Probabilistic Forecasting in Supply Chains: Lokad vs. Other Enterprise Software Vendors, July 2025

By Léon Levinas-Ménard

Last modified: July 23rd, 2025

Executive Summary

Lokad’s Unique Probabilistic Approach: Lokad pioneered the use of probabilistic forecasting in supply chain optimization, moving beyond point estimates to model full distributions of demand and supply uncertainties. Since 2012, Lokad has built its platform around this concept – estimating entire probability distributions (not just single forecasts or a few quantiles) and using them to optimize decisions ¹ ². This enables forecasting all sources of uncertainty (e.g. demand and lead time variability) and computing optimized inventory or production plans that account for those uncertainties ². The result is a decision-centric approach that turns forecasts into optimized actions under uncertainty, rather than treating forecasting and planning as separate steps.

Competitors’ “Probabilistic” Claims vs Reality: In the wake of Lokad’s lead, many major supply chain software vendors have adopted the language of probabilistic forecasting – but largely without the same substance. Vendors such as ToolsGroup, Blue Yonder, o9 Solutions, SAP IBP, RELEX, and Kinaxis now sprinkle terms like “probabilistic” or “stochastic” in their marketing. However, a closer look reveals key differences:

ToolsGroup (SO99+): Longstanding “stochastic” pioneer, but modern use of probabilistic is partial. They generate demand distributions (quantile forecasts) for inventory, yet they do not forecast lead times as random variables, relying on fixed lead-time inputs ³. Notably, since 2018 ToolsGroup touted “probabilistic forecasts” while still touting MAPE accuracy improvements – a contradiction, since MAPE does not apply to probabilistic forecasts ⁴. This suggests their probabilistic push was more buzzword than fundamental change.
Kinaxis (RapidResponse/Maestro): Historically focused on deterministic, in-memory planning. Only recently (2022–2023) has Kinaxis embraced probabilistic methods by partnering with Wahupa (for probabilistic inventory optimization) and acquiring an AI forecasting firm. Kinaxis even filed 2023 patents for ML-based quantile forecasting ⁵ ⁶, indicating a shift toward probabilistic techniques. However, these features are new and unproven, essentially providing quantile forecasts (prediction intervals) rather than the full distribution modeling Lokad does. The Wahupa initiative (probabilistic multi-echelon engine) thus far has had limited visible impact, highlighting the challenges of retrofitting probabilistic logic into an established platform.
Blue Yonder (formerly JDA): A giant in supply chain software with legacy (deterministic) planning engines. Blue Yonder’s recent Luminate platform messaging mentions “autonomous” and “probabilistic” forecasting, but evidence is scarce that its core engines truly output full probability distributions ⁷. In practice, BY appears to stick to traditional point forecasts (time-series methods like ARIMA) with safety stock formulas, perhaps adding ML-driven adjustments. There is no indication that Blue Yonder models lead time uncertainty or produces the kind of end-to-end probabilistic optimization Lokad does ⁸ ⁹. The “probabilistic” terminology thus seems mostly a branding of minor enhancements to a fundamentally deterministic approach.
SAP IBP (Integrated Business Planning): SAP’s planning suite inherited a probabilistic engine through its 2013 acquisition of SmartOps (which performed multi-echelon inventory optimization using demand variability models). In theory, IBP for Inventory can account for demand variability and even some supply variability ¹⁰. In practice, SAP’s focus is on integration and process; probabilistic forecasting isn’t emphasized in their messaging ¹¹. Most SAP IBP deployments still use point forecasts and user-set safety stocks; lead times are typically fixed inputs (with optional buffers) rather than system-forecasted uncertainties ¹¹. Thus, while the capability exists deep in the software, SAP has not operationalized probabilistic methods as a core differentiator, and many users may not leverage those advanced features.
o9 Solutions: A newer platform that markets a “Digital Brain” for supply chain planning. o9 focuses on real-time integrated planning (demand, supply, finance) with scenario analysis and big data integration. Its strength is unifying data and planning silos, but not specifically probabilistic forecasting of the kind Lokad specializes in ¹². o9’s AI/ML features mostly aid predictive analytics and what-if simulations, rather than producing explicit probability distributions for every variable. In short, o9 provides a broad planning toolkit (with graph-based data models and fast what-if recalculations), whereas Lokad provides deeper uncertainty modeling and optimization. Companies seeking full probabilistic rigor in forecasts may find o9’s approach more incremental (augmented point forecasts) compared to Lokad’s comprehensive stochastic optimization ¹².
RELEX Solutions: A fast-growing, retail-focused vendor known for high-frequency forecasting and automated replenishment (popular for grocery and fashion). RELEX touts “AI-driven” demand forecasting and real-time analytics, but it does not output full probability distributions for demand as Lokad does ¹³. Forecasts are improved via machine learning (and short-term demand sensing) but essentially remain enhanced point forecasts. Critically, RELEX does not treat lead times as probabilistic either ¹⁴ – users input lead times and perhaps manual variability factors. Thus, despite modern cloud architecture and AI claims, RELEX relies on conventional methods (with some ML) under the hood ¹⁵. Its use of “probabilistic” terminology is minimal; the emphasis is more on near-term responsiveness than on modeling every uncertainty mathematically.

Bottom Line: Lokad’s probabilistic approach remains highly differentiated. It stands out for genuinely incorporating uncertainty end-to-end – forecasting all relevant distributions and optimizing decisions accordingly. Other vendors have, to varying degrees, borrowed the buzzword or added pieces of probabilistic tech (often just quantile forecasts or risk buffers) to appear up-to-date. Yet, as detailed in this report, none match the depth of Lokad’s approach in practice. Most still fall back on deterministic planning paradigms with only superficial nods to uncertainty (e.g. using a few quantiles or “safety stock” logic rather than true stochastic optimization). Recent developments – such as Kinaxis’s patents and partnerships or Blue Yonder’s AI rebranding – show the industry’s recognition that probabilistic methods are the future. However, the substance lags behind the marketing for these incumbents. Executives evaluating “probabilistic” supply chain solutions should therefore scrutinize whether a vendor’s offering is merely using the word or truly embracing the probabilistic paradigm pioneered by Lokad.

Introduction

Over the last decade, probabilistic forecasting has emerged as a pivotal innovation in supply chain planning. Unlike traditional deterministic forecasts that provide a single expected value, a probabilistic forecast provides a range of possible outcomes with associated probabilities ¹⁶. This is crucial for supply chain optimization: decisions like how much inventory to stock or how to schedule production must account for variability in demand, supplier lead times, transportation delays, and other uncertainties. By 2012, Lokad recognized that classic forecasts (even “best guess” or median forecasts) were insufficient for such decisions, since they ignore the risk of higher or lower outcomes. Lokad began producing quantile forecasts in 2012 – essentially biased forecasts aimed at specific service levels or cost trade-offs – and by 2015 evolved toward forecasting entire distributions (via “quantile grids”) ¹⁷ ¹⁸. In 2016 Lokad fully embraced probabilistic forecasting, explicitly estimating full demand distributions rather than single points ¹⁹ ²⁰. This was followed by developing stochastic optimization techniques that take those distributions as inputs to compute optimal decisions under uncertainty ²⁰ ²¹.

Today, Lokad’s approach (sometimes branded “Forecasting & Optimization” or “Quantitative Supply Chain”) tightly integrates forecasting with decision optimization ²² ²³. Using its domain-specific language Envision, Lokad models the uncertainties (e.g. demand variability, supplier reliability, lead time distribution, etc.) and the business constraints (inventory costs, service targets, capacity limits), then produces an optimized plan (order quantities, production schedules, etc.) that maximizes the expected performance given the uncertainties ²⁴ ². The key is that uncertainty is not an afterthought – it’s baked into the computations. This contrasts with traditional tools that often generate a single forecast, then add buffers (safety stock, safety time) in a heuristic way. Lokad’s success in the M5 forecasting competition in 2020 (achieving #1 accuracy at SKU level worldwide) further demonstrated its leadership in predictive modeling ²⁵. More importantly, Lokad argues that accuracy alone isn’t enough – one needs to make optimal decisions from those probabilistic forecasts ²⁶ ²³.

As Lokad has gained traction with this probabilistic approach, other supply chain software vendors have taken note. In the mid-2010s, terms like “stochastic”, “probabilistic forecasting”, and “AI-driven planning” started appearing in competitors’ brochures. By late 2010s, a number of vendors began marketing some form of probabilistic capability – at least in name. The challenge for supply chain executives is cutting through this marketing: how do these vendors define “probabilistic” and how do their solutions differ from Lokad’s? This report examines Lokad’s approach versus the approaches of several leading vendors: ToolsGroup, Kinaxis, Blue Yonder (BY), SAP IBP, o9 Solutions, and RELEX Solutions. We focus on how each vendor addresses uncertainty in forecasting and planning:

Do they produce full probability distributions or just point forecasts / single-number predictions?
If distributions are mentioned, are they only for demand, or also for lead times and other factors?
Have they demonstrated their forecasting prowess (e.g. in competitions or published metrics), or is it mainly buzzwords?
How do they incorporate uncertainty into decision-making (e.g. true stochastic optimization vs. simple safety stock formulas)?
Are their “AI/ML” claims backed by technical specifics, or are they retrofitting old methods with new terminology?

By exploring these questions, we can understand the gap between Lokad and its competitors when it comes to probabilistic supply chain optimization. Below, we detail each vendor’s stance and capabilities, highlighting how “probabilistic” is interpreted in practice – and whether it lives up to the promise that Lokad exemplifies.

Lokad’s Probabilistic Approach: Full Distributions & Decision Optimization

Lokad’s philosophy is that better decisions come from better understanding of uncertainty. Concretely, this means forecasting the full probability distribution of future demand (and other uncertainties) and then directly computing the decisions that optimize metrics (like service level, cost, or profit) with those distributions as inputs. Several elements make Lokad’s approach unique:

Early Adoption and Innovation: Lokad was nearly a decade ahead of the market in pushing probabilistic forecasts. As early as 2012, Lokad publicly championed forecasting beyond averages – introducing quantile forecasts tailored to business goals ²⁷ ¹. By 2015–2016, Lokad moved to full probabilistic forecasting, meaning for each item it produces an entire probability distribution of demand over the lead time (or any horizon of interest) ¹⁹ ²⁰. This was a radical shift from the industry norm of generating one number per item-period. Lokad’s investment in this area made it “one of the few vendors to truly implement probabilistic forecasting (demand and supply) and true stochastic optimization” in practice ².
All Sources of Uncertainty Modeled: Unlike most tools that only model demand uncertainty (and treat supply lead times or other factors as fixed), Lokad explicitly models every significant stochastic factor. For example, if supplier lead times vary, Lokad will forecast a lead time distribution (e.g. a 10% chance a lead time extends to 8 weeks instead of the average 6 weeks). If there is manufacturing yield uncertainty or transportation delay risk, those too can be modeled as probabilistic inputs. Lokad’s documentation emphasizes forecasting both demand and supply uncertainties and feeding both into the optimization ². This comprehensive approach means the resultant decisions (like how much inventory to keep) protect against all major variability, not just demand swings. In contrast, a vendor who ignores lead time variability might understock because they never anticipated a late supplier (or overstock “just in case”, not knowing the actual risk distribution).
Probabilistic Inputs to Optimization (Not Just Forecasts): Crucially, Lokad doesn’t stop at forecasting distributions; it uses them in mathematical optimization models to derive decisions. Lokad’s Envision platform allows crafting an objective function (e.g. maximize expected profit or minimize total cost) that is evaluated under the full range of probabilistic outcomes ²⁶ ²³. Techniques like Stochastic Discrete-Event Simulation and Stochastic Optimization (Lokad introduced methods like Stochastic Discrete Descent in 2021 to solve these problems ²⁰) compute the best decision by weighing thousands of possible future scenarios (drawn from the forecast distribution). This yields recommendations like order X units now (or set reorder point to Y), with known probabilities of stockout or overstock based on the forecast. It’s a holistic forecast-to-decision pipeline: data → probabilistic forecast → optimized decision. Many vendors, by contrast, either provide forecasts and leave the rest to planners, or use simplistic rules (like “safety stock = Z * σ of demand”) that are not true optimizations.
Transparency and Tailoring: Lokad has made an effort to white-box its approach. Envision is a fully programmable engine where a company can adjust the model to its realities. For instance, if obsolescence is a concern, one can model a probability that demand drops to zero after a certain date; if there’s a chance of supplier failure, one can incorporate that scenario. This flexibility ensures the “probabilistic model” isn’t a black box – it’s understandable and adjustable, which contrasts with some vendor’s one-size-fits-all black-box AI. Moreover, Lokad’s results and methods have been documented in detail on their website and YouTube lectures (supply chain “Lokad TV”), reflecting a level of technical transparency not common in enterprise software ²⁸ ²⁶.
Proven Performance: Lokad’s credibility in probabilistic forecasting is backed by external benchmarks. One highlight often cited is Lokad’s performance in the M5 competition (a global forecasting contest) where it achieved top accuracy at SKU/item level ²⁵. This matters because it’s objective evidence that Lokad’s forecasting technology is state-of-the-art. Additionally, by being cloud-native and fully automated, Lokad ensures that these advanced techniques can run at scale without manual intervention – daily or weekly re-optimization can happen “hands-off,” which is essential for practicality. In short, Lokad pairs cutting-edge science with automation, aiming to eliminate the traditional trade-off between sophisticated models and ease-of-use.

In summary, Lokad’s probabilistic approach means true end-to-end stochastic planning: granular uncertainties are forecast, then directly converted into decisions that optimize outcomes under uncertainty. This is not just a module or feature, but the core of Lokad’s platform. The rest of this report will use Lokad’s standard as a yardstick to examine how other vendors have (or have not) incorporated probabilistic forecasting.

Competing Vendors and their “Probabilistic” Approaches

As probabilistic methods gained attention, other supply chain software vendors responded in various ways. Some have roots in stochastic techniques but may not have advanced them recently; others have bolted on new capabilities or simply rebranded aspects of their existing tools with probabilistic terminology. Below we explore each major vendor’s approach:

ToolsGroup – Early Stochastic Innovator, But Quantile-Focused and Demand-Only

Background: ToolsGroup is a veteran in supply chain planning (founded 1993) and is often credited with early use of stochastic models, especially for spare parts forecasting. In the 1990s, ToolsGroup introduced methods that challenged purely deterministic planning by statistically modeling demand variability (e.g. using Poisson or other distributions for intermittent demand) ²⁹. This heritage means ToolsGroup has long talked about probabilistic forecasting – indeed, their flagship Service Optimizer 99+ (SO99+) system has for years computed “statistical safety stocks” based on variability. In modern marketing, ToolsGroup continues to claim “probabilistic demand forecasting (also known as stochastic forecasting) as the cornerstone of effective planning” ³⁰.

Probabilistic in Practice: Despite the legacy, ToolsGroup’s current approach appears limited in scope compared to Lokad’s comprehensive probabilistic framework. Notably:

Demand Distributions, but Static Supply Parameters: ToolsGroup does generate demand distributions rather than single-point forecasts. For example, SO99+ can produce a “stock-to-service” curve, which essentially shows the demand probability distribution over the lead time and the service level attained for a given stock level ³¹. This implies ToolsGroup’s engine simulates or analytically derives the distribution of demand during a replenishment lead time – a useful probabilistic output. However, ToolsGroup stops short of full probabilistic modeling because it treats lead time as an input, not as a forecasted uncertainty. In ToolsGroup’s documentation, lead times are listed among “supply parameters” that the user provides to the model ³². There is no indication that SO99+ will analyze historical supplier performance and output a probability distribution of lead time. At best, the user might enter a mean and a standard deviation for lead time, and the tool will factor that into safety stock calculations deterministically ³³. This means ToolsGroup is ignoring a major source of uncertainty – if a supplier’s lead time occasionally doubles due to disruptions, ToolsGroup’s optimization might not fully account for that risk because it isn’t natively forecasting that scenario ³⁴. In contrast, Lokad or a “full” probabilistic approach would explicitly model a 10% (for example) chance of lead time doubling and adjust stock recommendations accordingly. The absence of lead-time probabilistic forecasting led one analysis to conclude: “ToolsGroup fails the full probabilistic test – it mentions lead times only as static inputs, not as forecasted uncertainties” ³².
2018 “Probabilistic” Push and MAPE Confusion: ToolsGroup began heavily advertising “probabilistic forecasts” around 2018, likely in reaction to market trends. However, this marketing push was undermined by an apparent lack of understanding (or transparency) in how to measure probabilistic forecast performance. Specifically, ToolsGroup’s materials from that time paired claims of probabilistic forecasting with claims of improved MAPE (Mean Absolute Percentage Error) ⁴. MAPE is a metric for point forecast accuracy – it measures the deviation of a single forecast number from actual demand. For a probabilistic forecast (which produces a distribution or multiple quantiles), MAPE is not applicable ³⁵. You cannot compute “the MAPE of a distribution forecast” without first reducing it to a point estimate (defeating the purpose of probabilistic output). Thus, ToolsGroup’s boast of better MAPE alongside probabilistic forecasting either indicates they were still effectively using a point forecast (perhaps the median) for error measurement, or it was a marketing oversight. Either way, it raised skepticism: “Such an obvious mix-up suggests that ToolsGroup’s probabilistic initiative might be more buzz than reality” ⁴. In other words, ToolsGroup may have added probabilistic features (like outputting a range or some quantiles) but continued evaluating and thinking in deterministic terms internally.
Focus on Quantiles (Service Levels): From available information, ToolsGroup’s use of probabilistic data is largely in service of target service levels. A planner using ToolsGroup typically sets a desired service level (say 95% or 99%), and the system computes the required stock levels to achieve that, given the demand variability. This is done by looking at the upper quantiles of the demand distribution. For example, if 95% service is the goal, ToolsGroup will ensure inventory covers the 95th percentile of demand over lead time. This is effectively quantile forecasting: computing the 95th percentile demand. While useful, it’s a limited form of probabilistic forecasting – the system might compute one or two quantiles (e.g. P50 and P95) to set safety stocks, rather than leveraging the full distribution shape for a cost-optimized decision. In contrast, Lokad could consider the entire distribution and the costs of overstock vs stockout to choose an optimal quantile dynamically (maybe it’s the 87th percentile that maximizes expected profit, for instance), rather than just hitting a fixed service target.
No Broad AI/ML Breakthroughs: ToolsGroup integrates machine learning in its forecasting, but reviews indicate these are relatively standard techniques (regression, time-series models with perhaps some ML adjustment). ToolsGroup’s claims about “AI” have been met with skepticism ³⁶ ³⁷ – with experts noting that public info on ToolsGroup shows “pre-2000 forecasting models” in use ³⁸ (like Croston’s method for intermittent demand). The company hasn’t demonstrated any new probabilistic algorithms (for example, no evidence of deep learning-based probabilistic forecasts or participation in contests). Thus, its probabilistic forecasting can be seen as an extension of its classic statistical models (adding more outputs), rather than a new paradigm.

Verdict on ToolsGroup: ToolsGroup remains a credible, proven solution for inventory optimization with a long track record. It embraced probabilistic concepts decades ago, but the breadth of that embrace appears limited today. It models demand uncertainty and computes safety stocks accordingly, which is valuable. Yet by ignoring explicit lead time forecasting and by mixing probabilistic outputs with deterministic thinking (e.g. MAPE), ToolsGroup’s “probabilistic” approach lacks the completeness and rigor of Lokad’s ³². As one critical review put it, ToolsGroup’s use of modern buzzwords doesn’t always match the underlying reality – “mixing modern buzzwords with old-school techniques” ³⁸. Companies should understand that ToolsGroup will help determine inventory targets for a given service level under demand uncertainty, but it will not necessarily quantify all risks or re-optimize decisions under every stochastic scenario the way a full probabilistic optimization would.

Kinaxis – Deterministic Origins with New Quantile Forecasting and Probabilistic Inventory Add-ons

Background: Kinaxis is known for its RapidResponse (recently rebranded “Kinaxis Maestro”) platform, which for years has been a leader in supply chain planning, particularly Sales & Operations Planning (S&OP) and scenario analysis. Kinaxis’s hallmark is a fast, in-memory computation engine that allows planners to run real-time simulations – e.g., “what if demand jumps 10%” – and see impacts across the supply chain instantly. Historically, Kinaxis did not focus on being a forecasting engine itself; rather, it would consume forecasts from elsewhere or use simple methods, and emphasize integration and speed in re-planning supply, capacity, and inventory. Its planning was largely deterministic – scenarios were single versions of the truth, and safety stocks or buffers were set by planners or basic rules.

Shift toward Probabilistic Methods: In the last few years, Kinaxis has clearly recognized the industry trend (and customer demand) for more advanced forecasting and uncertainty management. Several developments illustrate this shift:

Wahupa Partnership (2022): In May 2022, Kinaxis announced a partnership with Wahupa – a small software firm specializing in probabilistic Multi-Echelon Inventory Optimization (MEIO) ³⁹. Wahupa’s engine is designed to quantify uncertainty (in demand and supply) and optimize inventory buffers across complex networks using probabilistic models. The idea was to embed Wahupa’s probabilistic MEIO inside Kinaxis’s platform (as a “solution extension”) ⁴⁰ ⁴¹. This would give Kinaxis customers a way to calculate inventory targets with probabilistic logic, rather than the traditional deterministic or rule-based methods. Kinaxis even co-authored blog content with Wahupa’s CEO discussing probabilistic vs. deterministic planning, showing they were evangelizing this approach ⁴² ⁴³. However, integrating Wahupa seems to have faced hurdles. By late 2023/early 2024, industry observers noted that this alliance had made little visible progress – the probabilistic MEIO capability was not yet widely referenced in Kinaxis case studies or user stories. Unconfirmed reports (and the tone from Lokad’s research) suggest the Wahupa integration did not gain traction and may have been quietly deprioritized or shelved, essentially a “small fiasco.” The Kinaxis 2024 Annual report still lists Wahupa as a partner, but there’s scant evidence of success. This underscores how challenging it can be for an established planning platform to bolt on a new probabilistic optimization engine not originally built into its data model ⁴⁴ ⁴⁵.
Rubikloud Acquisition (2020): Kinaxis acquired Rubikloud, a machine learning company that focused on retail demand forecasting and AI, in 2020. Rubikloud’s technology presumably included modern ML forecasting which could produce not just point forecasts but also prediction intervals (a basic probabilistic output). This acquisition was an early sign of Kinaxis trying to bolster its demand forecasting capabilities with AI. By 2023, Kinaxis began marketing a suite called “Planning.AI” which integrates advanced ML forecasting into the planning platform ⁴⁶. While details are limited, it’s likely that this enables Kinaxis to generate forecasts with associated confidence ranges (quantiles) for demand, feeding into its planning scenarios.
Patent for ML-based Quantile Forecasting (2023): In 2023, Kinaxis filed a U.S. patent application for a method of forecasting using machine learning models and determining a set of quantiles for an unknown demand sample ⁵ ⁴⁷. The patent describes training tree-based models on historical data and then generating synthetic values to compute a range of quantiles as the output forecast ⁴⁸ ⁴⁹. The goal is explicitly to “optimize inventory based on the set of quantiles” ⁵⁰. In essence, this is a formalization of quantile forecasting with ML – the result would be something like: for a given product, the model might output P50, P75, P90 demand forecasts (or any specified quantiles), and those can be used to decide inventory levels. That Kinaxis pursued a patent here indicates they are developing in-house IP for probabilistic demand forecasting (though, notably, quantile forecasts are still a subset of full distribution forecasts and a comparatively “weak” form of probabilistic modeling, as the user query correctly noted).
Cautious Messaging & Evolution: Kinaxis has been somewhat careful in its marketing around AI and probabilistic forecasting. Unlike some competitors, it hasn’t blanketed its branding with “AI magic” claims. Instead, it often emphasizes a combination of human and machine intelligence, and the term “concurrent planning” (real-time what-if). However, as it has added these new capabilities, Kinaxis is now more openly discussing uncertainty. The fact that a Kinaxis blog delved into probability theory concepts is a positive sign ⁴². Still, Kinaxis acknowledges it’s on a journey: “moving toward more decision automation under uncertainty, albeit from a deterministic legacy” ⁵¹. As of early 2025, the probabilistic features are nascent. Kinaxis has not, for example, participated in open forecasting competitions, nor published technical white papers demonstrating the efficacy of its new probabilistic engines. So the proof of maturity is limited ⁵².

What Kinaxis Currently Offers: With these changes, Kinaxis now offers (or will soon offer) two main probabilistic elements:

Probabilistic MEIO (via Wahupa): An inventory optimization that can compute optimal buffers by considering demand variability across multiple echelons. If successfully implemented, this would be analogous to the functionality of a ToolsGroup or SmartOps, but potentially more advanced if it uses Monte Carlo simulations or similar. It would answer questions like “Given the distribution of demand at each node and uncertainty in supply, what safety stock should I keep at each location to achieve X service or minimize cost?”.
ML-Based Demand Forecasting with Quantiles: An AI forecasting module (from Rubikloud/Planning.AI) that produces not just a forecast but also a range (e.g. a high-low band). This can improve planning by giving a sense of risk (e.g. there’s a 10% chance demand exceeds the P90 forecast). Planners or algorithms can then adjust plans accordingly (e.g. higher production to cover that tail if the risk is worth mitigating).

However, it’s important to note that these capabilities may currently be separate and not deeply unified in the platform. Kinaxis’s strength remains the ability to rapidly recalc plans given any new inputs. But if those inputs (forecasts, safety stock parameters) are now probabilistic, Kinaxis must ensure its UI and process can handle that (for example, presenting planners with not one plan but distributional outcomes). This is non-trivial.

Verdict on Kinaxis: Kinaxis is catching up to the probabilistic trend. It has made tangible moves (partnerships, acquisitions, R&D) to incorporate uncertainty modeling. Yet, at present, Kinaxis’s approach could be characterized as adding quantile forecasting and stochastic add-ons to a fundamentally scenario-based planning tool. It is not (yet) a paradigm where every calculation is done in distribution form internally. The company itself acknowledges the transition – they still emphasize that planners use the tool to evaluate scenarios and that automation is applied in a controlled way ⁵³ ⁵⁴ (e.g. auto-executing certain planning decisions if within thresholds, rather than fully hands-off). In short, Kinaxis’s probabilistic capabilities are incremental improvements rather than a wholesale reinvention. They provide more insights into uncertainty than before, but the core planning logic remains to be proven at the level of sophistication of Lokad’s approach. Prospective users should monitor how deeply integrated the Wahupa MEIO becomes (if at all) and whether Kinaxis can demonstrate that its quantile forecasts lead to better supply chain outcomes. As of 2025, one might say Kinaxis is probabilistic on paper, with patents and partnerships, but the substance is still building up.

Blue Yonder – Legacy Planning Giant Adopting Probabilistic Terminology Without Deep Changes

Background: Blue Yonder (BY), formerly known as JDA Software, is a heavyweight in supply chain software, providing solutions for demand forecasting, supply planning, merchandising, and more. It has a long lineage: JDA had acquired Manugistics and i2 Technologies, two big players from the early 2000s, inheriting their technologies. In 2020, JDA rebranded to Blue Yonder and was subsequently acquired by Panasonic. Blue Yonder’s modern platform is called Luminate, which aims to incorporate AI and a cloud architecture on top of these legacy modules.

Use of “Probabilistic” and AI in Marketing: Blue Yonder’s marketing in recent years has leaned heavily into buzzwords like “autonomous planning,” “cognitive supply chain,” and AI/ML. They have explicitly mentioned “probabilistic forecasting” in some contexts – for instance, the Azure Marketplace description for Blue Yonder’s Demand Forecasting mentions “autonomous and probabilistic forecasts” ⁵⁵. Blue Yonder also published blog articles on topics like probabilistic forecast evaluation (calibration, sharpness) ⁵⁶, indicating their data science teams are aware of these concepts. However, the key question is how much of this is theoretical discussion versus implemented product.

Reality in Products: The evidence suggests that Blue Yonder’s core forecasting and planning approach remains largely deterministic, with some enhancements:

Forecasting Engine: Blue Yonder’s Demand Planning (now part of Luminate Planning) traditionally uses time-series algorithms (like exponential smoothing, ARIMA) possibly augmented with machine learning for demand sensing. Blue Yonder has open-sourced or referenced certain tools like “tsfresh” (for feature extraction on time-series) and a library called “PyDSE” and “VikOS” which relate to ARIMA and optimization ⁵⁷. An analysis of Blue Yonder’s open-source contributions noted they rely on decades-old methods (ARIMA, regression) despite the AI marketing ⁵⁸ ⁵⁷. This indicates that under the hood, Blue Yonder is not using state-of-the-art probabilistic algorithms (like deep learning quantile regressors or probabilistic graphical models); instead, it’s likely using tried-and-true forecasting methods and maybe layering ML to adjust or select models.
Probabilistic Outputs: Does Blue Yonder actually produce probabilistic forecasts (distributions)? There’s little public evidence of it producing full distributions by default. They have talked about “dynamic safety stock” which implies recalculating safety stock levels based on forecast volatility – that might be misconstrued as probabilistic forecasting. One likely scenario: Blue Yonder produces a baseline forecast and an error distribution (e.g., computes forecast error variance). It could then compute, say, a P90 demand over lead time to set safety stock (similar to ToolsGroup’s approach). That would be a quantile, but not a full distribution report. Blue Yonder’s own literature on “cognitive inventory” essentially reframed probabilistic inventory optimization (stock levels based on probabilities) but provided “no technical support” for how it was different ⁵⁹. In summary, Blue Yonder knows the right words – they acknowledge that instead of static safety stocks you should account for demand variability dynamically (probabilistically). But the actual implementation likely falls back on traditional safety stock formulas (which assume demand variability is normally distributed, or use simple lookup of percentiles from a normal/Poisson distribution using forecast mean and variance).
Lead Time Uncertainty: We found no mention of Blue Yonder forecasting lead times or treating lead time as stochastic in its planning solutions. It’s safe to assume BY, like most, treats lead time as a fixed parameter (maybe with a cushion added by planners). Therefore, Blue Yonder also fails the full probabilistic criteria by ignoring supply-side uncertainty explicitly ⁹. It focuses on demand forecasting (and even there, mostly point forecasts).
Integration vs Innovation: Blue Yonder is essentially a collection of many modules. Some modules (like their ESP – Enterprise Supply Planning, or IO – Inventory Optimization) date back to Manugistics or i2 algorithms. These likely include multi-echelon inventory optimization that was state-of-art in early 2000s (which did consider demand variance analytically). But Blue Yonder’s challenge is that it’s not one unified engine but “a haphazard collection of products, most of them dated” ⁶⁰. They’ve attempted to modernize by slapping an AI layer (Luminate) on top, but often that amounts to dashboards and minor ML-driven improvements, rather than rewriting the core engines. So any probabilistic claims must be viewed in light of this patchwork: one part of BY might produce a forecast confidence interval; another part (the inventory optimization) might use a classic formula; another part might simply be deterministic supply planning. The consistency is questionable, and integrating a truly probabilistic workflow end-to-end would require significant refactoring which hasn’t been evident.
Vague AI Claims: Analysts have pointed out that Blue Yonder’s AI claims are vague and unsubstantial. For example, BY has mentioned using ML to augment probabilistic models ⁵⁷, but no details on what algorithms or how well they work. Blue Yonder did acquire some AI startups (like Blue Yonder GmbH, a German AI firm, which ironically is where they got the new name from, as JDA acquired that company in 2018). They also partner with some university research. But none of this has translated into clear, published breakthroughs in supply chain probabilistic forecasting. The marketing remains one level above the technical reality.

Verdict on Blue Yonder: Blue Yonder is very much an example of “buzzwords ahead of reality.” They use terms like probabilistic, cognitive, autonomous, but when pressed, their solutions seem to implement fairly standard forecasting and planning techniques ⁵⁷ ⁶¹. To be fair, Blue Yonder has immense domain experience and a broad suite – their strength is being able to cover everything from demand forecasting to fulfillment, with lots of industry-specific capabilities (like specialized retail planning features). But in the narrow context of probabilistic forecasting: Blue Yonder knows of it and has probably added some quantile-based safety stock logic and ML forecasting enhancements. It has not demonstrated the kind of probabilistic optimization that Lokad does (where every decision is derived from simulated scenarios). An internal critique summarized it well: Blue Yonder’s use of “probabilistic” in cognitive inventory was essentially rebranding, “rehashed… with fancy terms” but no new algorithms ⁵⁹. Companies considering Blue Yonder should not assume they will get a cutting-edge probabilistic forecasting engine; rather, they will get a solid, if somewhat dated, forecasting toolset with a modern UI, and any probabilistic benefits will come from incremental improvements (like more frequent forecast updates, perhaps some automated exception management) rather than fundamental stochastic optimization.

SAP IBP – Powerful but Complex Suite with Probabilistic Roots (SmartOps) Largely Underutilized

Background: SAP’s Integrated Business Planning (IBP) is the successor to SAP’s APO (Advanced Planning & Optimization) and incorporates various modules for demand forecasting, supply planning, and inventory optimization. SAP, being an ERP giant, often competes on the promise of an end-to-end integrated platform (from finance and sales all the way to supply chain execution). Historically, SAP’s planning tools were mostly deterministic: APO demand planning gave point forecasts; APO inventory planning computed safety stocks using simple formulas or at best single-stage calculations. Recognizing a gap, SAP made strategic acquisitions: SmartOps in 2013 (which was a leading multi-echelon inventory optimization company known for probabilistic models), and earlier SAF AG in 2009 (a demand forecasting firm). SmartOps, in particular, brought in a probabilistic engine that could optimize inventory across multiple echelons (locations) to meet service levels at minimum cost ¹⁰. Essentially, SmartOps mathematically modeled demand uncertainty (and to some extent lead time variability via some assumptions) to recommend safety stock. This became SAP’s IBP for Inventory module.

Current State of Probabilistic Features in SAP:

Demand Forecasting: SAP IBP has a Demand module which can use advanced statistical methods and even machine learning (SAP has an Analytics Cloud that can be used for forecasting, including techniques like gradient boosting, etc.). However, SAP usually presents these as improving forecast accuracy (MAPE, etc.), not as providing full probability distributions. The typical output is still a single forecast (with perhaps some consensus process around it). The notion of probabilistic demand forecasts isn’t front-and-center; SAP rarely if ever uses that term in marketing. So, while SAP IBP’s demand planning could produce an error measure, it doesn’t natively output a distribution for consumption by the rest of the system.
Inventory Optimization (SmartOps): In principle, SmartOps inside IBP does use probabilistic models. SmartOps was known for computing probability distributions of inventory positions and solving for the optimal stock levels. It considered both demand variability and supply variability (the latter in a limited way, often using a “safety time” or a variability factor on lead time). If IBP for Inventory is implemented, a company can input forecast error distributions and lead time variability and get back recommended safety stocks at each location that achieve a target service level with minimum inventory. This is a form of stochastic optimization, albeit one that’s oriented to service levels (much like ToolsGroup’s approach, but multi-echelon). However, few SAP customers actually deploy this to its full extent. One reason is complexity: IBP is modular, and not every implementation includes the inventory optimization piece (some might just do S&OP and demand planning). Another reason is usability: the SmartOps algorithms might be a black box and require a lot of statistical data that companies find hard to provide or maintain (e.g., you need a good sense of your forecast error distribution by item, which not everyone has readily).
Probabilistic Emphasis (or Lack Thereof): SAP does not emphasize probabilistic forecasting in its messaging, as noted. They sell IBP on integration (“one version of the truth”), on scenario planning, on being cloud-based, etc., rather than claiming they have the best forecasting algorithms. In fact, SAP’s reputation is more about breadth than depth: “function-rich but not algorithmically advanced” ⁶². This seems to be an admission that, while SAP has the pieces (like SmartOps’s math), it hasn’t pushed them further. A critique mentioned that even with acquisitions like KXEN (a predictive analytics firm SAP bought in 2013), SAP’s forecasts aren’t necessarily better than using traditional methods ⁶³. A research note pointed out that modern ML methods have not clearly outperformed the older stat models in this domain, implying SAP’s integration of those acquisitions did not yield dramatic improvements ⁶³.
Lead Time and Other Uncertainty: SAP’s standard planning assumes fixed lead times. If one uses the inventory optimizer, one can input a lead time variability (or just inflate the lead time to a higher percentile). But SAP doesn’t automatically learn a lead time distribution from data. It expects users to provide a “service time” (like a percentile lead time) as a parameter if needed. So again, like others, SAP fails to internally generate probabilistic forecasts for supply; it relies on configuration.
Complexity and Fragmentation: One major drawback is that SAP’s solution is fragmented due to multiple acquisitions. The demand forecast might come from one module (with one set of assumptions), and the inventory optimization from another. If not perfectly aligned, you might feed a point forecast from IBP Demand into IBP Inventory, which then internally assumes a normal distribution of demand with a given standard deviation. If the forecast error doesn’t follow that assumption, the inventory outputs might be suboptimal. A quote from a review: “enterprise software isn’t miscible through M&A” – SAP’s pieces didn’t seamlessly mix ⁶⁴. So unless an implementation is carefully tuned, the probabilistic aspect can get lost or produce inconsistent results. In many cases, companies with SAP end up simplifying – e.g., using IBP to calculate some safety stock using a rough stat model and not fully utilizing SmartOps, or even turning off the optimizer because it’s too hard to trust/tune, and instead setting safety stock via simpler rules.

Verdict on SAP IBP: SAP IBP theoretically has probabilistic optimization capabilities (via SmartOps) that put it closer to Lokad’s camp than some others. But in practice, SAP does not actively champion or evolve these capabilities, and many customers may not realize or leverage them ¹¹. The probabilistic piece is effectively a bolt-on to satisfy feature checklists (for those who need multi-echelon inventory optimization). SAP’s main selling proposition is that it’s part of a comprehensive SAP ecosystem – not that it’s the most advanced analytics engine. Thus, compared to Lokad, SAP IBP’s use of probabilistic methods is incidental and somewhat stagnant. Companies considering SAP IBP for probabilistic forecasting should ensure they specifically implement the inventory optimization module and have the significant expert services (or consultants) needed to calibrate it. Indeed, it’s said that to succeed with SAP’s advanced planning features, one needs “the very best integrators – plus a few years” ⁶⁵. This underscores that SAP’s probabilistic tech, while present, is buried under complexity and not readily delivering value out-of-the-box.

o9 Solutions – Integrated “Digital Brain” Platform with Limited Emphasis on Probability Distributions

Background: o9 Solutions is a younger vendor (founded 2009 by former i2 executives) that has quickly gained a reputation as a “next-generation” planning platform. o9’s focus is on creating a unified model of the business (Enterprise Knowledge Graph) and performing real-time planning across demand, supply, and financial domains. It’s very buzzword-friendly: o9 markets concepts like a Digital Twin of the organization, real-time scenario planning, and recently even touts generative AI features. Given its origin, o9 is somewhat like a spiritual successor to i2 Technologies, emphasizing integrated planning over siloed tools.

Approach to Forecasting and Uncertainty: o9 certainly uses advanced analytics for forecasting, but it does not appear to champion probabilistic forecasting as a unique selling point. Notable points:

Predictive Analytics with ML: o9 provides statistical forecasting and machine learning for demand sensing. Their case studies mention using a variety of data (point of sale, weather, web searches, etc.) to improve short-term forecasts ⁶⁶. This implies o9 is tackling the accuracy side of forecasting – trying to get better point forecasts by incorporating more signals (the “sensing” approach). There is mention of scenario simulation: o9’s “Supply Sensing” monitors external factors and can simulate their impact on supply chain ⁶⁶. While this acknowledges uncertainty (you simulate various scenarios), it’s still in a deterministic way (each scenario is a what-if, not a probability distribution over many what-ifs automatically considered).
Enterprise Knowledge Graph: o9’s use of a graph-based model (which captures relationships between products, locations, customers, etc.) is powerful for scenario analysis. For example, if a certain component is delayed, the graph can quickly show all affected products and let planners recompute a plan. But again, this is not inherently probabilistic – it’s a structural data model that aids speed and insight.
No Evidence of Probabilistic Outputs: We haven’t seen evidence that o9 outputs full probability distributions for forecasts. Their public materials (and a Lokad skeptical review) suggest that o9 sticks to point forecasts and deterministic optimization, albeit very fast and integrated ¹². They highlight what-if analysis – which implies you can manually vary inputs to see different outcomes, rather than the system automatically quantifying uncertainty. Essentially, o9 might equip the user to explore different possibilities, but it doesn’t automatically tell you the probability of each scenario or optimize for the best expected outcome. The heavy use of the term “Digital Twin” and “real-time scenario planning” points to an interactive planning mentality: planners generate scenarios, the system quickly calculates results (e.g., if demand increases by 10%, how does inventory look?). That’s slightly different from Lokad’s approach where the system itself is crunching through thousands of scenarios in the background to pick an optimal plan.
Comparison with Lokad: A direct comparison noted that o9 emphasizes breadth (unifying many planning functions, incorporating IoT data, etc.) whereas Lokad emphasizes depth in quantitative optimization ¹². o9’s strength is giving large enterprises a one-stop platform that connects all data and plans (demand, supply, finance) and allows cross-functional collaboration. Lokad’s strength is solving specific numerical problems (like inventory or production optimization) with as much mathematical rigor as possible (probabilistic modeling, custom algorithms). In other words, o9 provides a wider planning lens – but possibly uses conventional forecasting inside – while Lokad provides a narrower but more advanced analytical lens. As the Lokad review put it, o9 is about “synthesizing various planning functions into a unified framework, while Lokad generates precise, actionable recommendations through deep analytical automation” ¹².
Recent Developments: o9 is certainly adding AI elements (they mention generative AI to query the system, etc.), but it’s unclear if they are adding probabilistic forecasting. It’s possible o9 might eventually incorporate some probabilistic forecasting library or at least allow integration with Python/R where one could do such modeling. But as of now, their differentiation does not lie in having invented new forecasting algorithms; it’s more in how they deliver the software (cloud-based, real-time, user-friendly analytics).

Verdict on o9: o9 Solutions is a formidable planning tool for organizations that want a single platform to do many things quickly and collaboratively. However, when it comes to probabilistic forecasting specifically, o9 does not appear to lead or innovate in that area. It likely provides “good enough” forecasting using ML and focuses on making the results immediately usable in planning scenarios. The mention of probabilistic forecasting is almost absent in their public content. If one were to implement o9 and desired probabilistic forecasts, one might have to generate those outside the system or via custom code. In summary, o9’s value proposition is agility and integration, not advanced stochastic optimization. Companies that prioritize a rigorous probabilistic approach might complement o9 with external data science efforts or consider whether a specialized tool (like Lokad) is more suitable for that particular need.

RELEX Solutions – Retail Specialist with AI Claims, Using Deterministic Forecasting + Fast Replanning

Background: RELEX Solutions is a Finnish company (founded 2005) that has grown rapidly, especially in the retail and grocery sector. RELEX’s platform covers demand forecasting, replenishment, allocation, and planogram optimization, with a particular focus on handling fresh food (perishables), promotions, and store-level planning. RELEX often wins with retailers due to its ability to do very granular, high-frequency planning (store/SKU daily forecasts, intraday, etc.) and its user-friendly analytics (real-time dashboards of inventory levels, etc.). They market themselves as highly automated and “AI-driven.”

Probabilistic Aspects (or Lack Thereof): RELEX’s approach to forecasting and inventory is, at its core, focused on speed and frequency rather than explicit stochastic modeling:

Frequent Re-forecasting (Demand Sensing): RELEX emphasizes their ability to sense demand changes and update forecasts quickly. For example, ingesting yesterday’s sales, weather changes, social media trends, etc., to adjust short-term forecasts (this is often termed demand sensing, a buzzword in itself). This can reduce forecast error in the short horizon (e.g., better react to a sudden drop or spike in sales). However, this is still producing a single updated forecast (just more frequently or with more data), not a distribution of forecasts. RELEX’s claim of “AI-driven forecasting” usually refers to using machine learning models (like gradient boosting, neural nets, etc.) on recent data to improve accuracy ¹³. Nowhere do they claim they output full probability distributions for each SKU’s demand. Indeed, an analysis of RELEX found “no evidence of generating full probability distributions for demand the way Lokad does” ¹³. It appears RELEX’s validation of success is still via traditional accuracy metrics (how close the forecast is to actuals), which is a deterministic mindset.
Inventory Optimization Method: RELEX does multi-echelon replenishment and allocation (especially push allocation for promotions, etc.), but how does it buffer against uncertainty? It likely uses traditional safety stock calculations behind the scenes. The system knows lead times and demand volatility (e.g., via a moving average of forecast error or a service level setting) and computes safety stock per item-location accordingly (like many systems do). The lead times in RELEX are inputs, not random variables the system forecasts ¹⁴. If a user knows a certain supplier is inconsistent, they might adjust the lead time or its safety margin manually. RELEX’s documentation doesn’t highlight stochastic optimization or Monte Carlo, so we infer it does what legacy tools did: for each SKU, decide on safety stock using an assumed distribution (perhaps normal or Poisson) based on forecast error, to achieve a target service level. This is a deterministic formulaic approach, not a simulation of many outcomes. In fact, RELEX’s heavy orientation to in-memory processing (OLAP-style) suggests it prioritizes fast querying over deep computation ⁶⁷ ⁶⁸. As one comment noted, their architecture being like an OLAP cube is “at odds with network-wide optimization” for things like substitutions or complex stochastic problems ⁶⁹. The design that gives super fast dashboards might not be ideal for running large Monte Carlo simulations on uncertainty – it may instead rely on analytical (simpler) methods that fit in memory.
Claims vs. Evidence: RELEX’s marketing uses the buzzwords: “autonomous supply chain, AI-driven, machine learning, digital twin,” etc. ⁷⁰. But they rarely specify which algorithms or provide case studies quantifying the impact of their AI beyond anecdotes. Third-party analyses have cast doubt: e.g., noting that RELEX’s forecasting tech “appears to be pre-2000 models” (meaning they likely use classical methods with some ML wrapper, rather than fundamentally new approaches) ¹⁵. They also likely still use concepts like MAPE or bias to measure forecast quality internally, which is a sign of deterministic thinking.
Handling of Uncertainty in Specific Scenarios: RELEX does excel in some areas that involve uncertainty, but often through business rules or heuristic approaches rather than probabilistic math. For example:
- Fresh Food and Expiry: RELEX has features to manage expiration dates on products, ensuring first-expiry-first-out, suggesting replenishment to avoid waste, etc. ⁷¹. While this deals with uncertainty of demand before spoilage, the solution is more heuristic (monitor days of supply vs shelf life) than probabilistic forecasting of spoilage risk.
- Promotion Effects and Cannibalization: RELEX can model promotion uplifts and perhaps some cannibalization (one product stealing sales from another), but the analysis suggests RELEX’s OLAP base makes it hard to do sophisticated modeling of these interactions ⁶⁹. They might just let planners manually adjust or use simple regression for promo uplift.
- “Intelligent” Order Batching: They mention things like optimizing truckloads or forward-buying for discounts ⁷². These are valuable features that optimize cost given constraints, but again, likely assuming known demand (or at least mean demand) rather than explicitly hedging uncertainty.
No Full Stochastic Optimization: Crucially, RELEX does not advertise anything akin to Lokad’s stochastic optimization. There’s no talk of solving an expected cost minimization under uncertainty; it’s more about ensuring high availability (they often tout clients reaching ~98-99% in-stock). In fact, some of their claimed results (99%+ availability) are viewed skeptically in the industry because average on-shelf availability tends to be lower ⁷³. It implies that maybe they achieved that in controlled pilots, or it’s an optimistic scenario, not necessarily broadly reproducible. This again hints that their approach might not fundamentally change the game on uncertainty, just improve execution and monitoring to reduce stockouts.

Verdict on RELEX: RELEX is a fast, responsive supply chain tool especially suited for retail, where daily agility is important. It automates many decisions (store replenishments, for example) with ease of use. However, its scientific depth in probabilistic forecasting is limited. It leverages ML for better forecasts, but not to produce the kind of probability distributions that would feed a Lokad-style optimizer ¹³. It largely relies on the same safety stock logic and forecast improvements that others do, albeit executed very efficiently at scale and frequency. For a retailer, RELEX might yield better shelf availability simply by re-planning more often and using recent data (a pragmatic approach), whereas Lokad might achieve it by fundamentally optimizing stock levels with probabilistic calculations. They’re different philosophies: one is “keep fixing the plan as reality unfolds” (RELEX’s near-real-time mindset), the other is “plan optimally from the outset given what could happen” (Lokad’s ex-ante optimization). Both can coexist, but if the question is about probabilistic forecasting per se, RELEX is not a trailblazer there. It has copied the buzzwords of AI, and perhaps implicitly uses some probabilistic concepts (like variance for safety stock), but it does not offer the kind of explicit probabilistic solution that Lokad does.

Conclusion

In the supply chain software market, “probabilistic” has become a bit of a buzzword, much as “AI” and “machine learning” have. Lokad stands out as the vendor that genuinely built its solution around probabilistic principles – treating uncertainty as a first-class citizen in both forecasting and optimization. Since pioneering this approach in 2012–2015, Lokad influenced the narrative of the industry. Competitors have, to varying degrees, borrowed the terminology and acknowledged the importance of uncertainty, but their solutions reveal a substantial gap in implementation.

To summarize the contrasts:

Lokad provides full probabilistic forecasts (demand and supply) and uses them in a custom-built optimization engine. It explicitly quantifies uncertainties and produces decisions that are optimized for expected outcomes ². In other words, it “walks the talk” of probabilistic supply chain management.
Others (ToolsGroup, Blue Yonder, SAP, RELEX, o9, Kinaxis) mostly provide partial or pseudo-probabilistic features:
- ToolsGroup and SAP have had probabilistic inventory modeling components for years, but they focus on demand variability and still treat key inputs (like lead time) as fixed or simplistic, which limits their probabilistic rigor ³ ¹¹.
- Blue Yonder and RELEX, despite new AI branding, continue to rely on traditional forecasting methods and deterministic planning logic, with “probabilistic” appearing mainly in marketing or minor tooling (like recalculating safety stocks more dynamically) ⁸ ¹³.
- Kinaxis and o9, as modern platforms, have begun integrating probabilistic ideas (Kinaxis via partners/patents, o9 via scenario flexibility) but are still largely evolving from deterministic foundations. Kinaxis’s quantile forecasts are a step forward, yet they represent forecasting at a few percentile points rather than the entire distribution, which is a weaker substitute for full probabilistic forecasts ⁵⁰ ⁴⁹.

A telling sign of superficial adoption is when vendors claim “probabilistic” or “stochastic” but then continue to use deterministic metrics or processes. The earlier example of ToolsGroup using MAPE for probabilistic forecasts is one such red flag ⁴. Similarly, any vendor that ignores lead time uncertainty or does not mention how they evaluate probabilistic accuracy likely hasn’t truly embraced the paradigm. Lokad’s criticism of these inconsistencies is backed by examples across the industry: “Claims of probabilistic forecasts are advertised alongside MAPE reductions, which is inconsistent… vendors sprinkle terms like ‘stochastic’ in brochures while remaining in a deterministic world.” ⁴ ⁷⁴. The substance often lags the slogans.

From a recent developments perspective, it’s clear that Lokad’s competitors are now paying lip service to probabilistic methods – which can be seen as a validation of Lokad’s approach. However, the degree of actual capability varies. Companies evaluating solutions should dig into questions like: Does the tool forecast a distribution or just a few scenarios? Does it optimize decisions using Monte Carlo simulation or just fixed safety stock formulas? Can it account for all major uncertainties or only demand? The answers often reveal that many vendors still essentially do what they always did, with a thin probabilistic veneer.

In conclusion, Lokad vs. the rest on probabilistic supply chain optimization is a story of depth vs. breadth, and of early innovation vs. catch-up. Lokad offers a deep, quantitatively rigorous solution purpose-built around uncertainty. The other major vendors offer broader supply chain planning suites where probabilistic features are add-ons or evolving components – in many cases, these are reactive measures to market demand for more advanced analytics, without overhauling their core methodologies. Executives should be wary of “probabilistic” as a buzzword and look for concrete technical evidence of how a vendor implements it. As of 2025, Lokad remains distinctive in providing true probabilistic forecasting and optimization, while most competitors either stick to deterministic approaches or provide only limited probabilistic capabilities that fall short of Lokad’s comprehensive treatment of uncertainty ³² ⁸.