Probabilità, Non Scenari

Joannes Vermorel

La supply chain viene spesso descritta come l’arte di bilanciare domanda e offerta. Io la vedo in termini più netti: ogni ordine di acquisto, ogni lotto di produzione, ogni variazione di prezzo è una scommessa su un futuro che non controlliamo. Impegniamo inventario, capacità, liquidità e attenzione oggi, sperando che il mondo di domani ricompensi tali impegni invece di punirli.

Because these decisions are bets, the way we think about the future is not a side issue. It is the core of the discipline. In my book Introduction to Supply Chain, I argue that the point is not to preserve harmony around a plan, but to commit resources where the expected, risk‑adjusted return is highest under uncertainty. Everything else is decoration.

Yet most of the supply chain world still works with a planning logic that treats the future as a handful of carefully scripted “scenarios.” My own view, formed over nearly two decades at Lokad, is that this scenario-centric mindset is not just suboptimal; it is structurally at odds with what a supply chain actually is. Supply chains live in probability spaces, not in storyboards.

In questo saggio voglio spiegare perché ritengo che la previsione probabilistica sia il linguaggio naturale della supply chain, e perché il mainstream incentrato sugli scenari – soprattutto come si concretizza in molti processi IBP e S&OP – non possa essere corretto con semplici ritocchi per diventare adeguato. Deve essere sostituito.

Come il mainstream inquadra il futuro

If you look at how large companies plan today, the pattern is remarkably consistent across industries and software vendors.

There is usually a baseline forecast, often a single time series per product family or region, generated by a demand-planning module. Around this baseline, Integrated Business Planning (IBP) processes orchestrate a monthly or quarterly ritual where sales, operations, and finance negotiate a “consensus” or “one” plan. Software suites such as SAP Integrated Business Planning explicitly present scenario planning and what‑if simulations as core capabilities: planners are encouraged to run alternate demand or offerta scenarios, compare them on dashboards, and select the one they find most appropriate.

Conceptually, the future lives as a small set of named worlds: base case, optimistic case, pessimistic case, perhaps a disruption scenario or two for good measure. Forecasts inside those worlds are deterministic; uncertainty is handled implicitly through service-level targets, safety-stock formulas, and some judgment. Once the favored scenario has been chosen, the organization is expected to “align” to it and to measure adherence with familiar KPIs.

The process feels structured and collaborative. It produces plans that can be explained in PowerPoint. It satisfies the understandable human desire to tell ourselves a story about tomorrow.

But it is a poor way to think about a system that makes, moves, and prices millions of units under relentless variability.

L’incertezza irreducibile delle supply chain

In manufacturing, it is often possible to buy your way to stability. Invest in better machines, tighter tolerances, more sensors, and defect rates go down. The randomness shrinks.

Supply chains do not offer this comfort. The uncertainty that matters lives in human behavior and in politics: demand shifts, competitor moves, promotions, price changes, upstream shortages, strikes, regulatory surprises. You can improve your data and your models, and you absolutely should, but you will never purchase a deterministic future.

If you accept this, it leads naturally to a different question. Instead of asking, “What is our plan for next quarter?” you start asking, “Given everything we know today, how does the future likely distribute itself, and how should we bet?”

È qui che entra in gioco la previsione probabilistica.

Cos’è veramente la previsione probabilistica

La previsione probabilistica viene spesso fraintesa come un modo sofisticato di tracciare intervalli di confidenza attorno a una previsione tradizionale. Non è questo il punto.

Per una supply chain, una previsione probabilistica assegna probabilità a tutti gli esiti che contano: quante unità potremmo vendere la prossima settimana, quanto tempo potrebbe effettivamente impiegare un fornitore a consegnare, quanto è probabile che un prodotto venga reso, con quale frequenza una macchina chiave potrebbe guastarsi. Invece di avere un solo numero previsto per la domanda e un “tempo di consegna” medio, otteniamo distribuzioni di probabilità complete per entrambi.

Questa prospettiva non è nuova in statistica, ma rappresenta un vero cambiamento di paradigma per la pratica della supply chain. In lezioni e interviste ho descritto la previsione probabilistica come uno dei cambiamenti più importanti in oltre un secolo di scienza delle previsioni, non perché la matematica sia esotica, ma perché ci permette finalmente di collegare le previsioni al processo decisionale in modo sensato.

Once we have distributions instead of single numbers, we can do something that scenario planning cannot: we can evaluate a decision across all plausible futures at once.

Decisioni come scommesse economiche sulle distribuzioni

Consider a very simple example. You are deciding how much of a product to order for next month.

If you rely on a single forecast, you get a number – say, 1,000 units – and you add some safety stock “just in case.” You might also review a couple of scenarios: what if demand is 20% higher, what if a supplier is late. Each scenario feels like a different world. You adjust the order quantity up or down, negotiate, and eventually pick a number.

Se ti affidi a previsioni probabilistiche, la logica è diversa. Hai una distribuzione per la domanda; forse esiste una modesta probabilità che le vendite siano estremamente alte, e una probabilità non trascurabile che siano molto basse. Hai una distribuzione per il tempo di consegna; alcune consegne sono puntuali, altre no. Hai quantità economiche associate agli esiti: il margine che guadagni se vendi, la penalità (esplicita o implicita) se hai una rottura di stock, il costo del capitale immobilizzato nell’inventario, il rischio di obsolescenza.

Now you can compute the expected economic outcome of ordering 800 units, 900 units, 1,000 units, and so on. Each order quantity becomes a bet whose payoff is evaluated across the whole probability space, not just a few handpicked scenarios. You can favor decisions that perform well on average and are robust to the tails, rather than ones that look good in a PowerPoint scenario but crumble when reality deviates slightly from the script.

This is not limited to inventory. The same logic applies to network design, capacity reservation, assortment curation, and even pricing. Everywhere we allocate scarce resources under uncertainty, we can ask the same question: given the probabilistic view of the future and our economic assumptions, which decision has the highest expected, risk‑adjusted return?

Scenario planning, as usually practiced, has no such calculus. At best, it offers a few isolated snapshots and leaves management to eyeball the trade‑offs.

Perché pochi scenari non sono sufficienti

Il primo problema della pianificazione per scenari è la granularità. Le moderne supply chain operano a una scala che rende l’approccio per scenari quasi comicamente grossolano. Potremmo avere milioni di combinazioni SKU–località–tempo su cui decidere, ciascuna con il proprio profilo di domanda, il proprio profilo di tempo di consegna e la propria sensibilità a prezzo e promozione. A fronte di questa realtà a grana fine, impostiamo forse tre o quattro narrazioni sul futuro.

Even if those narratives were perfect, they would still be too few. But they are not perfect. They are the result of judgment, politics, and habit. Which scenarios get written down is itself a random process, influenced more by organizational anxieties than by statistical evidence.

The second problem is that scenarios are rarely assigned explicit probabilities. We have “base,” “upside,” “downside,” but we do not say whether the upside case is five percent likely or fifty. IBP literature speaks warmly about continuous scenario planning, but in practice it means more and faster simulations, not calibrated probability distributions.

The third problem is that scenario planning tends to operate at a high level of aggregation. We run scenarios on total revenue, total capacity, perhaps a few key customers or product families. Meanwhile, the actual economic damage in a supply chain is done by local mismatches: the one component that is missing and grounds the aircraft, the one fashion item that is grossly overbought, the one region that is persistently under‑served. Those failures almost never show up clearly in aggregate scenario charts.

Scenarios appeal to us because they are narrative, nameable, and discussable. They fit our cognitive limitations. But supply chains are not narrative objects; they are stochastic systems with many degrees of freedom. A handful of stories cannot do justice to that reality.

Il costo nascosto della ricerca della precisione e degli scenari

For decades, companies have poured effort into improving forecast accuracy as if it were obviously aligned with better economics. We measure MAPE and similar metrics, launch Forecast Value Add initiatives, and celebrate small improvements as victories. Yet on the ground, the correlation between “better” accuracy and better P&L is often weak, sometimes negative.

In other work I have argued that this fixation on accuracy is a large, slow distraction. You can improve accuracy by predicting more zeros on intermittent demand and, in the process, starve your supply chain of the inventory it needs. You can improve accuracy by aggressively following the latest sales signal and inadvertently amplify bullwhip effects. You can generate impressive scenario decks that drive poor execution.

The deeper problem is that “accuracy” and “scenario coverage” are planning metrics that live in a world of numbers detached from prices. They evaluate how close forecasts are to realized quantities or how neatly scenarios cover a manager’s concerns, but they say nothing about the economic consequences of decisions. A small forecast error on a critical spare part can be much more damaging than a large forecast error on a slow‑moving accessory, yet accuracy metrics weight them similarly.

La previsione probabilistica ci consente di riportare l’intera discussione al centro dell’economia. Invece di chiederci quanto siamo vicini al numero realizzato, ci chiediamo se le nostre decisioni, valutate sull’intera distribuzione degli esiti, generino buoni rendimenti corretti per il rischio. La pianificazione per scenari, anche quando è adornata con strumenti sofisticati, ci incoraggia a ottimizzare le storie anziché il denaro.

Come il pensiero probabilistico cambia la pratica

Embracing probabilities instead of scenarios has several practical consequences.

In primo luogo, ci costringe a essere espliciti su dove risieda realmente l’incertezza. Domanda e tempo di consegna non sono “parametri” da fissare in un sistema di pianificazione; sono variabili casuali che devono essere modellate e continuamente aggiornate man mano che arrivano nuovi dati. Ciò vale tanto per l’affidabilità a monte e i resi quanto per la domanda dei clienti.

In secondo luogo, richiede di attribuire un valore monetario agli esiti. Una previsione probabilistica senza una lente economica è solo marginalmente migliore di una previsione deterministica. Dobbiamo sapere quanto costa una rottura di stock, quanto costa l’eccesso di inventario, quanto vale una vendita persa, come valutare la capacità inutilizzata. Queste quantità sono imperfette e talvolta scomode da stimare, ma sono l’unico modo per confrontare sensatamente le decisioni.

Third, it naturally leads to automation. Once you can compute the expected, risk‑adjusted payoff of a replenishment order, a price change, or a transfer, there is no reason to debate every decision in meetings. You can let software issue thousands or millions of small decisions every day, while humans focus on shaping the economic model, validating the assumptions, and handling the situations where the model’s confidence is low.

This is very different from an IBP process whose center of gravity is a monthly scenario workshop. Instead of a planning ritual that periodically re‑authorizes a plan, we get an economic engine that continuously arbitrates trade‑offs under uncertainty.

Dove gli scenari hanno ancora il loro posto

I am not arguing that companies should ban the word “scenario.” Imagination is essential in any complex endeavor. Boards and executives need narratives to reason about long‑term investments, strategic risks, regulatory shifts, and technology changes.

However, in a probabilistic supply chain, scenarios play a different role. They are not handcrafted futures that the planning system must follow. They are illustrations drawn from, or constrained by, the underlying probabilistic model.

If we want to explore a severe but plausible downside, we do not invent it from scratch; we stress the distributions in ways that are consistent with history and expert knowledge, and we let the same decision engine compute the consequences. If we want to show the upside potential of a more aggressive pricing strategy, we use the probabilistic model to simulate how demand might respond, and we quantify the range of possible outcomes.

In this way, scenarios become pedagogical views on a probabilistic reality, not substitutes for it.

Andare oltre le supply chain incentrate sugli scenari

The mainstream playbook has had a long run: deterministic forecasts, safety stocks, monthly consensus plans, and a handful of scenarios on top. It has brought some degree of structure, but it has also locked many organizations into a way of thinking that is increasingly incompatible with the complexity and volatility of modern supply chains.

L’alternativa non ha nulla di mistico. Si tratta semplicemente di prendere sul serio l’incertezza e di esprimerla nell’unico linguaggio che scala: la probabilità. Una volta che trattiamo domanda, tempi di consegna e altri fattori chiave come variabili casuali; una volta che attribuiamo prezzi agli esiti; una volta che valutiamo le decisioni come scommesse su distribuzioni complete anziché come risposte a poche storie; la pianificazione per scenari, nel suo senso tradizionale, inizia ad apparire per ciò che è: un meccanismo di compensazione per l’assenza di un vero motore probabilistico e incentrato sulle decisioni.

We do not need more scenarios. We need better probabilities, and a supply chain that knows what to do with them.