Zero-Shot Inventory Forecasting with Multi-Agent Orchestration and LLMs

The Challenge of Inventory Forecasting

Traditional inventory forecasting methods, whether classical time-series models (ARIMA) or advanced deep learning approaches (DeepAR), present significant operational burdens. They require extensive per-SKU model training, hyperparameter tuning, feature engineering, and continuous maintenance. This complexity scales linearly with catalog size, diverting engineering resources from quality improvements to infrastructure management. Furthermore, converting forecasts into purchase orders typically relies on manual business rules, leading to inconsistencies and auditability issues.

Solution Overview: Multi-Agent LLM Architecture

The proposed architecture combines two core capabilities: Amazon Chronos2 for zero-shot probabilistic demand forecasting and a multi-agent orchestration system built with the Strands Agents SDK and Amazon Bedrock AgentCore. This system addresses the limitations of traditional methods by automating the entire pipeline from forecast to purchase order.

End-to-End System Layers

• Data Layer: Amazon S3 serves as the single source of truth for historical sales, future covariates (e.g., promotions), and business rules (e.g., safety stock). This design makes new product onboarding a data concern rather than a code change.
• Inference Layer: Amazon SageMaker Serverless Inference hosts the Chronos2 endpoint for zero-shot time-series forecasting. This is the only external model inference in the pipeline.
• Orchestration Layer: Four LLM agents (Supervisor, Preprocessing, Forecasting, Reporting) built with Strands Agents SDK and deployed on Amazon Bedrock AgentCore. These agents use Claude on Bedrock for reasoning and invoke deterministic tools for all computational tasks.

Benefits of Chronos2 and Multi-Agent Design

Chronos2 excels in this context due to its zero-shot generalization, meaning new SKUs require no model training, and its robust covariate support for both past and future features (like promotions). This enables powerful what-if scenario analysis by comparing forecasts under different covariate conditions. The multi-agent approach, in contrast to a monolithic LLM, offers modularity, testability at the component level, and graceful error handling, making the system scalable and auditable.

date,sales,promotion,day_of_week,is_weekend,price
2024-01-01,120,0,1,0,29.99
2024-01-02,95,0,2,0,29.99
...
2024-01-20,140,0,6,1,29.99
2024-01-21,,1,7,1,24.99   <-- Forecast horizon begins (sales is null)
2024-01-22,,1,1,0,24.99
2024-01-23,,0,2,0,29.99