基于风险评分与风险事件生存分析的ST预测

Quantitative Models and Construction 1. Model Name: Financial Report Fraud Detection Score - Model Construction Idea: This model identifies and quantifies the risk of financial fraud in corporate reports by detecting anomalies such as fabricated revenue, inflated profits, hidden liabilities, and other manipulative practices[20][39]. - Model Construction Process: - The model captures six types of anomalies: fabricated/inflated revenue, inflated profits, inflated assets, hidden liabilities, fund misappropriation, and benefit transfers[20]. - Scores are assigned based on the severity of detected anomalies, with higher scores indicating higher fraud risk[20]. - The model updates its scores in sync with financial report disclosures[20]. - Model Evaluation: The model demonstrates strong predictive power for ST events, with an AUC of 0.806, significantly outperforming the benchmark factor of net profit (AUC = 0.656)[5][34][38]. 2. Model Name: Corporate Behavior Profiling Score - Model Construction Idea: This model evaluates the operational health of companies and quantifies their default risk based on comprehensive behavioral profiling[20]. - Model Construction Process: - The score is derived from operational data, with higher scores indicating better operational health and lower risk[20]. - The model is updated in alignment with financial report disclosure schedules[20]. - Model Evaluation: This model outperforms the Financial Report Fraud Detection Score, with an AUC of 0.865, and demonstrates stronger predictive power for ST events[5][34][39]. 3. Model Name: Cox Proportional Hazards Model (Survival Analysis) - Model Construction Idea: This model predicts the probability of ST events by analyzing the impact of risk events (e.g., regulatory actions, shareholder actions) on the survival time of stocks[44][61]. - Model Construction Process: - The survival function $S(t)$ is defined as the probability of a stock surviving beyond time $t$, with $S(0) = 1$ and $S(t)$ decreasing over time[41][44]. - The Cox model formula is: $h(t, X) = h_0(t) \cdot \exp(\beta_1 x_1 + \beta_2 x_2 + \ldots + \beta_n x_n)$ where $h_0(t)$ is the baseline hazard, $X$ represents risk events, and $\beta$ are coefficients[61]. - Risk ratios (HR) are calculated as $HR = e^\beta$, with $HR > 1$ indicating risk factors and $HR < 1$ indicating protective factors[61]. - Risk events include regulatory letters (e.g., inquiry letters), shareholder actions (e.g., equity freezes), and regulatory measures (e.g., public censure)[45][61]. - Model Evaluation: The model achieves an AUC of 0.810 and demonstrates flexibility in incorporating new risk events, making it suitable for high-frequency monitoring[6][73][75]. --- Model Backtesting Results Financial Report Fraud Detection Score - AUC: 0.806[5][34] - Recall Rate: Average recall rate exceeds 90% for ST predictions, with a minimum recall rate above 80%[5][30]. Corporate Behavior Profiling Score - AUC: 0.865[5][34] - Recall Rate: Average recall rate exceeds 90%, with a minimum recall rate above 80%[5][31]. Cox Proportional Hazards Model - AUC: 0.810[6][73] - Recall Rate: Recall rates improved significantly after 2019, averaging around 85%[6][71]. --- Quantitative Factors and Construction 1. Factor Name: Financial Report Fraud Detection Score - Factor Construction Idea: Quantifies the risk of financial fraud based on anomalies in financial reports[20]. - Factor Construction Process: - Scores are assigned based on six types of anomalies, with higher scores indicating higher fraud risk[20]. - Factor Evaluation: Demonstrates strong predictive power for ST events, with an AUC of 0.806[5][34]. 2. Factor Name: Corporate Behavior Profiling Score - Factor Construction Idea: Measures operational health and default risk based on corporate behavior[20]. - Factor Construction Process: - Scores are derived from operational data, with higher scores indicating better operational health[20]. - Factor Evaluation: Outperforms the Financial Report Fraud Detection Score, with an AUC of 0.865[5][34]. 3. Factor Name: Risk Events (Cox Model) - Factor Construction Idea: Analyzes the impact of risk events on stock survival time[44][61]. - Factor Construction Process: - Risk events include regulatory letters, shareholder actions, and regulatory measures[45][61]. - Risk ratios are calculated to determine the significance of each event[61]. - Factor Evaluation: Provides high-frequency tracking and flexibility, with an AUC of 0.810[6][73]. --- Factor Backtesting Results Financial Report Fraud Detection Score - AUC: 0.806[5][34] - Recall Rate: Average recall rate exceeds 90% for ST predictions, with a minimum recall rate above 80%[5][30]. Corporate Behavior Profiling Score - AUC: 0.865[5][34] - Recall Rate: Average recall rate exceeds 90%, with a minimum recall rate above 80%[5][31]. Risk Events (Cox Model) - AUC: 0.810[6][73] - Recall Rate: Recall rates improved significantly after 2019, averaging around 85%[6][71]