Hierarchical Industrial Demand Forecasting with Temporal and Uncertainty Explanations

This paper introduces a novel interpretability method for large hierarchical probabilistic time-series forecasting that addresses structural and uncertainty challenges, successfully explaining state-of-the-art industrial models to enhance stakeholder trust and decision-making through real-world case studies and semi-synthetic evaluations.

The Gist

Imagine you are the captain of a massive, futuristic cargo ship. You have a computer system that predicts exactly how much fuel, food, and spare parts you'll need for the next month. This system is incredibly smart, but it's also a "Black Box." It gives you a number, but it won't tell you why it chose that number.

If the computer says, "You need 500 tons of fuel," you might ask:

• "Is it because the weather is getting stormy?"
• "Is it because we're visiting a new port?"
• "Or is it just guessing?"

In the real world, big companies (like chemical manufacturers) face this exact problem. They use complex AI to predict demand for thousands of products, organized in a giant family tree (a hierarchy). For example: Chemical Company → Region → Factory → Specific Product.

The problem is that the AI is so complex and deals with so much uncertainty (probabilities) that no one can explain its logic. This paper introduces a new tool called HIEREINTERPRET to open that black box.

Here is the paper explained in simple terms, using some creative analogies.

---

1. The Problem: The "Too Big to Understand" Tree

Imagine the company's data is a giant, multi-story family tree.

• The Hierarchy Problem: If you want to know why the "Grandparent" node (the whole company) needs more product, you can't just look at every single "Grandchild" node (individual products) at once. That's like trying to hear every conversation in a stadium of 10,000 people all at once. It's too noisy and too slow.
• The Probability Problem: The AI doesn't just say "We need 100 units." It says, "We probably need 100, but it could be anywhere between 80 and 120." Most explanation tools only work on single, definite numbers, not on these "ranges of possibilities."

2. The Solution: Two Magic Tricks

The authors created a method to make the AI explain itself using two clever tricks.

Trick #1: The "Subtree" Shortcut (Simplifying the Family Tree)

Instead of trying to connect every single person in the family tree to everyone else (which is chaotic), the new method says: "Let's just look at the immediate family."

• The Analogy: Imagine you want to know why your Great-Grandfather is happy. Instead of asking every single cousin, aunt, and neighbor in the world, you just ask your Dad, who asks his Dad, who asks his Dad.
• How it works: The method breaks the giant tree into small, manageable "sub-trees." It calculates the importance of a variable by passing the "importance score" down the chain, step-by-step, from parent to child.
• The Result: This makes the math much faster and much clearer. It stops the "noise" of the whole stadium and lets you hear the specific conversation that matters.

Trick #2: The "Quantile Translator" (Translating the Uncertainty)

The AI speaks in "probabilities" (a foggy cloud of possibilities), but the explanation tools only understand "definite facts" (clear, solid ground).

• The Analogy: Imagine the AI is a weather forecaster saying, "There is a 90% chance of rain, but maybe 70%." The explanation tool is a person who only understands "Yes, it rained" or "No, it didn't."
• How it works: The authors invented a translator. They take the "foggy" probability cloud and slice it into specific layers (like 70%, 90%, and 95% confidence levels). They treat these slices as if they were definite facts.
• The Result: Now, the explanation tool can look at the "90% slice" and say, "Ah, I see! The reason the AI is worried about rain is because of the wind speed." It turns the fog into a clear picture.

3. The Proof: Did it Work?

The authors tested this on a massive dataset from The Dow Chemical Company (which tracks over 10,000 products) and other public datasets.

• The Test: They created "fake" scenarios where they knew exactly what should be important (like a fake storm or a fake price hike).
• The Result: The new method was significantly better at finding the right answers than old methods.
  • For definite predictions, it was 62% more accurate.
  • For uncertain (probabilistic) predictions, it was 26% more accurate.
  • It was also much faster, cutting down the time needed to explain the AI from over 100 minutes to just 2 minutes in some cases.

4. Real-World Stories (Case Studies)

The paper shows three cool examples of how this helps real people:

1. The Pandemic Shift: The AI noticed a sudden jump in demand for home-furnishing items. The explanation tool showed the AI was looking at the "upward trend" starting in late 2019. It confirmed the AI knew the pandemic was changing people's habits.
2. The Economic Indicator: The AI predicted a drop in packaging demand. The tool revealed it was reacting to a specific economic number (the Consumer Price Index) dropping. This helped business leaders understand why sales might dip.
3. The Lost Customer: A major customer stopped buying from the company. The AI became very "uncertain" about future sales. The explanation tool showed the AI was confused because it was trying to balance old data (with big peaks) and new data (flat lines). This told the business, "Don't trust the prediction yet; the situation is too unstable."

The Big Takeaway

This paper gives us a flashlight for the dark room of industrial AI.

Before, companies had to trust the AI blindly. Now, with HIEREINTERPRET, they can ask, "Why did you make that prediction?" and get a clear, logical answer. This builds trust, helps managers make better decisions, and ensures that when the AI says "Order more fuel," they know exactly why.

Technical Summary

1. Problem Statement

The paper addresses the critical lack of interpretability in state-of-the-art Hierarchical Time-Series Forecasting (HTSF) models, particularly within industrial demand forecasting contexts (e.g., supply chain management for chemical companies).

• The Gap: While modern HTSF models (often deep learning-based) achieve high accuracy and scalability by modeling complex temporal patterns and hierarchical constraints, they operate as "black boxes."
• Specific Challenges:
  1. Hierarchical Complexity: Industrial data involves thousands of time-series with strict hierarchical relationships (e.g., product $\to$ category $\to$ region $\to$ total). Existing interpretability methods (like LIME or Integrated Gradients) fail to handle these high-dimensional, interdependent structures efficiently, often producing noisy or suboptimal results.
  2. Probabilistic Outputs: State-of-the-art HTSF models output probability distributions (uncertainty) rather than deterministic point estimates. Most existing explainability tools are designed for deterministic outputs and cannot directly explain distributional forecasts.
  3. Evaluation Difficulty: There are no public datasets with "ground truth" explanations, making it difficult to validate if an interpretability method is actually correct.
• Key Research Questions (RQs):
  • RQ1: Which input variables (internal or external) contribute most to the prediction?
  • RQ2: Which specific time steps in the history are most critical?
  • RQ3: Why do predictions change when input data or distributions shift?

2. Methodology: HIEREINTERPRET

The authors propose HIEREINTERPRET, a framework that adapts general interpretability techniques for hierarchical and probabilistic tasks through two novel modifications:

A. Subtree Approximation (Addressing Hierarchy)

Instead of treating the entire hierarchy as a flat multivariate problem (which is computationally expensive and ignores structural constraints), the authors decompose the problem:

• Concept: They assume that the importance of a distant node in the hierarchy can be approximated by the chain of importance scores between adjacent nodes.
• Mathematical Formulation: For a path $x_i \to x_{v1} \to \dots \to x_j$, the global importance $I(x_i, x_j)$ is approximated as the product of adjacent importance scores:
  $$I(x_i, x_j) \approx I(x_i, x_{v1}) \times I(x_{v1}, x_{v2}) \times \dots \times I(x_{vm}, x_j)$$
• Algorithm: A Breadth-First Search (BFS) traversal is used to propagate importance scores from a target node up and down the tree. This reduces computational complexity from $O(N^2)$ (all-pairs) to $O(N \cdot T_{max})$, where $T_{max}$ is the max number of children, while preserving hierarchical coherency.

B. Non-Parameterized Translation (Addressing Probabilistic Outputs)

To explain probabilistic models that output distributions $\hat{P}$ rather than single values:

• Concept: The method transforms the probabilistic model into an equivalent deterministic model by extracting quantiles from the output distribution.
• Implementation: Instead of explaining the mean or the full distribution, the method treats specific quantiles (e.g., 70th, 90th, 95th) as deterministic targets.
• Process:
  1. Sample from the forecast distribution $\hat{P}$.
  2. Calculate specific quantiles $q(\tilde{X})$.
  3. Apply standard interpretability methods (like Integrated Gradients) to these quantile values.
  4. This allows the system to explain not just the "expected" value but also the drivers behind uncertainty and tail risks.

C. Synthetic Benchmark Construction

To evaluate these methods, the authors created a benchmark with ground truth explanations:

1. Generation: Created univariate synthetic series with known anomalies (e.g., sine waves, variance shifts).
2. Hierarchy: Embedded these anomalies into hierarchical structures with defined dependencies (same-series, cross-series, cross-level, and external variable dependencies).
3. Integration: Merged these synthetic patterns into real-world datasets (Dow Chemical, M5, Tourism-L, Wiki) to test robustness against real-world noise while retaining known ground truths.

3. Key Contributions

1. Generalized Solution: First framework to adapt a wide range of explainability methods (LIME, IG, FO, SG) for both hierarchical and probabilistic forecasting tasks.
2. Novel Techniques:
   • Subtree Approximation: Efficiently handles hierarchical constraints without flattening the data.
   • Quantile-Based Deterministic Alternative: Enables explanation of arbitrary probability distributions without needing to know the underlying parametric form.
3. New Benchmark: Established a semi-synthetic benchmark with ground-truth explanations derived from real-world industrial data distributions.
4. Real-World Validation: Validated the approach on proprietary data from The Dow Chemical Company (10,000+ products) and public datasets.

4. Experimental Results

The method was evaluated on four datasets (Dow, M5, Tourism-L, Wiki) across four HTSF models (HAILS, HierE2E, SHARQ, DeepAR).

• Deterministic Forecasting (Point Estimates):

  • Cross-Variable Importance: Subtree approximation improved External Variable Detection Accuracy (EVDA) by 40.4% and Importance Accuracy Score (IAS) by 62.0% on the Dow dataset compared to baselines without subtree approximation.
  • Same-Variable Importance: Showed consistent improvements (e.g., +12.3% on Dow, +8.0% on M5).
  • Best Performer: Gradient-based methods (Integrated Gradients) combined with Subtree Approximation yielded the best results.

• Probabilistic Forecasting (Uncertainty):

  • The method successfully adapted to quantile-based explanations.
  • Subtree approximation improved performance in roughly 80% of probabilistic cases.
  • The 90th quantile showed the most significant relative improvement, indicating the method effectively explains tail risks and uncertainty drivers.

• Scalability & Efficiency:

  • Time Complexity: The proposed method reduced runtime significantly. For the full Dow hierarchy, explanation time dropped from 6,114 seconds (naive approach) to 123.5 seconds (Subtree Approximation).
  • Scalability: Performance degradation was minimal as hierarchy depth increased, unlike baseline methods which suffered significant accuracy drops.

5. Significance and Case Studies

The paper demonstrates that explainability is not just an academic exercise but a practical necessity for industrial adoption.

• Case Study 1 (Trend Detection): The model correctly identified that a shift in housing product demand (post-2019 pandemic) was driven by historical upward trends in the data, aligning with domain expert knowledge.
• Case Study 2 (External Drivers): The system identified that a sudden drop in packaging demand was driven by a specific economic indicator (Consumer Price Index), validating the model's ability to link external variables to internal forecasts.
• Case Study 3 (Uncertainty Explanation): When a major customer relationship ended, causing periodic peaks to disappear, the model correctly attributed the resulting high forecast uncertainty to the conflict between old (periodic) and new (non-periodic) data patterns.

Conclusion:
This work bridges the gap between high-performance industrial forecasting and the need for transparency. By providing a scalable, accurate, and probabilistic-aware interpretability framework, it enables stakeholders to trust, debug, and strategically utilize hierarchical forecasting models in complex supply chain environments.