Granular Forecasting

Generates predictions at the stock-keeping unit (SKU) × location intersection rather than at product-category or regional aggregates. This atomic resolution captures local demand heterogeneity—a specific store's preference for a particular flavor or size—that gets averaged away in top-down forecasts. The approach directly models cannibalization effects where one product's promotion steals demand from a sibling SKU in the same aisle, a dynamic invisible to category-level models.

Ensures mathematical consistency across all aggregation levels through forecast reconciliation algorithms. A granular forecast for every individual store must sum exactly to the regional forecast, which must sum to the national total. Techniques like optimal reconciliation using the MinT (Minimum Trace) estimator adjust base forecasts to satisfy all aggregation constraints simultaneously while minimizing the mean squared error at every level of the hierarchy.

Addresses the statistical challenge of sporadic demand with frequent zero-sale periods common in spare parts, luxury goods, and slow-moving SKUs. Traditional exponential smoothing fails on these series. Granular approaches employ specialized methods:

• Croston's method: Separately models demand intervals and demand sizes
• Negative binomial distributions: Captures overdispersed count data
• Zero-inflated models: Explicitly models the excess zero probability mass

Solves the new product forecasting problem where no sales history exists. Granular systems leverage transfer learning from analogous products using attribute-based embeddings—mapping a new SKU's features (price point, category, brand, seasonality profile) into a latent space where similar items cluster. This enables probabilistic demand estimates from day one, refined through online learning as the first transactions stream in.

Incorporates high-dimensional exogenous features at the granular level that lose meaning when aggregated:

• Price elasticity per SKU per store
• Promotional calendars with lift coefficients
• Local events (concerts, weather, road closures)
• Competitor openings within a trade area
• Social media sentiment geo-tagged to specific locations These covariates feed into deep architectures like the Temporal Fusion Transformer, which learns variable selection and attention patterns specific to each time series.

Outputs full probability distributions rather than point estimates at the granular level, enabling risk-aware inventory decisions. Techniques include conformal prediction for distribution-free coverage guarantees and quantile regression for estimating specific service-level targets (e.g., the 95th percentile). Proper scoring rules like the Continuous Ranked Probability Score (CRPS) evaluate whether the predicted distributions are well-calibrated—neither overconfident nor underconfident—across thousands of individual time series.

A structured collection of time series organized by aggregation constraints, such as product SKUs rolling up to categories or regional demand summing to national totals. Granular forecasting operates at the bottom level of this hierarchy.

• Top-down: National forecast disaggregated to stores using historical proportions
• Bottom-up: Individual SKU forecasts summed to create category predictions
• Middle-out: Combines both approaches for optimal accuracy

The key challenge is ensuring forecast reconciliation—mathematical coherence where lower-level predictions sum exactly to higher-level totals.

The process of adjusting forecasts generated at different levels of a hierarchy to ensure mathematical coherence. Without reconciliation, the sum of individual store forecasts will rarely equal the national forecast.

• MinT (Minimum Trace): Optimal reconciliation method that minimizes forecast error variance
• OLS reconciliation: Simple ordinary least squares adjustment
• BU (Bottom-Up): Uses only granular forecasts, summing upward

Reconciliation is critical because aggregated forecasts often have lower relative error than granular ones, but granular forecasts capture local patterns that aggregates miss.

The use of real-time downstream data—such as daily point-of-sale signals, weather patterns, or social media sentiment—to adjust short-term granular forecasts. This dramatically reduces latency in supply chain response.

• Shortens the planning horizon from weeks to days or hours
• Incorporates leading indicators rather than lagging shipment data
• Enables micro-segmentation of demand by store, channel, and time-of-day

Demand sensing transforms granular forecasting from a passive planning exercise into an active operational tool for immediate replenishment decisions.

A demand pattern characterized by sporadic demand occurrences interspersed with frequent zero-demand periods. This is the dominant pattern at granular levels for slow-moving SKUs, spare parts, and aftermarket supply chains.

• Croston's Method: Separately estimates demand interval and demand size
• SBA (Syntetos-Boylan Approximation): Corrects Croston's positive bias
• TSB (Teunter-Syntetos-Babai): Handles obsolescence risk with dynamic probability updates

Standard exponential smoothing fails catastrophically on intermittent data, making specialized methods essential for accurate granular forecasts of long-tail inventory.

A change in the distribution of input features between the training environment and the production environment. At the granular level, this is particularly dangerous because local patterns can shift silently.

• A store's demographic catchment changes due to new housing developments
• A product's seasonal pattern shifts due to competitor exit
• A regional promotion strategy alters baseline demand structure

Detection methods include monitoring KL divergence between training and inference feature distributions. Mitigation requires online learning or frequent retraining cycles to prevent silent model degradation.

A machine learning paradigm where the model updates continuously as new data streams arrive, enabling granular demand forecasts to adapt in near real-time to shifting consumer behavior.

• Stochastic gradient descent updates model weights with each new observation
• Sliding windows discard stale data to prioritize recency
• Forgetting factors exponentially decay the influence of older observations

Online learning is essential for granular forecasting because local demand patterns evolve constantly—a store's customer base, a product's lifecycle stage, and competitive dynamics all shift over time.