Every supply chain eventually faces a decision that must be made before the information needed to make it exists. A fashion buyer commits to winter coats in February. A semiconductor firm reserves foundry capacity quarters in advance. A bakery decides at four in the morning how many croissants to bake. In each case demand reveals itself only after the quantity is locked, and whatever is left over or missing carries a cost that cannot be undone.

The newsvendor problem is the formal model of this predicament. It is named for the street vendor deciding how many papers to buy each morning. It reduces inventory decisions to their essential structure: a single ordering opportunity, a random demand, and unequal penalties for guessing too high or too low. The answer does not come from intuition or from the forecast alone. It comes from the precise ratio between those two penalties.

That simplicity is deceptive. Newsvendor logic underpins base-stock policies, capacity planning, airline overbooking, supply contracts, and increasingly the machine learning systems that forecast demand at scale. Understanding its derivation is not an academic exercise. It is the grammar from which most inventory reasoning is built. What follows develops that grammar from the cost structure up. It then traces how the logic governs service levels and how it extends into the multi-period, multi-product networks that real supply chains operate.

Begin with two costs. The overage cost, Co, is what you lose on each unit ordered but not sold: purchase cost minus whatever salvage value remains. The underage cost, Cu, is what you lose on each unit of demand you fail to meet: the forgone margin, plus any goodwill or substitution penalty a stockout imposes. Everything else in the problem, including price, forecast, and supplier terms, matters only insofar as it shapes these two numbers.

The optimal quantity follows from marginal reasoning. Consider whether to add one more unit to an order of Q. That unit sells only if demand exceeds Q, which happens with probability 1 − F(Q), earning Cu. It goes unsold with probability F(Q), costing Co. Keep adding units while the expected gain exceeds the expected loss. The stopping point is where they balance: F(Q*) = Cu / (Cu + Co). This ratio is the critical fractile.

The result is notable both for what it says and for what it leaves out. The optimal order is not the forecast mean or the most likely outcome. It is a specific quantile of the demand distribution. If demand is normal with mean μ and standard deviation σ, then Q* = μ + zσ, where z is the standard normal value at the critical fractile. The mean anchors the decision. The cost asymmetry and the uncertainty determine how far to deviate from it.

A worked case clarifies the mechanics. A product costs $20, sells for $50, and salvages at $5. Underage cost is $30 and overage cost is $15, so the critical fractile is 0.67 and z is roughly 0.43. With a forecast mean of 1,000 units and a standard deviation of 300, the optimal order is about 1,130 units. Cut the salvage value to zero and the fractile drops to 0.60, pulling the order down toward 1,075. Small shifts in economics move the answer materially.

The deeper insight concerns σ. Under normal demand, expected profit at the optimal order equals the riskless margin on mean demand minus a mismatch cost of (Cu + Co) multiplied by φ(z) multiplied by σ. Uncertainty is therefore not merely an inconvenience to forecast around. It is a direct, quantifiable cost that scales linearly with the spread of demand. This reframes investments in postponement, demand pooling, and faster replenishment. Each is valuable precisely to the extent that it shrinks the effective standard deviation the newsvendor must hedge against.

Takeaway

The right order quantity is a quantile, not a forecast. The forecast tells you where demand is centered, and the ratio of underage to overage cost tells you how much of its uncertainty you should be willing to cover.

The critical fractile has a direct operational meaning. It is the optimal in-stock probability, the likelihood that demand during the season will not exceed available inventory. Service level is therefore not a policy to be set by aspiration or industry benchmark. It is an output of product economics. A firm that mandates a uniform 95 percent target across its portfolio is, by construction, overstocking some items and understocking others.

Consider what this means across different margin structures. A premium fragrance with a 70 percent gross margin and reasonable salvage options may warrant an in-stock probability above 90 percent. A thin-margin item with steep markdown losses might justify only 40 percent. When the fractile falls below one half, the optimal order sits below mean demand. The firm should then expect, and accept, stockouts in most seasons, because carrying more inventory would destroy more value than it protects.

This is the analytical foundation of Marshall Fisher's distinction between functional and innovative products. Functional goods, with stable demand and modest margins, call for efficient, lean supply chains. Innovative goods, with volatile demand and high margins, carry large underage costs and high critical fractiles, which justify buffer inventory and responsive capacity. The newsvendor makes Fisher's prescription precise. The appropriate supply chain design follows from where a product's fractile lands and how wide its demand distribution is.

Service level must also be distinguished from fill rate, the fraction of demand actually satisfied. The two diverge sharply. Expected lost sales equal σ multiplied by the standard normal loss function evaluated at z, a quantity that shrinks quickly as z rises. In the earlier example, a 67 percent in-stock probability still yields a fill rate near 93 percent, because most stockouts clip only the tail of demand. Executives who conflate the two metrics routinely misjudge their actual customer exposure.

Underage cost is also where measurement discipline matters most. Lost margin is visible. Lost customer lifetime value, channel substitution, and contractual penalties are not. Firms that omit these terms systematically understate Cu, depress their critical fractiles, and drift toward chronic under-service. In the same way, ignoring markdown cascades and disposal costs understates Co. The optimization is only as good as the cost accounting that feeds it. That makes the cost model, rather than the forecast, the most consequential input.

Takeaway

Service level is a consequence you derive, not a target you choose. A uniform in-stock goal across products with different economics guarantees that some will be overstocked and others starved.

The single-period model generalizes more readily than its assumptions suggest. In a multi-period setting where unmet demand is backordered, leftover inventory is carried forward at holding cost h rather than salvaged, and shortfalls incur backorder cost b. The optimal policy becomes an order-up-to level S satisfying F(S) = b / (b + h). Here F describes demand over the protection interval, meaning lead time plus review period. The newsvendor fractile reappears intact, now governing the base-stock policy.

Arthur Veinott's work on myopic policies explains why. Under fairly general conditions, including stationary costs, backlogging, and the freedom to order each period, solving an independent newsvendor problem in each period is optimal for the full horizon. Lost sales, fixed ordering costs, and nonstationary demand complicate matters, introducing (s, S) policies and dynamic programming. Even there, newsvendor bounds provide tight approximations and serve as starting points for the heuristics used in commercial planning engines.

In multi-product contexts, the binding constraint is usually shared capacity, budget, or shelf space. Adding a Lagrange multiplier λ on that constraint modifies each product's fractile. The effective underage cost falls by λ multiplied by the resource the item consumes. Products with high margin per unit of constrained resource keep high service levels, and the rest are rationed. When demand for different products can be pooled, for example through a common component or a regional warehouse, aggregate σ narrows and profit rises. This is the economic basis for postponement and centralization.

The framework also extends across organizational boundaries. When a retailer faces the newsvendor problem at a wholesale price above the manufacturer's cost, its underage cost shrinks, and it orders less than an integrated chain would. This is double marginalization. Buyback, revenue-sharing, and quantity-flexibility contracts restore the system-optimal fractile by reshaping each party's Co and Cu. Gérard Cachon and Martin Lariviere's analysis of revenue sharing in video rental is the canonical demonstration that contract design is newsvendor design.

The newest extension is data-driven. Rather than forecasting a mean and then applying a fractile, modern approaches train models to predict the critical quantile directly, using quantile loss functions weighted by Cu and Co. Gah-Yi Ban and Cynthia Rudin's work on the big-data newsvendor shows this can outperform the estimate-then-optimize pipeline when demand depends on rich features. The decision logic has not changed since Edgeworth first framed it in 1888. What has changed is the machinery for estimating the distribution it requires.

Takeaway

Many sophisticated inventory, capacity, and contract structures are newsvendor problems in disguise. When a new system seems intractable, ask what the overage and underage costs are, and over what interval the uncertainty resolves.

The newsvendor problem endures because it isolates one irreducible feature of supply chain decision-making: commitment under uncertainty with unequal consequences for error. Its answer, a quantile set by the ratio of underage to overage cost, turns what looks like a forecasting problem into an economic one.

For next-generation network design, the implications are structural. Every investment in visibility, responsiveness, or flexible capacity can be evaluated by how it changes three quantities: Co, Cu, and the effective σ over the commitment horizon. Traceability that lowers disposal risk, autonomous replenishment that shortens lead times, and analytics that sharpen distributional estimates all pull on these same levers.

The architecture changes, but the logic does not. Supply chains that build the critical fractile into their planning systems, contracts, and service policies will place inventory where its marginal value is highest. Those that rely on uniform targets and point forecasts will keep paying a mismatch cost that never appears as a line item.