From ARIMA to Deep Learning: How the Forecasting Stack Evolved

Supply chain demand forecasting using AI has undergone three distinct phases of development since the earliest networked forecasting systems appeared in the early 2000s. Each phase expanded both the algorithmic sophistication and the organizational scope of what forecasting systems could do — moving from rule-based, retrospective reporting to anticipatory, self-correcting prediction engines that now integrate dozens of external data signals simultaneously.

The foundational era, spanning 2002 to 2014, was defined by collaborative, network-based frameworks. Restaurant Services, Inc. introduced global forecast distribution across outlet networks in 2002, while SAP SE added constraint-aware sourcing logic in 2006. A pivotal 2014 academic contribution on proactive supply chain performance management introduced KPI-linked predictive data mining — signalling the transition from retrospective reporting to anticipatory management. These early systems established the architectural vocabulary that AI-augmented platforms would later extend.

Between 2017 and 2022, academic and industrial attention intensified around machine learning benchmarking. Two critical literature studies — one in 2021 on LSTM networks for time-series forecasting, another in 2022 benchmarking deep learning methods for Supply Chain Management 4.0 — demonstrated that LSTM consistently outperformed ARIMA on real supply chain transaction data. Kinaxis filed a series of active US patents on ML-based lead-time forecasting during this period, incorporating weather and financial indicator features, while Target Brands and Coupa Software filed their first automated replenishment and demand forecasting patents.

Since 2023, the most significant shift has been a dual pivot: first, toward generative AI and multi-modal data fusion; second, toward sector-specific platforms for healthcare, defense, fashion, and FMCG. India has emerged as the most active filing jurisdiction by volume, concentrated in academic institutions and individual inventors, while US corporate assignees — Dell, HP Enterprise, Walmart Apollo, Honeywell — continue to file the commercially maintained patents most relevant to enterprise procurement decisions. According to WIPO, AI-related patent filings globally have grown substantially in recent years, with supply chain applications among the fastest-growing sub-categories.

Four Technology Clusters Defining AI Demand Forecasting in 2026

The AI supply chain demand forecasting patent landscape organises into four distinct technology clusters, each representing a different level of architectural maturity and a different primary innovation driver. Understanding which cluster a filing belongs to determines both its competitive positioning and its IP defensibility.

Cluster 1: Deep Learning Time-Series Forecasting (LSTM, Transformers, GRU)

Deep learning sequence models are the dominant algorithmic cluster in this dataset. LSTM networks capture long-range temporal dependencies that ARIMA and exponential smoothing cannot model, and transformer-based architectures extend this capability to multi-variate, multi-horizon demand prediction. A 2025 Indian filing from Pritam Roy claims a forecasting engine combining LSTM, Transformer models, and Langraph Retrieval-Augmented Generation (RAG) for multi-variable time-series demand prediction that also incorporates social sentiment data — reflecting how this cluster is absorbing techniques from adjacent AI sub-fields. The Nitte Meenakshi Institute of Technology filing (2025) uses LSTM and GRU networks fed by CRM data with real-time ERP integration and periodic model retraining, addressing one of the most common enterprise deployment challenges: model staleness.

Cluster 2: Multi-Signal Machine Learning with External Data Fusion

The second cluster covers systems that integrate IoT sensor streams, weather data, macroeconomic indicators, social media sentiment, and supplier performance metrics as forecasting features alongside historical sales records. Kinaxis Inc.’s ML forecasting module ingests historical lead time data, weather data, and economic indicator data through dedicated preparation pipelines — a design pattern that multiple 2024–2026 filings have elaborated on. JIS College of Engineering’s “Quantreal AI” system (2025, IN) integrates social media sentiment, stock market indicators, weather patterns, and retailer feedback into an adaptive demand prediction system optimised specifically for FMCG sectors. Dell Products L.P.’s temporal supply-related forecasting patent (active January 2026) addresses delivery commitment consistency using AI applied to supply-side visibility data for demand-supply matching.

Cluster 3: Multi-Scenario AI Planning and Automated Decision Support

The third cluster moves beyond single-point demand prediction toward probabilistic scenario generation — producing multiple demand outcomes that supply chain planners can evaluate and from which autonomous systems can trigger pre-defined responses. The most architecturally complete filing in this cluster is a 2024 WO patent by Marc Escande, which employs three distinct ML models — one each for demand signal processing, supply signal processing, and scenario generation — with company-attribute-based filtering. Walmart Apollo’s 2025 US patent introduces graph neural network approaches that represent supply chain entities as nodes and transactions as edges, enabling relational demand inference rather than purely time-series-based prediction. Target Brands’ 2023 replenishment simulation patent simulates per-epoch run states across supply chain network nodes using sampled demand distributions, enabling stochastic inventory policy optimisation.

Cluster 4: Explainable AI (XAI) and Generative AI

The newest and most rapidly emerging cluster combines two capabilities that address distinct enterprise barriers: explainability techniques that make forecasts auditable, and generative AI models capable of synthesising qualitative signals — emotional sentiment, political events, consumer behaviour — into demand predictions. Lovely Professional University’s 2025 hardware-software system integrates IoT sensors, RFID, and POS terminals with a forecasting engine featuring explicit SHAP, LIME, and counterfactual reasoning layers enabling real-time “what-if” analyses for pricing and promotions. SoftBank Group’s February 2026 filing — the most recent in this dataset — claims a generative AI system that analyses consumer social media posts, emotional sentiment, weather events, and political developments for real-time demand prediction, with a separate market trend prediction module incorporating political events and policy changes.

“With LSTM and Transformer-based forecasting maturing toward commodity status, generative AI with multi-modal signal fusion — consumer emotional sentiment, weather events, political developments — now represents the next defensible differentiation layer in supply chain demand forecasting.”

Who Holds the IP: Assignee and Jurisdiction Landscape

India dominates patent filing volume in this dataset with approximately 30+ filings, while the United States accounts for approximately 15+ filings that hold the highest concentration of commercially maintained, active patents. This asymmetry between IN (high volume, predominantly pending, institution and individual-led) and US (lower volume, largely active, large corporate assignees) defines where the commercially actionable IP white space currently sits — concentrated in the US and WO jurisdictions.

Dell Products L.P. is the most prolific single corporate assignee in this dataset, with at least 5 active US patents covering AI demand prioritization, temporal supply forecasting, delivery date generation, and data integration demand management — signalling a sustained, multi-year IP strategy in enterprise supply chain AI. Kinaxis Inc. holds 4+ active US patents on ML-based lead-time and supply chain design correction incorporating external data signals, backed by its commercially deployed RapidResponse platform. Strong Force VCN Portfolio 2019, LLC accounts for 4+ US and AU filings covering AI-driven predictive sourcing, value chain network modelling, robot fleet-demand coordination, and ML-driven out-of-stock resolution — reflecting a portfolio aggregation strategy. Target Brands holds 2 active US patents on stochastic replenishment simulation, and Walmart Apollo, LLC filed its graph neural network supply chain modelling patent in 2025.

Among established technology vendors, SAP SE holds 2 active US patents on constraint-aware demand forecast sourcing dating to 2006, Hewlett Packard Enterprise Development LP has 2 active US patents on ML customer demand forecasting with the most recent dated March 2026, and Honeywell International Inc. has 2 pending filings on AI-based product genealogy and supplier data mapping. Coupa Software Incorporated holds 2 filings (2022 and 2025) on automated changepoint-aware demand forecasting, and Wipro Limited has 2 pending filings on disruption-resilient ML demand prediction. According to EPO data, AI patent filings in logistics and supply chain have accelerated significantly since 2020, consistent with the concentration of activity seen in this dataset.

Where AI Forecasting Is Being Deployed: Sector-by-Sector Breakdown

AI supply chain demand forecasting is not a monolithic technology — it fragments by sector along the lines of domain-specific training data requirements, constraint sets, and regulatory expectations. The dataset reveals six distinct application verticals, each with characteristic IP patterns.

Retail and E-Commerce is the largest application sector. Filings address FMCG inventory, promotional event demand spikes, and cross-border e-commerce replenishment. Cognira LLC’s 2022 US patent specifically targets fashion and apparel with short, unpredictable product lifecycle forecasting — a vertical where demand volatility makes traditional statistical methods particularly unreliable. Academic work published in 2023 demonstrated XGBoost achieving an RMSE of 46.64% with a 9-minute computation time on cross-border e-commerce inventory data, outperforming competing algorithms for this high-SKU, high-velocity use case.

Manufacturing and Industrial Supply Chains are addressed by Kinaxis’s multi-feature ML lead-time variability system, Bennett University’s IoT-integrated ML for sensor manufacturing (2025, IN), and RK University’s neural predictive systems for broad industrial SCM optimisation (2025, IN). The common thread across these filings is that manufacturing supply chains have longer lead times and higher cost-of-error, making probabilistic forecasting with uncertainty quantification more valuable than point predictions.

Healthcare and Medical Supply Chains represent a specialised vertical where stockout risk has direct patient safety implications. SR University’s 2026 IN filing claims a system that analyses historical consumption data, clinical activity indicators, seasonal variations, and external influencing factors with dynamic model updating — addressing the regulatory and operational requirements that distinguish medical supply chains from commercial ones.

Defense and Government Logistics are the subject of a 2025 IN filing by Dr. Rupesh Shukla applying LSTM, reinforcement learning, and optimisation heuristics to Indian Air Force supply chain forecasting, including wartime surge simulation, supply disruption scenarios, and route optimisation for military logistics. This represents one of the more unusual application contexts in the dataset, where demand “spikes” are not promotional events but operational exigencies.

Technology Hardware and Enterprise IT are served by Dell’s sustained patent portfolio, HP Enterprise’s cross-functional ML demand forecasting system, and Honeywell’s AI-based product genealogy and supplier data mapping. These filings reflect the complexity of hardware supply chains with long component lead times, component obsolescence risk, and multi-tier supplier networks.

Food Service supply chains were among the earliest to be addressed by networked AI forecasting. Restaurant Services, Inc. established a foundational collaborative forecasting framework in 2002–2003 covering real-time accuracy feedback and network-based store-level forecasting — a framework that later AI-augmented systems have extended architecturally.

Emerging Directions: Generative AI, Graph Networks, and Autonomous Procurement

Six emerging directions define the near-term frontier of AI supply chain demand forecasting, each reflected in patent filings from 2024 to early 2026. These represent both the next wave of innovation and the highest-opportunity areas for IP strategy.

Generative AI for Demand Forecasting

The SoftBank Group filing from February 2026 — the most recent in this dataset — claims a generative AI system that analyses consumer social media posts, emotional sentiment data, weather events, and political developments for real-time demand prediction, with a market trend prediction module incorporating political events and policy changes. This is a qualitatively different capability from traditional ML forecasting: rather than fitting patterns in historical transaction data, generative AI can synthesise novel demand signals from unstructured sources that have no historical equivalent. The 2025 IN filing for GAI-SCOS (Generative AI-Driven Supply Chain Optimisation System) corroborates this direction with generative AI applied to demand forecasting, inventory management, and route planning simultaneously.

Explainable AI for Forecast Trustworthiness

Enterprise adoption barriers around “black box” AI models are driving investment in explainability. Lovely Professional University’s 2025 patent embeds SHAP, LIME, and counterfactual reasoning layers directly into the forecasting pipeline — enabling real-time “what-if” analyses for pricing and promotions decisions. This architectural choice reflects a broader shift: explainability is increasingly a procurement criterion and a regulatory expectation rather than an optional UX feature, particularly for regulated industries such as FDA-regulated medical supply chains where decision traceability is mandated.

Digital Twin Integration

Strong Force VCN Portfolio 2019, LLC’s 2025 US patent on value chain network planning using ML and digital twin simulation — alongside a 2020 academic paper on digital twin-driven supply chain planning — demonstrates convergence of ML forecasting with digital twin simulation environments. These systems enable dynamic synchronisation between physical inventory states and virtual demand models, allowing organisations to test policy changes against simulated demand distributions before committing to them operationally. This is particularly valuable for capital-intensive supply chains where the cost of a wrong inventory decision is high.

Graph-Based Supply Chain Modelling

Walmart Apollo’s 2025 US patent introduces graph neural network approaches that represent supply chain entities as nodes and transactions as edges, enabling relational demand inference rather than purely time-series-based prediction. This is architecturally significant: graph models can capture network-level demand propagation effects — where a disruption at one node affects demand signals at adjacent nodes — that LSTM and Transformer models operating on single time-series cannot. Only Walmart Apollo has filed in this specific architectural space within this dataset, suggesting limited competition for novel graph-based supply chain AI claims relative to the crowded LSTM cluster.

Autonomous Procurement Execution

Strong Force VCN Portfolio 2019, LLC’s 2025 US patent on ML-driven out-of-stock inventory resolution and contract negotiation extends forecasting into closed-loop autonomous procurement — where AI not only predicts demand gaps but executes procurement actions. This closes the loop between prediction and operation, transforming forecasting from a decision-support function into an autonomous execution layer. R&D teams entering this space should architect for end-to-end automation, not standalone forecasting modules.

Sector-Specific AI Forecasting Platforms

The dataset shows clear specialisation emerging for healthcare (SR University, 2026), defense logistics (Dr. Rupesh Shukla, 2025), and fashion (Cognira LLC, 2021–2022). Domain-specific training data, constraint sets, and regulatory requirements are driving vertical fragmentation — meaning that a generic ML forecasting platform is increasingly less defensible than one with deep domain-specific architecture. Research bodies including OECD have highlighted sector-specific AI adoption as a key driver of productivity gains in supply chain management.

“Graph-based and simulation-based forecasting represent underexplored IP territory — only Walmart Apollo and Target Brands have filed in these architectural spaces within this dataset, suggesting limited competition for novel claims relative to the crowded LSTM and Transformer cluster.”

Strategic Implications for R&D and IP Teams

The AI supply chain demand forecasting patent landscape in 2026 carries several concrete strategic signals for R&D leaders, IP strategists, and technology evaluators — derived directly from the filing patterns and architectural innovations documented in this dataset.

• Architect for end-to-end automation, not standalone forecasting modules. Leading filers — Dell, Strong Force VCN, Walmart Apollo — are patenting integrated systems where AI forecasts trigger autonomous inventory replenishment, supplier selection, and contract negotiation. The forecasting-to-execution loop is closing rapidly, and standalone prediction components will be increasingly commoditised.
• Treat explainability as a differentiation and defensibility vector. The emergence of XAI-specific filings embedding SHAP, LIME, and counterfactual reasoning reflects enterprise procurement criteria and regulatory expectations. IP strategists should evaluate XAI integration as a feature of architecture patents, not a UX layer added post-launch.
• Prioritise US and WO jurisdiction for maximum IP protection. The asymmetry between IN (high volume, mostly pending, institution and individual-led) and US (lower volume, predominantly active, large corporate assignees) means new entrants seeking commercial defensibility should concentrate filing resources in the US and WO jurisdictions.
• Graph-based and simulation-based forecasting offer genuine IP white space. Only Walmart Apollo (graph neural networks) and Target Brands (stochastic simulation) have filed in these architectural spaces within this dataset — a substantially lower competitive density than the LSTM and Transformer forecasting cluster.
• Generative AI and multi-modal signal fusion are the next defensible differentiation layer. With LSTM and Transformer-based forecasting maturing toward commodity status, the filings by SoftBank (generative AI, emotional sentiment, political events) and JIS College of Engineering (multi-spectral data including stock market and social signals) point toward competitive moats rooted in proprietary data pipelines and multi-modal model architecture patents.
• Sector-specific platforms carry regulatory moats. Healthcare and defense filings in this dataset reflect domain constraint sets — patient safety stockout risk, wartime surge simulation — that generic platforms cannot address without substantial domain-specific adaptation. Early specialisation in these verticals creates defensible positions that later entrants will find difficult to replicate.