Kexintong is a technology-credit research framework that brings patent evidence and financial resilience into the same assessment process. It is designed to help lenders examine not only what a technology company owns today, but also the quality of its innovation, the durability of its operations, and the way those signals should affect credit terms and monitoring.
The complete study covers a broader private company universe. This repository publishes a five-company reproducibility sample so that the core data flow and deterministic analysis can be inspected without distributing the full research dataset. The public sample demonstrates the method; it is not the definition or full scope of Kexintong.
Conventional credit analysis is strongest when value is visible in collateral, stable cash flows, and long operating histories. Technology-oriented companies often look different: important assets sit in patent portfolios, R&D precedes revenue, and short-term statements may understate long-term technical value.
Kexintong addresses that gap through two connected analytical views:
- Patent intelligence measures the scale, quality, technological relevance, growth, and semantic content of a company’s patent portfolio.
- Financial resilience measures cost structure, break-even capacity, sensitivity to operating shocks, and early-warning conditions.
Together, these views support borrower segmentation, differentiated pricing, and post-loan monitoring.
flowchart LR
P[Patent records] --> PS[Patent intelligence score]
F[Financial statements] --> BEP[Analytical BEP and resilience]
PS --> A[Technology credit assessment]
BEP --> A
A --> S[Borrower segmentation]
A --> C[Credit and insurance pricing]
A --> W[Early-warning monitoring]
A -. broader private research sample .-> M[GBM evaluation]
The public workflow follows the solid arrows. Machine-learning evaluation belongs to the broader research workflow and is not retrained on five observations.
- A six-factor patent score. Patent stock, patent-type quality, IPC heat, recent patent growth, NLP similarity, and frontier-keyword density are combined under one versioned weighting scheme.
- Technology-aware text analysis. Company patent text is compared with a frozen high-value patent corpus using TF-IDF and cosine similarity.
- An operating-resilience measure. Financial statements are converted to a consistent quarterly basis before cost-structure and break-even analysis.
- Decision-oriented outputs. Technology and financial signals are translated into borrower quadrants, pricing adjustments, and warning levels rather than stopping at a descriptive score.
- An auditable public interface. CSV schemas, provenance notes, sample-selection metadata, and quality checks are included alongside the code.
The underlying study was developed on a private sample of 50 A-share companies across 16 technology-related industries. That environment supports the complete patent, financial, and Gradient Boosting evaluation workflow.
This repository uses five representative listed companies to reproduce the deterministic portion of the framework:
- schema-documented data ingestion and financial-basis conversion;
- six-factor patent valuation;
- analytical break-even-point analysis;
- operating and raw-material sensitivity analysis;
- borrower portfolio classification;
- risk-based lending and insurance recommendations;
- time-series early-warning views.
Public scores are normalized within the five-company sample. They should therefore be read as workflow outputs, not as full-market ranks or replicas of the private 50-company results.
| Public asset | Rows | Role in the workflow |
|---|---|---|
| Company master | 5 | Company identity and technology-field mapping |
| Quarterly financial observations | 95 | Revenue, cost, expenses, R&D, and operating cash flow |
| Company patent records | 3,911 | Portfolio, legal-status, IPC, type, and text analysis |
| High-value patent reference | 198 | Frozen NLP comparison corpus |
| IPC market reference | 1,593 | Technology-track heat and concentration reference |
| Industry BEP reference | 16 | Industry-level operating benchmarks |
| Raw-material series | 5 years | Scenario and sensitivity reference |
The public package also includes JSON schemas, a quality report, selection metadata, and source-provenance documentation. The private full-sample workbooks are not part of this repository.
Detailed field definitions are available in docs/data_dictionary.md. Data-source and transformation notes are recorded in metadata/source_provenance.md.
| Component | Weight |
|---|---|
| Patent stock | 20% |
| Patent-type quality | 12% |
| IPC technology heat | 20% |
| Recent patent CAGR | 13% |
| NLP semantic similarity | 25% |
| Frontier-keyword density | 10% |
The weighted score is adjusted by a technology-industry multiplier and normalized to a 0–100 scale. The formal configuration is maintained in one scoring configuration source so the report, code, and generated workbook use the same weights.
Financial flow data is interpreted from the explicit value_basis field:
- cumulative year-to-date reports are differenced within each company and fiscal year;
- reported single-quarter values are retained;
- sequence-based inference is only a compatibility fallback when the metadata cannot be used.
The analytical safety margin is defined as:
BEP revenue = fixed costs / gross margin
BEP safety margin = observed revenue / BEP revenue
Cost behavior is estimated from the company’s quarterly history. The resulting resilience measure feeds the sensitivity, segmentation, pricing, and warning modules.
The complete study also evaluates a GradientBoostingRegressor using patent-derived and financial-structure features. The report gives the following five-fold cross-validation result for 49 usable modeling observations from the broader private research sample:
R² = 0.667 ± 0.113
This is a result of the broader private research sample. It is documented here for research context and is not generated by the five-company public run.
See docs/methodology.md for the operational distinction between Demo and Research modes.
The reproducibility workflow is organized around four output groups:
- Reports: patent scoring, BEP analysis, and credit recommendations;
- Figures: cost structure, patent/financial comparisons, sensitivity, and representative-company profiles;
- Interactive views: borrower portfolio positioning and BEP time-series monitoring;
- Website payload: a derived-only company assessment JSON for the separate public interface.
Curated reproducibility outputs belong in outputs/examples/. A local run writes fresh results to outputs/latest/, which is excluded from version control.
kexintong/
├── src/
│ ├── patent_scoring/ # Patent valuation and NLP module
│ └── bep_analysis/ # BEP, sensitivity, segmentation and pricing
├── scripts/ # Demo preparation and explicit run modes
├── data/
│ ├── demo/ # Five-company public sample
│ ├── reference/ # Frozen analytical references
│ ├── schema/ # Public data contracts
│ └── generated/ # Local compatibility inputs
├── metadata/ # Selection, quality and provenance records
├── docs/ # Methodology and data dictionary
├── tests/ # Repository and data-contract regression checks
├── outputs/
│ ├── examples/ # Curated public-workflow examples
│ └── latest/ # Local run output; not committed
└── report/
└── Kexintong_Report.pdf
python -m pip install -r requirements.txt
python scripts/run_demo.pyThe runner prepares internal compatibility files under data/generated/, executes the integrated patent and financial analysis, and writes current results to outputs/latest/. It also creates outputs/latest/public/company_assessments.json, a derived-only payload for the separate public website. The bundled financial series ends at 2025 Q3, while the patent records extend into March 2026; the export identifies these cutoff dates separately.
The private research entry point is separate:
python scripts/run_research.py --data-dir PATH_TO_PRIVATE_PREPARED_DATAThe patent valuation and NLP implementation lives under src/patent_scoring/. The public runner executes patent scoring first, followed by BEP, sensitivity, borrower-segmentation, pricing, early-warning analysis, and derived website-payload export on the bundled five-company sample.
- Five-company normalization does not reproduce the scale or ranking of the full research universe.
- The public package cannot reproduce private full-sample model training.
- TF-IDF provides an interpretable semantic baseline but does not capture the full contextual capacity of modern embedding models.
- The model result is proof-of-concept evidence from a limited research sample, not production validation.
- Outputs support research and credit analysis; they are not automated approval decisions or investment recommendations.
The research report documents the broader financing mechanism, quantitative validation, methodology, policy discussion, and appendices.
Read the full Kexintong report
Developed by Team Oyster Hour.
- Xiaoyu (Sophy) Xia — project framework, data sourcing, financial-resilience analysis, BEP modeling, and credit-decision design.
- Xiaolei Zhang — patent intelligence, NLP scoring, code integration, and interactive analytical outputs.