credit-risk-cot

Target Role

Role: Credit Risk Analyst / Credit Reviewer
Use Scenarios: Pre-loan due diligence, annual post-loan review, risk report quality inspection, credit approval defense — need to submit a logically complete risk analysis to the credit review committee
Output Purpose: Generate a structured reasoning chain as the data support and logical evidence for the risk conclusion
Decision Level: Information assistance, providing a logical basis for credit decisions, and must not be directly used as an approval conclusion
Execution Frequency: Executed on demand, once per credit application or risk event trigger

Data Sources

Required Data

Data Item	Source	Acquisition Method	Sensitivity Level
Raw credit data (financial statements/credit reports)	User upload	File upload	Internal
Target risk point	User-specified	Conversation input	Internal
Industry background information	User-provided / references/	Conversation input or file reading	Public
Industry benchmark data	references/industry-benchmarks.md	File reading	Public

Data Masking Rules

Unified Social Credit Code of the enterprise: show the first 6 and last 4, with * in between
Bank account number: show only the last 4 digits
Personal ID number: show the first 3 and last 4, with * in between
Detailed address of collateral: not fully displayed in the output, only the region and type are shown

Degradation Strategy

If the raw credit data is empty or missing:The analysis must be terminated immediately and no content output(core constraint)
If industry background information is missing: mark "industry benchmarking dimension not included" and perform individual analysis only
If the financial statements cover only 1 year: mark "insufficient data, trend analysis unavailable" and perform static analysis only
If the statement basis is unclear (unable to distinguish consolidated/parent company): mark "statement basis questionable, analysis based on the consolidated statement assumption"

Terminology

Easily Confused Terms	Meaning in This Skill
Consolidated statements vs. parent company basis	Consolidated statement data may only be compared with consolidated statement data, and parent company basis may only be compared with parent company basis; mixing them is strictly prohibited
DSCR (Debt Service Coverage Ratio)	Operating cash flow / current principal and interest due, not EBITDA/interest
Net cash ratio	Net operating cash flow / net profit, measuring profit quality
Construction in progress	Capital expenditures not yet completed and transferred to fixed assets; neither inventory nor fixed assets
Capacity utilization rate	Actual output / designed capacity; a key indicator for industry benchmarking
Aging	The time span from the date accounts receivable are recognized to the analysis point in time; not the number of overdue days

Execution Workflow

Read before write: Before beginning any reasoning, the following data confirmation steps must be performed:

Read and list all input data (financial statements, credit reports, target risk points)

Confirm the data time range, accounting standards (CAS/IFRS), and reporting scope (consolidated/parent company)

Run scripts/validate_financial_data.py to verify data integrity and cross-footing relationships

Proceed to Step 1 only after verification passes

Step 1: Signal Identification

Precisely cite from the raw credit data the preliminary anomaly signals related to the target risk point, and check item by item across the following dimensions:

Assets and liabilities: construction in progress/inventory/other receivables showing large year-over-year increases for multiple consecutive periods
Profitability: gross margin/net margin showing a continuous declining trend
Cash flow: operating cash flow continuously diverging from net profit (net cash ratio < 0.5)
Scale matching: inventory growth significantly higher than operating revenue growth; accounts receivable growth exceeding revenue growth by 50pct+

Anti-laziness instruction：

Do not skip any dimension, even if some dimensions "appear normal"
All cited data must be exactly consistent with the original text (including units and decimal places)
Must show the "intermediate stepping stones"; cannot directly say "A causes C"
When evidence is insufficient, acknowledge the insufficiency and do not fabricate data

Consult references/cot-signal-library.md to obtain the complete signal type library.

✅ Valid signal identified → proceed to Step 2
❌ No verifiable signal in the raw data → terminate the analysis of this risk point and output nothing
⚠️ Insufficient data for some signals → mark "(insufficient data, pending verification)" and proceed to Step 2

📋 Data source:user_upload(user-uploaded financial statements/credit reports)

Step 2: Industry Trend Verification

Obtain public data for the sub-industry, verify whether the enterprise's behavior is consistent with the overall industry trend, and distinguish "individual problems" from "industry systemic risk".

Consult references/industry-benchmarks.md to obtain industry benchmark data.

Key checkpoints：

Whether the growth rate of newly added capacity in the entire industry far exceeds the growth rate of downstream end demand
Capacity utilization rate, price trends, and overall industry profitability
If signs of oversupply have already appeared in the industry, the risks faced by the enterprise will be significantly amplified
✅ Industry data obtained successfully → proceed to Step 3
⚠️ Industry data unavailable → mark "industry benchmarking dimension not included" and proceed to Step 3

📋 Data source:reference（references/industry-benchmarks.md）

Step 3: Policy Driver Analysis

Study relevant national and local industrial policies and determine the current direction of policy impact on the industry (supportive/neutral/tightening).

Consult references/policy-regulatory-framework.md to obtain the list of regulatory documents.

Key judgment points：

Whether there were previously encouraging subsidies, low-threshold market entry, or excessively loose approvals
Whether the current policy direction has turned toward tightening, and the quantitative impact on the enterprise's profitability center
✅ Policy analysis complete → Proceed to Step 4
⚠️ No explicit policy document available for citation → Mark "(insufficient policy information)", proceed to Step 4

Step 4: Technology Substitution Possibility Assessment

Assess the life cycle position of the target technology route and evaluate the substitution threat of new technologies to the current core business.

Assessment Dimensions：

Whether new technology solutions pose a substitution threat to the current mainstream technology
Conversion efficiency, production cost, and industrialization maturity of the new technology
Estimated substitution time window and confidence level explanation
✅ Assessment complete → Proceed to Step 5
⚠️ Technical information insufficient to determine substitution trend → Mark "(inferred)", proceed to Step 5

Step 5: Financial Consequence Sensitivity Calculation

Establish a simplified financial model, conduct stress testing on the worst-case scenario, and quantify the path of risk transmission to repayment capacity.

Consult references/financial-stress-test-params.md to obtain stress test parameters.

Calculation Elements：

Assume the industry average capacity utilization rate remains below the break-even point for the next 3 years
Estimate additional depreciation and amortization expenses, and recognize corresponding asset impairment provision amounts
Calculate the combined erosion effect on the enterprise's EBITDA / net profit / operating cash flow
Compare the DSCR under the stress scenario with the credit facility covenant constraints

Output Format: Present in a chain expression of "Revenue decline X% → Gross margin change Y pct → Net profit decrease Z% → DSCR = W".

✅ Calculation complete → Proceed to Step 6
❌ Financial data insufficient to build a model → Mark "insufficient data, unable to perform stress test", proceed to Step 6

Step 6: Comprehensive Judgment and Credit Recommendation

Combine the analysis results of the above five steps to form a closed-loop causal chain judgment, align with the target risk points, and provide credit decision reference.

Judgment Logic：

If the enterprise is at the forefront of technology iteration and possesses strong R&D capabilities → risk may transform into opportunity, credit may be moderately maintained
If the enterprise is engaged in follow-the-trend capacity expansion and lacks a core technology moat → treat as a high-risk client, recommend tightening credit

Credit Decision Reference Output：

Prudently assess long-term solvency and provide specific thresholds (e.g., "recommend reducing the limit when DSCR < 1.2")
Consider reducing the credit limit or shortening the credit term
Require supplementary collateral or add financial covenant clauses (covenant)
Or place the enterprise on a key watch list and set post-loan monitoring trigger indicators

📋 Executing Entity:ai→human(AI-generated reasoning chain → confirmed by credit review personnel) 📋 Confirmation Mechanism:approve(the reasoning chain conclusions must be manually reviewed and confirmed)

Output Format (Output Format)

Logical Thinking Chain Derivation

### 逻辑思维链推导

#### 推理链

**【数据源1】** 只可引用原文数据，不可自行杜撰
- 依据：[完整复制原文内容]
- 来源：[标注数据来自哪个报表/哪个年份]

**【推导】** 基于数据源的逻辑推演
- 现象：数据表明...
- 推导：这一变化意味着...
- 影响：进而导致...

**【结论】**
综上，推导出...

**【基于因果链的逻辑推导步骤】**
1. 信号识别：[结论]
2. 行业趋势验证：[结论]
3. 政策驱动力分析：[结论]
4. 技术替代可能性评估：[结论]
5. 财务后果敏感性测算：[链式表达]
6. 综合判断与授信建议：[具体阈值]

### 最终综合结论
综合上述所有推理链，【原始信贷数据】充分支撑了【目标风险点】，核心逻辑为：...
（若证据不足，则仅输出："证据不足，无法构建有效推理链"）

Output Field Specifications

Field	Type	Description
Data source citation	string	Must indicate the source (report type + year + account item)
Derivation process	array	[phenomenon, derivation, impact] three-part format
Six-step conclusions	array	[Step 1-6 conclusions]
Uncertainty annotation	enum	Must mark "(inferred)" or "(insufficient data, to be verified)"
Quantitative threshold	string	Credit recommendations must include specific numerical values (e.g., DSCR < 1.2)

📋 This output can be parsed and used by the credit-risk-classification and post-loan-management Skills.

Compliance Red Lines (Constraints)

Terminate when no data: If the raw credit data is empty or missing, or the reasoning process shows "cannot verify" or "insufficient data", the analysis of that risk point must be terminated immediately, without outputting any content (including the title), and skipped directly.
Fabricating data prohibited: Only numbers explicitly appearing in the raw credit data may be used; fabricating any numbers is prohibited. Cited numbers must be exactly consistent with the original text (including units and decimal places).
Return promises prohibited: Under no circumstances may definitive expressions such as "expected recovery", "expected to improve", or "recovery rate estimated at X%" appear.
Data guessing prohibited: Missing data must be handled according to the degradation strategy; industry averages must strictly not be used to replace real data (industry averages are only used for benchmarking comparisons).
Consistency of reporting basis: Consolidated statement data may only be compared with consolidated statement data, and parent-company-basis data may only be compared with parent-company-basis data; mixing them is strictly prohibited.
Data timeliness annotation: If cited industry benchmark data exceeds its marked validity period, "⚠️ Industry benchmark data may be outdated" must be annotated in the output.
Full coverage of six steps: Each CoT output must go through all 6 steps of the derivation framework; if any step is missing, the reason for skipping must be explained in the report.
Balanced positive and negative evidence: Each risk point must simultaneously present evidence supporting the risk and exceptional circumstances that may weaken the risk, avoiding one-sided reinforcement.
Explicit annotation of uncertainty: Conclusions based on inference (not original-text data) must be annotated with "(inferred)", and data-missing sections must be annotated with "(insufficient data, pending verification)".
Unauthorized approval prohibited: This Skill only generates reasoning chains and credit recommendations, and must not replace manual approval or automatically execute credit adjustments.

Audit Trail

After each reasoning chain generation is completed, generate an audit log audit/{企业简称}_{日期}_cot_audit.json：

{
  "skill_name": "credit-risk-cot",
  "skill_version": "1.0.0",
  "execution_time": "YYYY-MM-DDTHH:mm:ss+08:00",
  "customer_id": "[脱敏]",
  "target_risk_point": "[用户指定的风险点]",
  "model": "claude-opus-4-7",
  "operator": "[姓名]（工号：[工号]）",
  "steps": [
    {
      "step": "数据确认与验证",
      "executor": "ai",
      "data_source": {"type": "user_upload"},
      "result": "通过/不通过",
      "duration_seconds": 0
    },
    {
      "step": "信号识别",
      "executor": "ai",
      "data_source": {"type": "user_upload"},
      "result": "通过/不通过/预警",
      "signals_found": 0
    },
    {
      "step": "行业趋势验证",
      "executor": "ai",
      "data_source": {"type": "reference"},
      "result": "通过/降级"
    },
    {
      "step": "政策驱动力分析",
      "executor": "ai",
      "data_source": {"type": "reference"},
      "result": "通过/预警"
    },
    {
      "step": "技术替代评估",
      "executor": "ai",
      "data_source": {"type": "context"},
      "result": "通过/预警"
    },
    {
      "step": "财务敏感性测算",
      "executor": "ai",
      "data_source": {"type": "user_upload"},
      "result": "通过/降级"
    },
    {
      "step": "综合判断与授信建议",
      "executor": "ai→human",
      "data_source": {"type": "context"},
      "confirmation": {"type": "approve", "approved_by": "[姓名]", "role": "信审人员"}
    }
  ],
  "warnings": ["如有"],
  "data_termination_triggered": false,
  "references_used": ["references/cot-signal-library.md", "references/industry-benchmarks.md", "..."]
}

Audit logs must be retained for ≥ 3 years.

Gotchas

#1: Failure to terminate when there is no data leads to fabricating reasoning chains

Symptom: Key account data is missing from the raw data, but the reasoning chain continues to output and cites non-existent numbers
Cause: The core constraint "terminate when no data" was ignored, and the model inferred and supplemented data on its own
Solution: Strictly enforce the termination condition of step 1—if data is missing, immediately terminate the analysis of that risk point and output no content

#2: Mixing consolidated statements and parent-company basis

Symptom: Parent-company income statement data is cross-used with consolidated balance sheet data, resulting in incorrect ratio calculations
Cause: The reporting basis was not confirmed at the beginning of the execution process, and the basis was switched arbitrarily during analysis
Solution: In the "read first, then write" stage at the beginning of the Workflow, the reporting basis must be confirmed and kept consistent throughout; consult references/financial-stress-test-params.md for the basis validation rules

#3: Merging aging analysis leads to missed judgments

Symptom: Simply merging different aging buckets (within 1 year / 1-2 years / 2-3 years / over 3 years) into a single total conceals abnormal growth in long-aging balances
Cause: Ignoring the requirement in the specific scenario specification to "list precise values year by year and aging bucket by aging bucket"
Solution: When aging analysis is involved, precise values must be listed year by year and aging bucket by aging bucket; merging or vague treatment is not allowed

#4: Industry benchmarking error — financial enterprises use manufacturing indicators

Symptom: Using manufacturing risk indicators such as the Altman Z-score for financial enterprises such as banks/insurers/brokerages, resulting in severe distortion
Cause: Failure to identify the enterprise's industry classification at the beginning of the execution process, directly applying the default analysis framework
Solution: Before Step 1 signal identification, first confirm the enterprise's industry; for financial enterprises, skip manufacturing indicators such as the Z-score and switch to the CAMELS framework (see references/industry-benchmarks.md）

Quality Requirements (Quality Requirements)

After each reasoning chain is generated, perform a self-check item by item against the following 8 items; output only after all pass:

Zero data fabrication: All cited numbers must be fully consistent with the original data (including units and decimal places); if there is no textual basis, output must be terminated and no speculative supplements may be added. See Constraints #1、Constraints #2。
Complete intermediate stepping stones: The reasoning chain must include a complete transmission path (signal → phenomenon → cause → impact → risk conclusion); jumping directly from signal to conclusion is prohibited. See Workflow Step 1。
Full coverage of the six steps: Every CoT output must go through all 6 derivation framework steps; if any step is missing, the reason for skipping must be explained in the report. See Constraints #7。
Explicit labeling of data sources: Every cited data item must be labeled with its source (report type + year + account), format: "(from 2023 consolidated income statement - operating revenue)". See Output Format output field specification.
Consistency of reporting basis: Consolidated statement data may only be compared with consolidated statement data, and parent-company-basis data may only be compared with parent-company-basis data; mixing is strictly prohibited. See Constraints #5、Terminology。
Quantitative conclusions preferred: The final credit recommendation must be accompanied by specific quantitative thresholds (e.g., "when DSCR < 1.2, recommend reducing the limit to 70% of the existing credit line"); purely qualitative statements are prohibited. See Workflow Step 6。
Balanced positive and negative evidence: Each risk point must present both the evidence supporting that risk and the exceptional circumstances that may weaken that risk, avoiding one-sided reinforcement. See Constraints #8。
Explicit labeling of uncertainty: Conclusions based on inference (not original textual data) must be labeled "(inference)", and data-missing sections must be labeled "(insufficient data, pending verification)". See Constraints #9、Output Format output field specification.

Examples (Examples)

Example 1: Standard scenario — reasoning chain for abnormal growth in construction in progress

User input：

Original credit data: the 2021-2023 consolidated balance sheets of a chemical enterprise show that construction in progress increased from 210 million to 870 million (+314%), while operating revenue decreased from 1.23 billion to 1.08 billion (-12%) over the same period. Target risk point: the enterprise has a risk of blind capacity expansion, and its long-term solvency is questionable. Industry: basic chemicals - polyurethane.

Skill execution process：

Data confirmation: read the financial statements → confirm the consolidated statement basis → verify the cross-footing relationships → passed
Signal identification: Precisely cited construction in progress growth of 314% and revenue decline of 12% → Identified "asset-liability anomaly + scale mismatch" signal → Passed
Industry trend verification: Consulted industry-benchmarks.md → Polyurethane industry capacity utilization fell from 78% to 65%, while new capacity across the industry grew by 45% → Industry systemic risk confirmed → Passed
Policy driver analysis: Consulted policy-regulatory-framework.md → The "dual carbon" policy tightens approval for high-energy-consumption chemical projects → Tightening direction → Passed
Technology substitution assessment: New eco-friendly materials create substitution pressure on traditional polyurethane → Estimated 3-5 years → Passed
Financial sensitivity calculation: Assuming capacity utilization falls to 55% → Revenue declines by 25% → EBITDA insufficient to cover interest → DSCR = 0.8 → Passed
Comprehensive judgment: This is bandwagon capacity expansion lacking a core technology moat → Recommend tightening credit, with reduction when DSCR < 1.2

Output summary: Construction in progress surged 314% while revenue fell 12%, forming a severe divergence; industry overcapacity is confirmed, policy tightening is compounded by technology substitution pressure, and the stress test DSCR = 0.8 is below the safety line. Recommend prudently maintaining credit and setting DSCR < 1.2 as the trigger for credit reduction.

Example 2: Insufficient data scenario—terminate output