ML & global forecasting models

Classic methods fit one model per item: thousands of SKUs means thousands of small models, each learning only from its own short history. A global model flips this. One model learns from all items together, so a new or sparse item can borrow patterns from similar ones.

The most common version in practice is gradient boosting (LightGBM, XGBoost). You turn forecasting into a table of features and let the model learn how they relate to sales:

• history features: sales last week, the same week last year, rolling averages
• calendar features: day of week, festivals, paydays, school holidays
• business features: price, discount depth, promotion flags, display space

That is why these models shine when demand is driven by things you can see coming, such as promotions and price changes. In the Walmart-based M5 competition (2020), most of the top entries were LightGBM-based. In the earlier M4 competition on mostly business and economic series without those drivers, combinations of statistical methods and a hybrid of exponential smoothing with a neural network did best.

Deep learning models such as DeepAR, N-BEATS and the Temporal Fusion Transformer push the global idea further, and newer foundation models are pre-trained on huge collections of time series so they can forecast a new series with little or no training.

They have limits. Tree models don’t extrapolate trends beyond what they’ve seen, every model needs clean data and careful backtesting, and on stable items without drivers a well-tuned exponential smoothing model is very hard to beat.

Step 1: turn history into a table

Each row is one item, at one location, in one period. The target is the sales you want to predict; the columns are features known at the moment the forecast is made.

Typical feature groups: recent sales and rolling averages, the same period last year, calendar effects (day of week, month, festivals, paydays), price and promotion plans, product attributes (category, brand, pack size) and location attributes.

Step 2: avoid leakage

Leakage means letting the model see information it won’t have when forecasting for real, which makes test results look far better than reality. The two classic forms:

• Horizon leakage: if you forecast four weeks ahead, “sales one week before” isn’t known yet for weeks two to four. Either use only lags of at least four weeks, or forecast one week at a time and feed the predictions back in (recursive forecasting).
• Future information: a promotion that happened but wasn’t planned when the forecast was made, or a rolling average that includes the week being predicted.

Step 3: choose what the model minimizes

• Squared error targets the mean, the usual choice for replenishment.
• For sales counts with many zeros, a Tweedie or Poisson loss fits the data better.
• For a P90 forecast, quantile (pinball) loss trains the model to hit that quantile directly.

Step 4: test it like a forecast

Split by time, never at random: train on the past, test on the future, and roll the cut-off forward several times (rolling-origin backtesting). Compare against strong simple baselines, such as seasonal naive and ETS, at the horizon the decision uses, and check bias by segment, not only overall accuracy.

Why one model for many items works

A single item may have only two years of noisy history, but ten thousand items together hold millions of observations of how prices, promotions, seasons and holidays move demand. A global model learns those shared effects once and applies them everywhere, including to new or sparse items. To make items comparable, sales are often scaled per item (for example, divided by each item’s average) before training.

Lessons from the M5 competition

The M5 competition used hierarchical Walmart sales data with prices, promotions and calendar events. The organizers’ findings (Makridakis and colleagues, 2022) included:

• Gradient-boosted trees trained across many series dominated the top places.
• Explanatory variables such as prices and events improved accuracy.
• Combining several models, trained on different subsets or with different settings, beat single models.
• Validation that mimicked the real forecasting setup was essential.

Deep learning and foundation models

Neural approaches such as DeepAR (probabilistic forecasts from a recurrent network), N-BEATS and the Temporal Fusion Transformer also learn across series and can output full distributions. Foundation models for time series are pre-trained on very large collections of series and can forecast a new one with little or no training. They are promising for cold starts and quick prototypes, but on rich retail data with known drivers, well-tuned tree models remain hard to beat, so test before switching.

What breaks in production

• Trends: trees predict within the range they have seen, so a fast-growing item gets flattened. Model the change rather than the level, or blend with a method that handles trend.
• Stability: ML forecasts can jump from week to week and make plans nervous. Smooth the output or limit near-term changes.
• Explainability: planners trust what they understand. Feature-importance tools, such as SHAP values, help explain why a forecast moved.
• Maintenance: feature pipelines, retraining, monitoring and data drift take real effort. A model that is slightly more accurate but breaks every month may not be worth it.