HSCode: Global Trade Classification

• HSCode is a comprehensive, hierarchical system standardized by the WCO to classify traded products, guiding customs tariffs and global trade policies.
• It employs a multi-tiered structure—from HS2 to HS6 and beyond—ensuring precise categorization and regulatory compliance across international markets.
• Automated HSCode prediction leverages machine learning, multimodal fusion, and hierarchical rule reasoning to enhance duty assignment accuracy and reduce misclassification.

The Harmonized System Code (HSCode) is the foundational, internationally standardized nomenclature developed by the World Customs Organization (WCO) for classifying traded products. It is a hierarchical, multi-digit system underpinning customs tariffs, trade statistics, and the regulation of cross-border product flows. HSCode assignment is critical for determining duties, enforcing trade regulations, and enabling compliance in global supply chains. The task of automating or benchmarking HSCode prediction has attracted extensive study, with research advancing across machine learning, multimodal modeling, benchmarking methodologies, domain-specific LLM adaptation, anomaly detection, and hierarchical rule reasoning.

1. Structure and Function of HSCode

The HSCode is a product classification schema structured in a multi-tiered hierarchy. The global standard comprises:

• The first two digits (HS2): Section-level headings (broad categories).
• Digits 3 and 4 (HS4): Product chapters, specifying subcategories.
• Digits 5 and 6 (HS6): Harmonized subheadings, recognized worldwide.
• Digits 7–10: National extensions, allowing countries (e.g., the U.S. Harmonized Tariff Schedule, HTS) to add further specificity.

Classification mandates strict rule-following: products are to be classified according to section and chapter notes, with recourse to hierarchical assignment rules when ambiguity is present. These rules, detailed in the WCO’s General Rules for Interpretation, dictate multiple cross-level, multi-hop decision processes based on product description, composition, use, and context (Yang et al., 22 Oct 2025).

2. Automated HSCode Assignment: Methodologies

Recent research has centered on automating HSCode prediction from product descriptions, leveraging advances in machine learning, deep neural architectures, and multimodal fusion.

Hierarchical Ensemble Methods

• A three-stage hierarchical ensemble model mirrors the code structure: a BERT-transformer classifier for the HS2 level, a similar network for HS4, and an unsupervised, distance-based approach for HS6. Each stage narrows the candidate space and captures finer semantic distinctions (Shubham et al., 2022).
• As an example, the BERT-based model infers class membership via softmax probabilities:

$P(y_i = j) = \frac{e^{z_j}}{\sum_k e^{z_k}}$

where $z_j$ is the logit for class $j$ at the current hierarchy level.

• Rule-based Named Entity Recognition (NER) extracts key product entities and relations by leveraging part-of-speech tagging and dependency parsing, feeding critical features into later stages.
• At the most granular level, cosine similarity between transformed product description vectors (sentence embeddings) determines the candidate HS6 code:

$$\text{cosine}(i, o) = \frac{\text{embedding}_i \cdot \text{embedding}_o}{\|\text{embedding}_i\| \|\text{embedding}_o\|}$$

• Knowledge graphs constructed from regulatory manuals (e.g., the WCO) serve as external knowledge overlays for auditing and fallback classification (Shubham et al., 2022).

Multimodal and Deep Learning Approaches

• Multimodal systems combine text and visual information through deep learning to improve prediction, particularly with noisy or ambiguous source data (Amel et al., 2024). Textual data (description, title, category) is embedded with models such as SimCSE, while product images are encoded via networks such as ResNet50, ViT, or CLIP’s image encoder.
• Fusion of modalities is critical:
  • Simple concatenation or low-rank tensor fusion (LMF) are baselines.
  • The MultConcat method fuses projected feature vectors using both concatenation and element-wise multiplication, producing a comprehensive and discriminative vector for classification.

Benchmarking and Evaluating Automated Solutions

• Multiple commercial and research systems (Zonos, Tarifflo, Avalara, WCO BACUDA, etc.) are benchmarked on critical metrics: accuracy at various hierarchy depths, speed, provision of rationale, and alignment with official codes (Judy, 2024).
• Notably, high accuracy and rationale transparency (as seen in Tarifflo) are essential for regulatory compliance, whereas fastest systems often lack transparency or detail.
• State-of-the-art LLM-based models fine-tuned for domain specifics (e.g., Atlas, based on LLaMA-3.3-70B) have been introduced, achieving 40% full 10-digit classification accuracy and demonstrating favorable five- to eight-fold inference cost reductions compared to proprietary models, though the benchmark remains challenging (Yuvraj et al., 22 Sep 2025).

Outlier Detection and Pattern Analysis

• HSCode misuse and evasion (e.g., illicit e-waste trade) are detected using K-Means-based outlier-aware segmentation of price-volume trade data, supplemented with logistic regression assigning a “Waste Score.” Finished goods with anomalous (scrap-like) trading signatures (e.g., electric generators under HS 8502) are flagged for policy review and enforcement (Ramli, 24 Sep 2025).

3. Challenges in HSCode Prediction and Application

Automated HSCode assignment is hindered by several factors:

• Hierarchy Imbalance and Data Sparsity: Deeper HS levels (e.g., HS6 or 10-digit national extensions) often contain highly imbalanced class distributions. The ensemble paradigm mitigates, but not eliminates, granularity loss by grouping rare codes under an “others” category (Shubham et al., 2022).
• Ambiguity and Noise: User input is frequently ambiguous or intentionally misleading, making nuanced NLP critical. Even advanced agents or multimodal systems can underperform against human experts in realistic noisy settings (Yang et al., 22 Oct 2025).
• Rule Complexity and Implicit Logic: HS rules feature ambiguous boundaries and implicit exceptions; agents must perform multi-step, hierarchical, cross-referential reasoning. Current LLM and agent-based systems remain substantially behind human-level accuracy (46.8% vs. 95.0% on realistic benchmarks) (Yang et al., 22 Oct 2025).
• Evolving Standards: The WCO and national authorities periodically revise codes and notes, creating a moving target for static models; “industry-level” hierarchy modeling and semi-supervised updating are proposed as partial solutions (Shubham et al., 2022).
• Anomalous Trade Patterns: Evasion via misclassification, especially with e-waste, is sustained by loopholes at the code definition level; automated pattern recognition frameworks are required to detect and address these cases (Ramli, 24 Sep 2025).

4. Performance Metrics and Benchmarks

Rigorous evaluation frameworks have been developed:

Solution	Full-Code Accuracy	Speed	Rationale
Tarifflo	89.22% (10-digit)	~30s/item	Comprehensive
Avalara	80.00% (10-digit)	Days–Weeks	Detailed
Zonos	44.12% (10-digit)	Near instant	Limited
WCO BACUDA	12.75% (6-digit)	Near instant	None
Atlas (LLaMA-3.3)	40% (10-digit)	Batch	Reasoned path

• Multimodal fusion has demonstrated improvements in top-1 accuracy (increase of 8.2% with image inclusion over text-only), with top-3 and top-5 accuracies of 93.5% and 98.2% respectively (Amel et al., 2024).
• Hierarchical ensemble models have achieved 16% higher accuracy compared to single-stage, flat classifiers, achieving 100% valid assignments (relative to 12% invalid in baselines) (Shubham et al., 2022).
• Human experts, rigorously annotating real-world e-commerce product datasets, are estimated to achieve ~95.0% 10-digit accuracy, far above automated systems on realistic, expert-level benchmarks (Yang et al., 22 Oct 2025).

5. Interpretability, Auditing, and Regulatory Implications

• The use of knowledge graphs for auditing provides interpretability and traceability by mapping extracted product entities and regulatory relationships, supporting both automated checking and post-hoc justification (Shubham et al., 2022).
• Some commercial systems substantiate assignments with detailed rationales, referencing statutes and court rulings, enhancing compliance and trust. Fastest systems lacking rationale are limited in regulatory contexts (Judy, 2024).
• Waste signature detection and the assignment of a Waste Score to HS codes expose regulatory evasion routes, enabling targeted inspection, policy adjustment, and resource prioritization (Ramli, 24 Sep 2025).
• Hierarchical, rule-based decision making is essential for expert-level agent performance but remains a bottleneck; rule-overriding logic, ambiguity management, and multi-hop inference are identified as areas for further development (Yang et al., 22 Oct 2025).

6. Future Directions

• Advances in Hierarchical and Industry-Level Modeling: The construction of hierarchical models at sector-specific levels, as opposed to broad chapter levels, is foreseen to alleviate imbalance and improve fine-grained accuracy (Shubham et al., 2022).
• Continuous Learning and Retrieval-Augmentation: Research is moving toward integrating dynamic, unsupervised or retrieval-augmented modules to maintain alignment with evolving regulatory standards and enhance reasoning performance (Yuvraj et al., 22 Sep 2025).
• External Data and Multimodality: Fusion with real-time shipment, contextual, and regulatory data is expected to further enhance accuracy and robustness (Shubham et al., 2022). Multimodal expansion beyond mere text-image pairing presents new opportunities and challenges.
• Advanced Benchmarking: The establishment of open, standardized benchmarks (e.g., Atlas and HSCodeComp), mirroring those in NLP (SuperGLUE, MMLU), is critical for objective progress, revealing gaps and incentivizing innovation (Judy, 2024, Yuvraj et al., 22 Sep 2025, Yang et al., 22 Oct 2025).
• Enhanced Rule Reasoning and Working Memory Integration: Development of agents capable of structured, stepwise application of tariff rules with persistent external memory is projected to close the remaining gap to human expertise, reducing both “premature closure” and “reasoning hallucinations” (Yang et al., 22 Oct 2025).

7. Impact on Global Trade and Policy

HSCode assignment profoundly impacts the efficiency, compliance, and integrity of global trade. Automated and benchmarked solutions promise streamlined clearance, error reduction, and dynamic adaptation to regulatory changes. However, as current research demonstrates, significant barriers remain before automated agents can fully replicate expert human judgment. The integration of auditing frameworks, anomaly detection (e.g., waste signature analysis), and transparent, standardized evaluation will be essential to the next generation of resilient, compliant, and efficient global trade classification.