Ask an operations leader what they want from a plan and they will not say "the mathematically optimal schedule given last month's averages." They will say something like: hit the service level, protect the margin, and do not let one bad day cascade through the week. That is a statement about distributions. It deserves an optimizer that speaks the same language.
Stochastic optimization asks a different question than its deterministic cousin. Not "what is the best plan if the forecast is right," but "what is the best plan given everything that could plausibly happen, weighted by how likely it is and how much it hurts." The output is a decision that has already been stress-tested by construction.
What changes in practice
Three things change when a network moves to stochastic planning. First, recourse becomes part of the plan. The optimizer knows a fallback exists, what it costs, and when it triggers, so it stops overpaying for protection in the base case. Second, tail risk becomes a dial rather than a surprise. You can explicitly choose how much expected cost to spend reducing the worst 5 percent of outcomes, which is a board-level conversation, not a dispatcher's guess. Third, buffers become targeted. Slack goes where the variance is, not everywhere.
Priced for the mean, or priced for the distribution.
Raise volatility and watch which policy holds.
Policy B buys more capacity and pays for it, then stops paying penalties.
The objection, answered
The standard objection is data: "we do not know our distributions." In our experience the objection has it backwards. Every network already contains its distributions, recorded in years of shipment events, delays, and costs. What networks lack is a planning layer that consumes that history as distributions instead of collapsing it to averages. Estimating a transit time spread from 100,000 historical legs is a solved problem. Refusing to use the spread is a choice.
The mathematics has existed since the 1950s. What changed is compute, data availability, and solvers that make scenario-based optimization tractable at operational scale and speed. The gap between networks that plan over distributions and networks that plan over averages is now a compounding commercial gap, visible in margin, service, and the size of the firefighting team.