Demand Forecasting: LLMs vs Classical Time Series – Hybrid Wins for 2026 Supply Chains

Classical Time Series Forecasting: Foundations and Limitations Classical time series forecasting has long been the backbone of supply chain planning, relying on statistical models like ARIMA, Exponential Smoothing (ETS), and Prophet. These methods excel at capturing trends, seasonality, and autocorrelation in structured historical sales data, making them reliable for stable demand patterns in logistics. For B2B leaders, classical approaches integrate seamlessly with ERP systems like SAP IBP or Blue Yonder, offering interpretable outputs and low computational needs. A typical workflow involves decomposing data into trend, seasonal, and residual components, then fitting models via auto-arima or similar tools. However, limitations emerge in volatile environments: Exogenous shocks : Pandemics, strikes, or promotions disrupt patterns, leading to high forecast errors (e.g., MAPE 20% during COV

ID). Unstructured data blindness : Ignores news, social media, or weather events that drive sudden demand spikes. Stationarity assumptions : Fails on non-stationary series with structural breaks. In logistics, where SKUs number in the millions, these gaps amplify inventory costs—overstock by 10-30% or stockouts by 15%. How LLMs Enhance Demand Forecasting with Unstructured Data Large Language Models (LLMs) revolutionize "demand forecasting LLMs vs classical" debates by processing multimodal inputs: text, images, and time series. Unlike classical models, LLMs like GPT-4 or Llama variants embed unstructured data—supplier emails, market reports, event calendars—into vector spaces for contextual forecasting. Key mechanism: Prompt engineering + fine-tuning . Feed LLMs with time series patches converted to text prototypes (e.g., "Sales: [100,120,90] during rainy season"), then query for predict

ions. This enables zero-shot forecasting, where models generalize from pre-trained knowledge without retraining. In AI supply chain forecasting, LLMs shine by: Incorporating real-time events (e.g., "Black Friday promo + port strike"). Handling multimodality: ERP logs + product images for visual demand cues. Scaling to high-dimensional data via attention mechanisms. This shifts logistics from reactive to proactive planning, reducing bullwhip effects in volatile chains. Key Research: EventCast, LLMForecaster, and Multimodal Frameworks Recent arXiv papers provide evidence for LLM time series forecasting: EventCast (arXiv:2403.06195, March 2024) : A hybrid framework injecting LLM-extracted event knowledge into e-commerce forecasts. On real datasets, it cuts MAE by 86.9% and MSE by 97.7% vs. event-agnostic models. . LLMForecaster / LLM4TSF (arXiv:2402.02795, Feb 2024) : Benchmarks LLMs across

8 billion observations, showing 10-20% gains in cross-domain tasks. Pre-alignment (adapting LLMs before forecasting) outperforms post-alignment, but compute costs rise 5-10x. . Time-LLM (arXiv:2310.01728, Oct 2023) : Reprograms LLMs by aligning patches with text prototypes, beating specialized models in few-shot settings (e.g., 15% MAPE reduction on ETTh datasets). Excels in zero-shot for unseen series. . Multimodal ERP study (Scientia Research, 2024): LLM embeddings from ERP text/images yield 35%+ MAPE drops vs. classical or ERP-only baselines. These validate LLMs for "hybrid demand forecasting" in logistics. Performance Comparison: Accuracy, Volatility, and Zero-Shot Gains Benchmarks reveal nuanced "demand forecasting LLMs vs classical" results: Scenario Classical (e.g., Prophet) LLM-Augmented Improvement :---------------- :------------------------ :------------ :-------------- Stable

demand Low error (MAPE 5-10%) Comparable Minimal Volatile (events) MAPE 20-40% MAPE 8-15% 35-97% (EventCast) Zero-shot Poor generalization Strong (Time-LLM) 20%+ LLMs outperform in high-volatility (e.g., holidays, disruptions) by 35-97% error reduction, per EventCast. In low-volatility, classical edges out due to efficiency. Zero-shot gains aid new SKUs, critical for e-commerce logistics. Caveat: Raw LLMs sometimes match ablations without LLM components, per recent critiques, emphasizing smart integration over direct use. Hybrid Approaches and Multi-Agent Orchestration Pure LLM or classical falls short; hybrids rule. LLM-classical hybrids use LLMs for feature engineering (e.g., event embeddings) fed into XGBoost or Prophet. Enter multi-agent platforms like LUMOS: An open-source orchestration layer where agents specialize—ClassicalAgent for trends, EventAgent (LLM-powered) for shocks, En

sembleAgent for fusion. LUMOS routes queries dynamically: stable series to classical, volatile to LLM. Benefits for supply chain AI agents: Modularity : Swap models (e.g., Time-LLM agent). Scalability : Parallel inference on GPU clusters. Auditability : Trace predictions to agents. In logistics, LUM