Time-Series Versus Causal Forecasting: Choosing The Right Method

Demand Forecasting Estimating future demand for products using sales history, seasonality, promotions, trends, and external factors. This article compares time-series and causal forecasting approaches, explains when each is preferable, and outlines hybrid and machine-learning options for complex catalogs.

Choosing the right forecasting method is a function of data availability, SKU behavior, and the business decision the forecast informs. Time-series methods rely on historical patterns, while causal methods model demand as a function of external drivers. Machine learning sits between these approaches and can incorporate both patterns and causal signals.

Time-Series Forecasting: Strengths And Limits

Time-series models use past demand values to predict future demand. They perform well when demand follows persistent patterns such as trends and seasonality and when external drivers are stable or already reflected in history.

• Strengths: Low data requirements beyond reliable historical demand, interpretable components (trend, seasonality), fast to compute and easy to scale across SKUs.
• Limits: Underperform when demand is driven by external events (promotions, price changes) or when product life cycles are short and history is limited.

Causal Forecasting: When To Use It

Causal models include explanatory variables — price, promotion intensity, marketing spend, macroeconomic indicators, or weather. Use causal methods when you can quantitatively link those drivers to demand and when those drivers will change in the forecast horizon.

• Strengths: Captures the effect of planned actions (like promotions) and external shocks; better for scenario planning and "what-if" questions.
• Limits: Requires reliable driver data and careful model specification; risk of omitted-variable bias if important drivers are missing.

Machine Learning And Hybrid Approaches

Machine learning models (random forests, gradient boosting, deep learning) can combine time-series features and causal variables, handling complex interactions. Hybrid strategies often yield the best operational results: use automated time-series models for stable SKUs and ML or causal models for promotional, new, or highly-variable SKUs.

• Strengths: High predictive power when large, clean datasets exist; adaptable to many features.
• Limits: Greater compute and data engineering needs, potential lack of interpretability, and risk of overfitting.

Choosing A Method: Practical Rules

Decisions should be pragmatic and tied to SKU segmentation and the forecasting horizon.

• Short Horizon, High Frequency: Time-series models with daily/weekly granularity typically work best for labour and picking planning.
• Promotion-Driven SKUs: Causal or ML models that include promotional variables outperform pure time-series models.
• New Products: Use judgmental forecasting with analogues and accelerate to causal or ML as data accumulates.
• Intermittent Demand: Special methods (Croston’s or bootstrapping) or a combination of judgmental adjustment and statistical smoothing.

Evaluating And Deploying Models

Standardize evaluation with holdout sets and operational KPIs (fill rate, inventory turns). Track model drift: re-evaluate parameters after major assortment changes or when forecast error systematically increases. Keep a fallback baseline (e.g., last-period or moving average) and compare new models against it before rollout.

• Backtest: Validate on historical periods that include promotions and other events.
• Monitor: Implement automated alerts for error spikes and retraining triggers.
• Explainability: Prefer models that allow you to explain drivers behind major forecast changes to commercial teams.

In short, the Demand Forecasting method you choose should match SKU behavior, data availability, and the operational decision it informs. Time-series methods are efficient for stable demand; causal and machine-learning approaches are superior when external drivers or promotions materially alter demand. Hybrid strategies and ongoing monitoring usually deliver the best, most practical outcomes.