Pharmaceutical Demand Planning: Balancing Forecast Accuracy, Expiration Risk, and Inventory Availability

At a clothing brand, a bad forecast can leave the wrong sizes in a warehouse until a markdown clears them. In pharmaceuticals, the same planning mistake has another dimension: a patient may not get a medicine when it is needed, while stock in the next building ages toward an expiry date. The planner is asked to prevent both outcomes with a forecast that will always be imperfect. That is why the job starts with a different question from “How accurate was last month’s number?” The useful question is “Which decision will this forecast change, and what happens if we are wrong?”

An aggregate forecast can look excellent while the actual network is in trouble. A 10 percent miss on a widely used medicine may be manageable if capacity is flexible and replenishment is fast. A smaller miss on a single-source, temperature-sensitive product may be far more serious. The relevant unit is often the presentation, strength, package, customer, location, and lot, not just the product family. Demand planning has to meet inventory in that level of detail, because that is where expiry, shipping rules, allocation, and patient access become real.

When shipments masquerade as demand

The first trap is using shipments as a clean measure of demand. During a shortage, a wholesaler may order more than usual to protect its own customers. A hospital may substitute another therapy. A distributor may have fulfilled only half its orders because supply was constrained. Shipments then reflect allocation decisions as much as patient need. A model trained only on those figures can learn the shortage and call it a new normal. Planners need a record of unfilled orders, backorders, substitutions, market events, channel inventory, and any changes in prescribing patterns. None of these signals is perfect. Together, they are more informative than a smooth historical curve.

The expiry clock changes the math

Forecasts also need two clocks. The commercial clock watches demand by week or month. The physical clock watches remaining shelf life by lot. A forecast of 1,000 packs next quarter is not enough if 600 available packs expire before they can reasonably be dispensed. For each lot, compare expected consumption before expiry with the quantity and location of stock. Then apply the customer’s minimum remaining shelf-life requirement and the actual transit time. Inventory can be technically unexpired yet commercially unusable. That distinction should be visible before a buyer releases another order.

First-expire, first-out is a useful operating rule, but it is not a complete plan. It helps choose which eligible lot to move first. It does not tell you whether a future production batch is too large, whether stock sits in the wrong region, or whether a customer would reject a short-dated shipment. A company can run FEFO perfectly and still write off inventory. It needs an expiry exposure forecast: what quantity will likely remain at each lot’s last usable date under a low, base, and high demand case? The low case often matters most for write-off risk, while the high case matters most for service risk.

Production has to see the downside

This is where demand and supply decisions should meet. Suppose a product sells about 100 packs per week, but demand has recently ranged from 60 to 140. A new batch of 2,000 packs may be cheap per unit because it spreads setup cost over more volume. Yet if the product has a short effective selling window, that cost advantage may evaporate into expiry losses. A smaller batch, a delayed start, or a different packaging sequence could be more economical. The choice depends on validated capacity, changeover cost, raw material shelf life, release lead time, and service commitments. A forecast alone cannot choose the answer; it has to feed a decision model with those constraints.

Planners should watch forecast bias separately from absolute error. A team can have reasonable average accuracy and still overforecast the same product every cycle. That repeated upward miss quietly builds aging stock. The opposite bias creates recurring shortages and emergency allocation. Measure both at the time the decision was made, using the forecast version that informed production or purchasing. Recomputing history with later knowledge makes the score look better and removes the lesson. For intermittent products, a percent error can become misleading when actual demand is tiny or zero; unit error, bias, service performance, and expiry exposure tell a fuller story.

The response to uncertainty should vary by medicine. A common product with several qualified sources may need a relatively lean buffer. A critical therapy with limited alternatives may justify more reserve inventory, even if holding cost is high. But the buffer must be tied to a specific risk: demand variability, production yield, quality release delay, supplier interruption, or transport disruption. If every risk is covered by a separate pile of stock, the network overinsures itself. If no one owns the service standard, it underinsures the products that matter most.

What the model cannot decide

There is a temptation to let a sophisticated model make this feel automatic. It can help detect a change in demand sooner, identify anomalous orders, and show how different assumptions change service and expiry outcomes. It cannot authorize an expiration date extension, waive a quality hold, or decide that a short-dated lot is acceptable for a patient. Those are controlled decisions. The US Food and Drug Administration notes that expiration dating is supported by stability data and that extensions require appropriate evidence and regulatory pathways. Treating an expiry date as a flexible planning variable would be an expensive category error.

The most useful planning meeting therefore has fewer slides on a single forecast number and more discussion of choices. Which products have rising demand and constrained capacity? Which lots will be stranded under the low case? Which accounts are likely to reject remaining shelf life? Which production orders can still be changed? Which decisions require quality, regulatory, medical, or commercial approval? An exception list that answers those questions can be short enough to act on. A dashboard full of every SKU in the network usually is not.

FDA’s current shortage guidance also makes clear that supply disruptions can come from manufacturing and quality problems, delays, discontinuations, and demand changes. That is a reminder that “better forecasting” alone cannot solve availability. It can give operators more time to see a shortage forming and choose among qualified options. The better ambition is to make risk visible early enough that a company can protect patients without flooding the channel with inventory that may never be usable. Accuracy matters. So do the clock, the lot, the release date, and the real cost of being wrong.