Demand forecasting is one of the most decisive capabilities for the efficiency and profitability of any organization. However, even companies that have already advanced in analytics and artificial intelligence often stumble over the same obstacles. Many of these mistakes are not in the models themselves, but in how data is structured, how variables are evaluated, or how forecasts are validated.
Below, we review the five most common forecasting mistakes and how to avoid them.
1. Believing that “more data” always improves forecasting
One of the most widespread myths is that adding more variables, more columns, or more data sources inevitably leads to better forecasts. In practice, the opposite often happens: more data does not mean more accuracy, especially if that data is poorly defined or does not provide real information to the model.
Irrelevant regressors introduce noise, distort patterns, and degrade performance. Worse still, they can cause a complex model to perform worse than a simpler one.
How to avoid it:
- Select only variables that make sense from a business perspective.
- Validate their impact through backtesting to prove they add real value.
- Use lightweight models such as LightGBM to identify which variables deserve to be included in more sophisticated models.
2. Working with insufficient or excessively granular historical data
This is a critical point: you cannot model what you cannot see. Time series without enough complete cycles (or with overly noisy structures) produce unstable predictions. This is especially common when forecasting at very granular levels (SKU, store, 15-minute intervals) with intermittent or sparse demand.
How to avoid it:
- Ensure a minimum of two full cycles to capture seasonality.
- Aggregate data when noise outweighs the signal.
- Use models designed for intermittent demand when necessary (such as Croston or Tweedie).
3. Building models without understanding the real demand drivers
Before training any model, the key question is: what truly drives this demand?
The best models combine internal patterns (trend, seasonality) with validated external regressors: prices, promotions, holidays, weather, macroeconomic indicators, or competitive signals. Projects that skip this step end up with blind models, unable to explain key behaviors or capture causal effects.
How to avoid it:
- Map and prioritize the factors that drive demand.
- Validate each regressor using rolling backtests.
- Remember that advanced M5 solutions demonstrated that external information significantly improves forecasts.
Forecasting stops being guesswork when cause–effect relationships behind each series are understood.
4. Using a single model for everything (or a different model for each series)
Both extremes are problematic. A single global model ignores heterogeneity across products, channels, or regions. Training thousands of independent models, on the other hand, is hard to maintain, costly, and not scalable. A balanced solution? Intelligent clustering.
How to avoid it:
- Group series by behavior, seasonality, product type, or context.
- Train specific models for each cluster.
- Or use a single model with cluster ID as a regressor, integrating series diversity into one architecture.
5. Failing to ensure hierarchical coherence and proper model validation
One of the most expensive mistakes is evaluating forecasts with incorrect methods: random validations (k-fold), misaligned metrics, or leakage between training and testing.
- Use rolling-origin cross validation instead of random K-Fold.
- Carefully review scalings, moving averages, and derived calculations to avoid leakage.
- Reconcile forecasts using top-down, bottom-up, or MinT approaches to ensure global coherence across SKU, category, store, and region levels.
How to avoid it:
- Use rolling-origin cross validation instead of random K-Fold.
- Carefully review scaling, moving averages, and derived calculations to avoid leakage.
- Reconcile forecasts using top-down, bottom-up, or MinT approaches to ensure overall consistency.
The most common errors in forecasting do not stem from technology, but rather from approach: how data is structured, which variables are included, how models are validated, and how hierarchies are constructed. Avoiding these five mistakes allows you to move from a reactive approach to a truly proactive one. When models cease to be black boxes and begin to reflect actual business behavior, forecasting becomes an advantage.