Stockout Probability Scoring

The model calculates a precise probability score (e.g., 87% chance of stockout within 4 hours) rather than a binary in-stock/out-of-stock flag. It ingests real-time inventory velocity, pending order volume, and supply chain lead times to project depletion curves. This allows the system to act before the shelf is empty, suppressing visibility when the probability crosses a defined risk threshold.

Scoring accuracy depends on fusing heterogeneous data streams into a unified risk calculation. Key inputs include:

• Current on-hand inventory and committed stock
• Real-time sales velocity at the specific SKU-location level
• Inbound shipment ETAs and historical carrier reliability
• Seasonality and demand spikes from local events
• Substitute availability to model demand transference The model weights these signals dynamically based on product lifecycle stage.

The probability score directly feeds into availability-weighted relevance engines. When a product's stockout probability exceeds a configurable threshold (e.g., >70% within the shipping window), the system automatically:

• Suppresses the item in search results and recommendation carousels
• Boosts overstocked substitutes with low stockout risk
• Adjusts dynamic pricing to slow demand velocity This prevents the broken user experience of clicking on unavailable products.

Unlike enterprise-wide inventory dashboards, this scoring operates at the fulfillment node level—individual stores, micro-fulfillment centers, or dark stores. A product may have a 5% stockout probability in one zip code and 92% in another. This granularity enables geofenced assortment rules that show different catalog availability based on the user's proximity to inventory, ensuring only locally purchasable items are surfaced.

Advanced implementations weight the probability score by the business cost of a stockout for that specific item. High-margin items or those with strong demand transference to substitutes may tolerate higher risk thresholds. Conversely, contractually obligated products or those with no viable substitutes trigger suppression at much lower probability scores. This transforms a purely statistical output into a profit-optimizing decision signal.

The model continuously recalibrates by comparing predicted stockout probabilities against actual outcomes. Key calibration metrics include:

• Brier Score: Measures the accuracy of probabilistic predictions
• Expected Calibration Error: Quantifies systematic over- or under-confidence
• Precision-Recall at threshold: Evaluates the trade-off between premature suppression and missed stockouts This feedback loop ensures the scoring remains reliable as demand patterns shift.

A ranking signal that directly consumes stockout probability scores to adjust product visibility in search and recommendation results. Items with a high probability of imminent stockout are down-weighted or suppressed entirely, while items with deep inventory receive a boost.

• Prevents customers from seeing products they cannot purchase
• Reduces bounce rates from frustrating 'out of stock' landing pages
• Integrates as a multiplicative factor in final ranking formulas

A predictive framework that estimates which alternative product a customer will purchase if their first choice stocks out. Stockout probability scores trigger preemptive substitution logic before the stockout actually occurs.

• Uses product affinity graphs to identify nearest substitutes
• Enables seamless 'pre-switching' of recommendations
• Preserves revenue by redirecting demand rather than losing it

A complementary mechanism that automatically increases visibility of overstocked or perishable items when stockout probability is near zero. Works in tandem with stockout scoring to balance sell-through across the catalog.

• Prevents dead stock accumulation
• Accelerates clearance of aging inventory
• Applies exponential boost factors as days-on-hand increases

A dense vector representation of a product that encodes real-time stock status alongside static attributes. Stockout probability scores are concatenated into the embedding space, allowing retrieval models to natively filter unavailable items.

• Eliminates post-retrieval filtering latency
• Enables approximate nearest neighbor search with inventory constraints
• Trained via multi-task learning on relevance and availability objectives

A short-term forecasting model that translates real-time downstream signals—point-of-sale data, website clicks, cart additions—into upstream inventory decisions. Stockout probability scoring is the output layer that converts demand forecasts into actionable risk metrics.

• Operates on hourly or sub-hourly time windows
• Detects demand spikes before they deplete safety stock
• Feeds both merchandising and supply chain systems

A reinforcement learning agent that incorporates remaining stock levels into its reward function. The bandit naturally ceases exploration of items with high stockout probability, avoiding the opportunity cost of promoting soon-to-be-unavailable products.

• Balances exploration-exploitation with inventory constraints
• Reward shaping penalizes recommending near-stockout items
• Converges to optimal assortment under supply limitations