From Distribution to Decision: How Probabilities Become Order Quantities

“Cool, But What Do I Actually Order?”

You’ve just built a probabilistic demand forecast. Instead of a single number, you now have a full distribution — percentiles, confidence intervals, maybe even a Monte Carlo sample of 10,000 possible demand scenarios for the next four weeks. It’s sophisticated. It’s statistically rigorous. It’s beautiful.

And your purchasing manager is staring at you asking: “Great, but how many units do I put on the PO?”

This is the question that trips up nearly everyone who encounters probabilistic forecasting for the first time. If the whole point is that demand is uncertain and a single number is misleading, how does a probability distribution help? You can’t ship a confidence interval to a warehouse. Your supplier doesn’t accept quantile functions as order quantities. At some point, someone has to commit to a specific number.

Here’s the punchline: you absolutely do land on a single number. Every time. The difference is how you arrive at that number — and that difference is worth millions.

The Old Way: Forecast, Then Buffer

The traditional approach to turning a forecast into an order quantity is straightforward:

1. Get a point forecast (e.g., “we expect to sell 10,000 units”)
2. Add safety stock using a rule of thumb (e.g., “add 20%” or “keep 2 weeks of supply”)
3. Place the order

The safety stock formula might look scientific — it might involve a service level target, a standard deviation, and a z-score. But in practice, it’s built on shaky assumptions: that forecast errors are normally distributed (they rarely are), that the standard deviation is stable over time (it isn’t), and that a single average error metric captures the risk profile (it doesn’t).

The buffer is a hedge against the forecast being wrong, but the hedge itself is poorly calibrated. You’re protecting against uncertainty with a tool that doesn’t actually understand the uncertainty.

The New Way: Optimize Directly Against the Distribution

With a probabilistic forecast, the logic is fundamentally different. You don’t forecast and then buffer. You feed the entire distribution into an economic optimization that directly computes the best quantity to order.

Here’s how it works conceptually:

For every possible order quantity Q, you can compute the expected economic outcome by evaluating Q against the full distribution of demand scenarios. If you order Q and demand turns out to be higher, you eat the cost of lost sales, expedited shipping, or damaged customer relationships. If you order Q and demand turns out to be lower, you eat the cost of excess inventory, markdowns, storage, or obsolescence.

The optimal Q is the one where these two forces balance out — weighted by their actual probabilities and actual costs.

That’s it. That’s the entire intellectual framework. The distribution gives you the probabilities. Your business context gives you the costs. The math gives you the number.

Why This Is Different from “Pick the Median”

A common misconception is that a probabilistic forecast just gives you more percentiles to choose from, and the “decision” is picking which percentile to use. Order at the 50th percentile if you’re balanced, the 90th if you’re conservative, the 75th if you’re somewhere in between.

This is better than a point forecast with an arbitrary buffer, but it’s still not right. Here’s why:

The right order quantity depends on the economics, not on a percentile. If stockout costs are 10x holding costs, the optimal order quantity might correspond to roughly the 91st percentile. If holding costs and stockout costs are roughly equal, the optimum might land near the 50th percentile. If the product is perishable and excess inventory goes to waste, you might order below the median.

But these aren’t arbitrary percentile choices — they’re the output of an optimization that considers the full shape of the distribution and the full asymmetry of the cost structure. The optimal quantity might not correspond to any clean percentile at all. It emerges from the math.

The distinction matters because it changes what you optimize. Picking a percentile is a heuristic. Optimizing against the distribution is a principled economic calculation.

The Newsvendor Intuition

The classic illustration of this idea is the newsvendor problem — one of the oldest and most elegant results in operations research.

Imagine you’re selling newspapers. You buy them in the morning at cost $c$ and sell them during the day at price $p$. Unsold papers are worthless at the end of the day. How many should you buy?

If demand follows some probability distribution $F$, the optimal order quantity $Q^*$ satisfies:

$$F(Q^*) = \frac{p - c}{p}$$

That’s it. The ratio on the right is called the critical ratio — it captures the cost asymmetry between ordering too much (you lose the purchase cost $c$) and ordering too little (you lose the profit margin $p - c$). The higher the margin relative to cost, the more you should order. The distribution $F$ tells you exactly where to set the target.

This isn’t the 50th percentile, or the 90th, or any predetermined level. It’s the percentile that minimizes your expected total cost given your specific economics. If your margin is 80%, the critical ratio is 0.80, and you order at roughly the 80th percentile of demand. If your margin is 40%, you order near the 40th percentile.

The newsvendor is a simplification — real supply chains have lead times, multiple periods, substitution effects, and capacity constraints. But the core insight generalizes perfectly: the optimal decision is determined by the interaction of the probability distribution and the cost structure, not by the distribution alone.

Real Supply Chains Are More Complex (But the Principle Holds)

In practice, the decision isn’t just “how many newspapers to buy.” It’s multi-SKU inventory allocation across a network, with lead time uncertainty, supplier minimums, container capacity constraints, shelf life limits, and demand that’s correlated across products and regions.

The newsvendor formula won’t solve that directly. But the principle scales:

Step 1: Generate the demand distribution. Not just point estimates — the full probabilistic forecast for each SKU, location, and time period. Include correlations if they matter (e.g., if a promotion on product A cannibalizes product B).

Step 2: Define the cost function. What does it cost to have one unit too many? One unit too few? These aren’t always obvious. Stockout cost might include lost margin, lost customer lifetime value, contractual penalties, or emergency procurement premiums. Holding cost might include warehousing, capital cost, shrinkage, and obsolescence. Sometimes the cost function is asymmetric in weird, nonlinear ways — and that’s fine. The framework handles it.

Step 3: Optimize the decision against the distribution. For each feasible order quantity (or allocation, or schedule), compute the expected cost across the distribution of scenarios. The best decision is the one with the lowest expected cost — or, if you’re risk-averse, the one that minimizes some risk-adjusted metric like CVaR.

In complex settings, this might involve simulation-based optimization, stochastic programming, or scenario-based heuristics. The specific technique varies. The logic doesn’t.

What Changes in Practice

When you shift from “forecast + buffer” to “optimize against the distribution,” several things change in how the organization operates:

Safety stock becomes a computed output, not an input

You stop setting safety stock targets manually. Instead, the optimization naturally produces order quantities that embed the right amount of protection. If demand uncertainty is high and stockout costs are severe, the system orders more. If the product is cheap to hold and expensive to run out of, it orders more. This happens automatically, SKU by SKU, without anyone hand-tuning parameters.

Service levels become an outcome, not a target

Traditional systems work backwards from a service level target (“we want 95% fill rate”) and compute how much stock that requires. A distribution-based system works forward from the economics and tells you what service level you’ll achieve. Often, you discover that some SKUs should have 99.5% service levels (because stockouts are catastrophic) and others should have 85% (because excess is worse than shortage). A uniform service level target is almost always wrong — it over-invests in low-risk items and under-invests in high-risk ones.

You can quantify the value of information

With a probabilistic framework, you can measure exactly how much a better forecast is worth. If reducing forecast uncertainty by 10% at the tails changes your optimal order quantity and saves $50K in expected cost, you know the forecast improvement is worth up to $50K. This makes R&D investment in forecast models a quantifiable decision, not a leap of faith.

Buyers and planners become decision reviewers, not decision makers

The system produces the order quantity. The human reviews it, applies judgment for factors the model doesn’t capture (an upcoming trade show, a supplier going through a merger, a hunch about a trend), and approves or adjusts. The human isn’t doing arithmetic — they’re applying contextual intelligence on top of a quantitatively sound baseline.

The Common Objection (And Why It’s Wrong)

The most frequent pushback is: “This is too complicated. Our team can barely manage a point forecast. You want us to generate distributions AND run economic optimizations?”

The answer is: this is exactly what software is for. The buyer doesn’t see a probability distribution. They don’t see a cost function. They see a recommended order quantity — a single number on their screen — with maybe a confidence range and a few what-if scenarios they can toggle. The complexity is in the engine, not the interface.

The irony is that the “simple” approach — point forecast plus manual safety stock — actually requires more human judgment and intervention, not less. Someone has to decide the safety stock multiplier. Someone has to review and override when the number “feels wrong.” Someone has to firefight when the buffer wasn’t enough. All of that mental labor gets replaced by a system that does the math correctly in the first place.

Simple for the user. Rigorous under the hood. That’s the goal.

The Bottom Line

Probabilistic forecasts don’t make decisions fuzzier. They make decisions sharper. The output is still a crisp, actionable number — order 4,237 units, schedule 12 workers, reserve 3 containers. The difference is that the number was computed by weighing every plausible future against its economic consequences, instead of plugging a guess into a formula and hoping the buffer holds.

You start with uncertainty. You end with a decision. The distribution is the bridge between the two — and crossing it correctly is the difference between supply chains that react and supply chains that perform.