Hierarchical Industrial Demand Forecasting with Temporal and Uncertainty Explanations
| Authors | Harshavardhan Kamarthi et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2603.06555 |
| Download | |
| Categories | cs.LG |
Abstract
Hierarchical time-series forecasting is essential for demand prediction across various industries. While machine learning models have obtained significant accuracy and scalability on such forecasting tasks, the interpretability of their predictions, informed by application, is still largely unexplored. To bridge this gap, we introduce a novel interpretability method for large hierarchical probabilistic time-series forecasting, adapting generic interpretability techniques while addressing challenges associated with hierarchical structures and uncertainty. Our approach offers valuable interpretative insights in response to real-world industrial supply chain scenarios, including 1) the significance of various time-series within the hierarchy and external variables at specific time points, 2) the impact of different variables on forecast uncertainty, and 3) explanations for forecast changes in response to modifications in the training dataset. To evaluate the explainability method, we generate semi-synthetic datasets based on real-world scenarios of explaining hierarchical demands for over ten thousand products at a large chemical company. The experiments showed that our explainability method successfully explained state-of-the-art industrial forecasting methods with significantly higher explainability accuracy. Furthermore, we provide multiple real-world case studies that show the efficacy of our approach in identifying important patterns and explanations that help stakeholders better understand the forecasts. Additionally, our method facilitates the identification of key drivers behind forecasted demand, enabling more informed decision-making and strategic planning. Our approach helps build trust and confidence among users, ultimately leading to better adoption and utilization of hierarchical forecasting models in practice.
Engineering Breakdown
Plain English
This paper addresses a critical gap in hierarchical time-series forecasting: while machine learning models have achieved good accuracy on demand prediction tasks across industries, their predictions remain black boxes. The authors introduce a novel interpretability method specifically designed for large-scale hierarchical probabilistic forecasting models that works in real supply chain scenarios. The approach reveals which time-series within a hierarchy matter most and identifies which external variables drive predictions at specific time points—insights that were previously inaccessible. The work bridges interpretability and uncertainty quantification, two areas that typically operate in isolation.
Core Technical Contribution
The core novelty is a domain-adapted interpretability framework that handles the unique challenges of hierarchical forecasting: reconciling predictions across multiple aggregation levels while maintaining uncertainty estimates. Rather than applying generic feature importance methods (like SHAP or attention weights), the authors built methods that respect the hierarchical constraints—where parent-level forecasts must align with aggregated child-level forecasts. The approach also handles probabilistic outputs (not just point estimates), extracting interpretability signals from distributions rather than single predictions. This is the first systematic treatment of interpretability for hierarchical probabilistic forecasting at production scale.
How It Works
The system starts with a hierarchical time-series structure where lower-level series (e.g., SKU-store combinations) aggregate into higher levels (product lines, regions, company-wide). A probabilistic forecasting model produces distributions over future values at all hierarchy levels. The interpretability method then attributes prediction changes to: (1) specific time-series within the hierarchy (which series contributed most to a given forecast?), (2) specific external variables at specific time steps (which promotion, weather, or inventory signal mattered?), and (3) temporal dependencies (how much does history versus recent context drive the forecast?). The approach respects hierarchical reconciliation constraints—ensuring that explanations for parent nodes align logically with explanations for child nodes. Uncertainty is preserved throughout: interpretability scores are computed per sample from the predictive distribution, not just for point estimates.
Production Impact
For supply chain teams, this enables root-cause analysis of forecast errors and anomalies without data scientists acting as intermediaries—a demand planner can now ask 'why did this store's forecast spike?' and get a direct answer about which SKUs drove it and which external signals mattered. This directly addresses the black-box problem that prevents adoption of sophisticated models in risk-averse industries; explainability becomes a first-class requirement, not an afterthought. The method integrates into existing probabilistic forecasting pipelines with modest computational overhead (interpretability computed post-hoc from model outputs). However, trade-offs include: the approach requires the underlying model to be probabilistic (not all demand forecasting systems use uncertainty quantification), and it adds latency if explanations must be computed in real-time at scale. For hierarchies with millions of nodes, computing per-series importance scores may require batching or approximation strategies.
Limitations and When Not to Use This
The method assumes the underlying forecasting model is well-calibrated and produces valid probabilistic outputs—garbage distributions will produce misleading explanations. It does not address why the base model makes systematic errors (e.g., if the model consistently underforecasts seasonal spikes, interpretability can only explain which series drove that underforecast, not fix it). The paper truncates mid-sentence in the abstract, leaving the full scope of contributions unclear; the presented work may be incomplete. Scalability to extremely deep hierarchies (10+ levels) or non-tree hierarchical structures (e.g., multiple cross-cutting dimensions like location × channel × product) is not discussed. The approach also doesn't handle dynamic hierarchies where relationships between series change over time, a real problem in retail when stores open/close or product lines are reorganized.
Research Context
This work extends the interpretability-for-time-series literature (building on temporal SHAP, attention-based attribution) into the hierarchical domain, where prior interpretability research mostly ignored the constraint that child predictions must sum to parents. It builds on recent advances in probabilistic forecasting (DeepAR, transformer-based models) and hierarchical reconciliation methods (ERM, CoRec), combining insights from both areas. The paper implicitly benchmarks against supply chain forecasting datasets, though specific accuracy comparisons aren't detailed in the abstract. This opens a research direction: how do we ensure interpretability explanations remain faithful when models are fine-tuned or retrained, and how do we build interactive explanation systems where users can drill into hierarchies dynamically?
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::