The State of “Probabilistic” Forecasting in Supply Chain (2025)

Joannes Vermorel

In 2012, at Lokad, we pivoted from single‑number time‑series to a probabilistic view of the world—first through quantile forecasting, then full predictive distributions for demand and lead time, and ultimately stochastic decision optimization. More than a decade later, very little has actually changed in the enterprise software market—except at Lokad. Most vendors now utter the word probabilistic; almost none have rebuilt their stack so that uncertainty is modeled, combined, and carried all the way into automated decisions.

A compact map of the market

This table codifies public claims and documentation as of November 2025; where vendors publish little technical detail, I err on the conservative side.

A sharper set of litmus tests

Over the years I have refined a few tests that separate mathematics from marketing.

First, uncertainty must be modeled where it hurts. A supply chain does not only face demand volatility; supply timing is itself a random variable. If a vendor does not estimate full distributions for demand and for lead time—then combine them coherently—probabilistic claims are beside the point. Merely inflating point forecasts with buffers is not probabilistic forecasting.

Second, the output must be decisions, not tuning knobs. A genuine system pushes distributions into an optimization layer that trades expected returns against costs and constraints, yielding quantities to order, allocations, and prices. If the “end‑game” is a safety‑stock percentile or a target service level, it is not a probabilistic decision system; it is a deterministic policy tuned by variance estimates.

Third, probability must survive contact with the process. If planners are expected to “edit” forecasts in a grid, we have already left the probabilistic realm. Human overrides can amend constraints or priorities; they do not reshape well‑calibrated distributions by hand.

Fourth, fat tails matter. Retail and spare parts are intermittent; lead times are heavy‑tailed. A vendor that quietly assumes Gaussians everywhere is not doing probabilistic forecasting; it is doing algebra with wishful moments. The practice insists on non‑Gaussian treatment.

Fifth, measurement must be probabilistic. If the headline KPIs are MAPE and classic accuracy, the incentives will favor point predictions and service‑level cosmetics. Proper scoring rules for distributions and, better yet, profit‑and‑loss‑grounded objectives, are what count.

Finally, transparency is non‑negotiable. Vendors should publish enough methodology for practitioners to see how probabilities become decisions, including how multiple uncertainties are composed and propagated.

Against that yardstick, here is where the market stands.

Vendor‑by‑vendor assessment

Lokad — the exception that proves the rule

Lokad’s stack is probabilistic by construction: demand and lead time are learned as full predictive distributions, combined inside programmatic decision models that output concrete actions—orders, allocations, pricing—under business constraints. We made a deliberate break with safety‑stock/service‑level constructs, favoring decision‑first optimization and fat‑tail handling; the objective is economic, not accuracy‑for‑its‑own‑sake. This journey started in 2012 (quantile forecasting), matured into full distributional models and stochastic optimization, and remains our dominant paradigm.

ToolsGroup — probabilistic demand, service‑level front‑end

ToolsGroup communicates clearly about full demand distributions and long‑tail handling. Yet those distributions feed a service‑driven MEIO: stock‑to‑service curves, dynamic safety stocks, and target service levels remain the lingua franca. In practice, this is a probabilistic modeling layer wired back into deterministic policy parameters and planner workflows.

Smart Software — credible distributions, classical policies

Smart’s long‑standing strength is demand‑over‑lead‑time distributions for intermittent items, often embedded via Epicor IP&O. The math is real; the operationalization is classical: choose service levels, set safety stocks, simulate policies. That is probabilistic input with deterministic output.

Epicor IP&O (Smart) — Smart inside, service outside

Epicor’s IP&O advertises probabilistic modeling of demand and lead time and stresses policy stress‑testing. Yet the levers exposed to users are service levels, safety stocks, and reorder logic; optimization is framed as service‑versus‑cost.

SAP IBP (SmartOps) — stochastic MEIO as a module

IBP for Inventory descends from SmartOps: multi‑stage optimization with forecast error and lead‑time variability to maintain customer service levels at minimum cost. It calculates safety stocks and service‑level targets; probabilistic elements exist, but as an add‑on to an otherwise planner‑driven process.

GAINS — optimization to service targets

GAINS markets MEIO that “considers uncertainty” in demand, supply, and lead time. The working interface, however, is explicit service‑level optimization and safety‑stock settings. Sophisticated, yes; probabilistic decisioning end‑to‑end, no.

Blue Yonder — probabilistic in name, service in practice

Blue Yonder’s pages speak of “autonomous” and “probabilistic” forecasts, yet the heart of inventory planning is granular service‑level segmentation and dynamic safety stock. Case studies and partner material reinforce a scenario‑ and service‑driven operating model rather than distribution‑native decisioning.

RELEX — retail focus, safety‑stock core

RELEX focuses on retail execution and now touts probabilistic modeling around True Inventory. But when it comes to replenishment and protection against uncertainty, its own materials still center on mastering safety stocks and achieving target service levels—a deterministic policy world calibrated by ML.

o9 Solutions — scenarios with service‑level math

o9’s “Digital Brain” is a capable scenario platform. Its MEIO pages describe optimal service‑level recommendations and continuous policy rebalancing; probability appears as event likelihoods and what‑ifs, not as distributions that directly drive optimization of expected economic outcomes.

Kinaxis — quantiles and a MEIO extension, still service‑first

Kinaxis has moved toward probabilistic elements: blogs discuss quantile forecasts, and Wahupa’s extension brings a probabilistic MEIO into Maestro. Yet even that extension advertises differentiated service levels and safety‑stock settings; the broader process remains scenario‑heavy and planner‑centric.

E2open — classic MEIO vocabulary

E2open explains MEIO in terms of optimizing inventory across nodes to meet service expectations, with demand, lead time, and service rippling through the network. It is the canonical service‑level MEIO story.

Coupa/LLamasoft — service‑level design

Coupa’s design suite (ex‑LLamasoft) is explicit: develop policies per SKU, calculate optimal safety stocks, and optimize to service‑level objectives. That is stochastic parameterization of deterministic policies.

Infor — probabilistic phrasing, deterministic levers

Infor materials mention “intelligent, probabilistic forecasting,” but the surrounding content emphasizes forecast accuracy, service levels, and safety‑stock replenishment. The levers made available to planners are not distribution‑native decisions.

Anaplan — safety‑stock apps with scenarios

Partner content and demos showcase inventory apps built around dynamic safety stock, service‑level balancing, and fast scenario modeling. Useful, yes; probabilistic forecasting in the strong sense, no.

What “genuine” really means

A genuine probabilistic supply‑chain system must clear a high bar. A minima, it estimates full distributions for demand and lead time; it composes these uncertainties—often with heavy tails—into a single stochastic view of each decision; it optimizes an economic objective under constraints, returning decisions rather than parameter targets; it measures itself with proper probabilistic scores or direct financials; and it automates execution so that humans manage priorities and constraints, not the shapes of distributions. At Lokad, we built precisely that. It is not a veneer on top of safety stock; it is a different architecture.

Closing thought

I welcome the industry’s newly found taste for the word probabilistic. But words are not the thing. Ten years on, most vendors still dress deterministic policies with stochastic cosmetics. Until uncertainty is combined and optimized into decisions—without safety‑stock and service‑level crutches—the claim remains marketing. Lokad stands apart because we removed those crutches a long time ago and learned to walk on probability itself.

Methodological note. This assessment relies on vendors’ public documentation and independent write‑ups. ToolsGroup openly discusses distributions yet routes them to service‑level MEIO; Smart/Epicor emphasize demand‑over‑lead‑time distributions feeding policy choices; SAP IBP/SmartOps documents service‑level‑driven multi‑stage optimization; GAINS surfaces service‑level optimization screens; Blue Yonder advertises “probabilistic” while centering on service segmentation and dynamic safety stocks; RELEX highlights safety‑stock mastery and probabilistic inventory accuracy; o9 spotlights service‑level‑optimized MEIO and scenarios; Kinaxis adds a Wahupa MEIO extension that still speaks the service‑level dialect; E2open and Coupa/LLamasoft describe classic MEIO; Infor and Anaplan frame outcomes in accuracy and service‑stock terms.

If a vendor wishes to dispute their placement, they are invited to publish the fine print: how they compose demand and lead‑time distributions; how those distributions are optimized into decisions; how they measure calibration; how they handle heavy tails; and how the process resists the temptation of human edits on probability shapes. That would be news.