An inventory shortage in a commercial distribution operation is a cost event. You lose the sale, you pay expediting fees to backfill, you absorb the customer service overhead of managing the failure. Painful, but recoverable. The same structural failure in a humanitarian supply chain, a relief depot running out of oral rehydration salts during a cholera outbreak or a staging area exhausting water purification tablets after a flood, is a different category of consequence entirely.
Despite the difference in stakes, the underlying forecasting logic that prevents both failures follows a common structure. The signal-to-noise problems are the same. The data quality challenges are similar. The failure modes, if you trace them back to their origin, are often structurally identical. Which is why it is worth understanding the forecasting problem in both contexts.
Why Shortages Happen in Operations With Enough Aggregate Supply
Most inventory shortages in logistics operations are not caused by absolute supply scarcity. The aggregate supply exists somewhere in the network. The failure is a distribution problem: the inventory is in the wrong location relative to where demand occurs, and the replenishment chain cannot move it fast enough when the gap becomes apparent.
This is the localization problem in inventory management. A regional distributor can have adequate total inventory across five warehouses while simultaneously stocking out at a single location that serves a high-demand corridor. The stockout is real even though no aggregate shortage exists, because the inventory in the other four locations is not accessible to the customers experiencing the gap.
In humanitarian contexts, the localization problem is compounded by the fact that replenishment routes may be disrupted exactly when demand spikes, which is the same event that caused the demand spike in the first place. A flood that increases demand for clean water also closes the roads that would normally be used to replenish the depot. The distribution failure and the demand surge are correlated, not independent events, which means standard safety stock calculations that treat demand volatility and supply disruption as independent risks will systematically underestimate the required buffer.
The Leading Indicators That Precede Stockouts
Reactive inventory management waits for stock levels to fall below a reorder point before triggering replenishment. The reorder point is set based on historical demand and lead time estimates. When those estimates are accurate and conditions are stable, this approach works reasonably well. When demand patterns change or lead times become unpredictable, the reactive model generates a consistent lag between when the shortage is becoming inevitable and when the replenishment order is triggered.
Predictive inventory management looks for leading indicators that a shortage is developing before the stock level falls to the reorder threshold. The most useful leading indicators depend on what you are managing.
For commercial distribution operations, the leading indicators for local stockouts include: unusual demand velocity in the past 5-7 days (a site that typically draws 40 units per day pulling 65 units per day for several consecutive days), a replenishment order that is overdue (a scheduled replenishment that has not arrived on time), and concurrent demand pressure at multiple sites served by the same upstream source (which signals potential supply constraint upstream before it is visible in any single site's inventory level).
For humanitarian supply chains, the leading indicators include all of the above plus infrastructure risk signals: weather events that may disrupt replenishment routes, population movement data that indicates demand is concentrating in a location (more people arriving at a community than the inventory buffer was sized for), and partner organization activity (if another organization stops taking draws from a shared inventory pool, it may mean their own supply chain is failing, not that demand has dropped).
Building a Forecast Model for Depot-Level Inventory
Forecasting at the depot or distribution point level requires a different approach than forecasting at the aggregate network level. Aggregate forecasts benefit from the law of large numbers: individual demand variability cancels out across many locations, and the aggregate demand is smoother than any individual site's demand. A depot-level forecast is based on low-volume, high-variance demand data from a single location, which is fundamentally harder to forecast accurately.
The approaches that work best at depot level are not sophisticated statistical models. They are models that incorporate domain knowledge about what drives demand at that specific location. For a commercial distribution warehouse, that might be: the seasonal pattern for the products stored there, the order patterns of the top 10 customers served from that depot, and any known demand commitments (contracted volumes, planned events, known large orders in the pipeline).
For a humanitarian field depot, the demand drivers are: current population count in the served area, current consumption per capita for the relevant supplies, the anticipated duration of the crisis phase, and whether any demand-boosting factors are present (disease outbreak, extreme heat, water source contamination). The uncertainty range on these estimates is larger than for commercial demand forecasting, which means the appropriate inventory buffer at a humanitarian depot is larger relative to average demand than for a comparable commercial setting.
The target safety stock for a field depot is not calculated the same way as for a commercial DC. It should account for the correlated disruption risk: the probability that the replenishment route is disrupted is not independent of the probability that demand spikes. Building in a buffer that covers both the demand variability and the supply disruption scenario simultaneously requires a larger reserve than treating them as independent.
When to Override the Forecast
Every inventory forecasting system requires a mechanism for human override, because quantitative forecasting models cannot capture all the signals that an experienced logistics coordinator can see. An NGO field coordinator who hears from a local contact that 2,000 people are expected to arrive at the depot site over the next three days has information that no model has. An experienced commercial dispatcher who knows that a major local manufacturer is starting a promotional period next week has information that predates the demand data that would make it visible to a model.
The practical design is not a fully automated forecast that dispatches replenishment orders without human review, but a system that surfaces the forecast and its confidence interval to the human decision-maker, flags cases where the model's confidence is low (high recent demand volatility, unusual activity that does not match historical patterns), and makes it operationally easy to override the model when the coordinator has information the model lacks.
An override should be an exception workflow that captures why the override was made, so that the information informing the override can be reviewed retrospectively. If the override turned out to be right and the model wrong, the override rationale is a signal that the model needs a new input. If the override turned out to be wrong, that is useful feedback for calibrating human judgment in future decisions. Both outcomes contribute to improving the overall forecasting system over time.
The gap between "forecast says we are fine" and "operations says we are running short" almost always contains actionable information, whether the operations perspective was right or wrong. Capturing that gap systematically, rather than treating each stockout as an isolated incident, is how inventory forecasting improves in a real logistics environment rather than just in academic settings.
