CEFAR Awards

Diamond Awardee: ZENG Lingteng

Project Title: Multi-Market Collateral Risk Assessment Using Machine Learning and Financial Engineering

Project Abstract:

This paper presents a research-driven decision-support system for securities financing collateral management developed in collaboration with Hang Seng Bank. The project addresses a practical business question: under publicly observable market information and explicit acceptance constraints, how can a collateral receiver assess whether an incoming basket is acceptable, whether a baseline haircut remains sufficient, and how market expansion changes that judgement. The final artifact integrates five risk dimensions - market, concentration, liquidity, foreign exchange, and bond credit risk - with advanced Value-at-Risk decomposition, multi-scenario stress testing, basket-index tracking, counterparty monitoring, and automated risk reporting.  Methodologically, the project combines financial engineering with two AI-enhanced components: a machine-learning style market-regime retrieval engine that extracts multi-dimensional feature vectors and ranks historically similar periods via a weighted similarity ensemble, and a large-model reporting layer that transforms structured risk outputs into traceable English supervisory summaries without fabricating facts. The system was implemented as a full-stack platform with a Vue.js front end, a Flask-based analytics layer, and MySQL-backed portfolio storage. Project evidence shows end-to-end support for portfolio generation or import, advanced risk assessment, three-market coverage (Hong Kong, Japan, Taiwan), multi-currency analysis, five stress-testing families, counterparty diagnostics, and exportable reports. The study contributes a defensible industrial prototype that bridges academic rigor and operational realism under real-world data constraints.

Gold Awardee: SUN Guoxuan

Project Title: Justice AI Compass: A Universal AI Framework for Deterministic Legal Document Intelligence and Risk Pricing

Project Abstract:

The contemporary judicial system faces a systemic crisis characterized by an overwhelming volume of disputes and a critical shortage of analytical tools, resulting in a "Black Box" of litigation that imposes unquantifiable uncertainty on both individual litigants and corporate entities. This research introduces Justice AI Compass, a universal artificial intelligence framework engineered to transform unstructured judicial documents into deterministic, actionable risk metrics. The framework implements a multi-tiered methodology to ensure high-fidelity legal intelligence. First, an AI-anchored document segmentation model is deployed to isolate core functional zones, effectively eliminating semantic noise through a constrained loss function that combines Cross-Entropy with Intersectionover-Union (IoU) penalties. Second, to address the challenge of Extreme Multi-label Classification (XMC) in legal provision prediction, we propose an Adaptive Dual-Stage LightXML architecture. This mechanism utilizes a dynamic "Recall and Rank" cascade to overcome the limitations of static negative sampling, ensuring high precision even in sparse label spaces. Finally, a Unified Multi-Task Pricing Engine, built upon a 12-layer Chinese RoBERTa backbone, concurrently predicts categorical joint liability, discrete fee intervals, and continuous fee burden ratios by harmonizing divergent loss functions, including Focal BCE and Huber Loss. Empirical evaluation on a foundational subset corpus of over 100 million judicial judgments demonstrates significant performance gains. The Adaptive Dual-Stage LightXML achieves a Precision@1 of 0.88, while the Unified Pricing Engine records a Task 2 accuracy of 0.829, representing a 22% improvement in Macro-F1 over traditional single-task baselines. Beyond algorithmic innovation, the Justice AI Compass provides a scalable socio-technical intervention that democratizes legal certainty for individual citizens (B2C), enables quantitative ROI calculation for enterprises (B2B), and facilitates systemic relief for overextended judicial resources (B2G). This work establishes a market-ready standard for the "Financial Truth"in litigation, shifting the paradigm from probabilistic text generation to deterministic legal risk quantification.

Silver Awardee: WANG Tianyi

Project Title: Liquidation-hunting Trading System on Aster Futures Strategy Design, System Architecture, and Risk-Controlled Execution

Project Abstract:

The rapid growth of cryptocurrency perpetual futures markets has introduced significant market microstructure phenomena, particularly forced liquidation events. These non-discretionary, margin-driven order flows create short-lived price dislocations that present systematic trading opportunities. This paper details the design, implementation, and evaluation of a real-time liquidation-hunting trading system built on the Aster decentralized perpetual futures exchange. The system employs a contrarian strategy formalized through a multi-layer signal filtering pipeline—incorporating volume significance thresholds, price proximity validation, VWAP-based directional gating, and temporal cooldown mechanisms—to transform noisy liquidation streams into actionable, risk-managed trading signals. A dual-process system architecture separates the web-based observability dashboard from the core trading service, ensuring operational resilience and separation of concerns. The project delivers a complete research-to-engineering pipeline, including a unified paper trading and live execution framework that minimizes behavioral divergence between testing and production environments. Key contributions include the formalization of a parameterized liquidation-contrarian strategy, a robust event-driven system methodology, and a practical validation framework for algorithmic trading deployment. Limitations including the paper realism gap and market regime sensitivity are acknowledged, with future directions proposed in adaptive thresholding and regime classification.