonalities.
Prophet-style decomposition models
Prophet, released as open source by Meta, popularised a pragmatic approach: model the series as a sum of a trend curve, seasonal components (weekly, yearly) and holiday effects, fitted in a way that analysts can tune with intuitive settings. Similar additive models exist in other libraries.
They shine on business series with strong weekly and yearly seasonality, missing days and known holidays, and they are approachable for analysts without deep statistics training. They are not a universal answer, though: on many benchmarks well-configured classical models or machine learning models do as well or better, and the trend extrapolation can behave oddly over long horizons. Treat them as one candidate, not a default.
Sales forecasting with machine learning: gradient boosting
For retailers and distributors with many products, locations and drivers, global machine learning models have become a strong choice. Instead of one model per series, you train a single gradient boosting model (LightGBM, XGBoost, CatBoost) across all series, turning the problem into supervised learning with features such as:
- Lag features – sales 7, 14 and 28 days ago.
- Rolling statistics – mean and standard deviation over recent windows.
- Calendar features – day of week, week of year, holiday flags, days to the next holiday.
- Drivers – price, discount depth, promotion flags, marketing spend, weather if relevant.
- Identifiers – product category, store, region.
A global model learns shared patterns: a new product can borrow behaviour from similar ones, and promotion effects are estimated from thousands of examples instead of a handful. The cost is careful feature engineering and a strict discipline about leakage – every lag and rolling feature must only use data available at the moment the forecast would have been made.
How to evaluate forecasts honestly
Random train/test splits are wrong for time series. Use backtesting (also called time-series cross-validation): train on data up to a cutoff, forecast the next horizon, move the cutoff forward, repeat. Score every model on exactly the same windows.
Choose metrics that match the decision:
| Metric | What it tells you | Watch out for |
|---|---|---|
| MAE | Average absolute error in units | Not comparable across products |
| MAPE | Average percentage error | Explodes when actual sales are near zero |
| WAPE | Total absolute error / total actuals | Good aggregate view for retail |
| Bias | Systematic over- or under-forecasting | Can hide behind a decent MAE |
| Quantile loss | Quality of P10/P90 ranges | Needed when you plan safety stock |
Bias deserves its own line. A model that is slightly but consistently too low will cause stock-outs month after month, even if its average error looks fine.
Demand forecasting pitfalls
- Forecasting sales instead of demand. When a product was out of stock, sales were zero but demand wasn't. Flag or correct stock-out periods, or the model learns that demand vanishes.
- Ignoring promotions. A promo spike treated as normal demand inflates the baseline; future promotions get under-forecast.
- Leakage. Using a feature like "monthly total sales" that includes days after the forecast date.
- One number, no range. Planning needs uncertainty. Produce intervals or quantiles, not just a point forecast.
- Wrong level of detail. Forecasting each product per store per day can be pure noise; sometimes forecasting weekly or by category and then splitting down works better.
- Overriding without tracking. Planners will adjust forecasts. Record the adjustments and measure whether they help.
A practical roadmap
- Clean the history: mark stock-outs, promotions and data errors.
- Build seasonal naive and exponential smoothing baselines.
- Set up backtesting with WAPE and bias.
- Add a global gradient boosting model with lags, calendar and price features.
- Produce prediction intervals and connect them to inventory rules.
- Monitor accuracy weekly and retrain on a schedule.
Final word
Good time-series forecasting for sales and demand is less about finding the cleverest model and more about honest testing, clean history and forecasts that carry their uncertainty. Get those right and even a simple model becomes something planners trust – which, in the end, is the only forecast that changes decisions.
