Reports

Final Project Report

Project No.: ARP_112023_08 Grant Number: L.015604-41-IADF-01 Project Duration: 12 months Funding: SNSF Leading House Asia - Bilateral Science and Technology Programme with China 2021-2024

---

Executive Summary

This project established a sustained Swiss-Chinese research collaboration in financial machine learning, specifically addressing graph-based credit risk assessment in consumer lending markets. Despite regulatory constraints preventing access to proprietary Chinese platform data, the research pivoted successfully to public peer-to-peer (P2P) lending datasets, yielding methodological contributions with broader applicability and enhanced reproducibility for the academic community.

The collaboration produced six peer-reviewed publications, multiple conference presentations, trained doctoral researchers, and established frameworks for continued bilateral research. The methodological innovation, a homophily-guided graph neural network for credit risk prediction, demonstrates statistically significant improvements over state-of-the-art baseline methods across multiple international datasets.

---

1. Research Activities and Scientific Contributions

1.1 Research Exchange Program

The bilateral exchange program facilitated knowledge transfer across institutional and cultural boundaries, enabling the synthesis of complementary expertise in statistical methodology (Chinese partners) and industry applications (Swiss partners).

Date	Activity	Participants	Location	Outcomes
Sep 2023	International Week lectures	Joerg Osterrieder, Yiting Liu, Lennart Baals	Shenzhen Technology University	Initial methodology transfer, network establishment
Sep 17-22, 2024	Research planning and methodology alignment	Yiting Liu, Jeffrey Chu	Beijing	Data pipeline specification, model architecture design
Sep 23-28, 2024	International Week lectures	Yiting Liu, Joerg Osterrieder	Shenzhen Technology University	Knowledge transfer to 50+ practitioners
Nov 2-3, 2024	Blockchain and DeFi Workshop	Jeffrey Chu	HKUST Guangzhou	Network expansion, methodology feedback
Dec 13-17, 2024	Big Data and AI Symposium	Yiting Liu	HKUST Hong Kong	Research presentation to 200+ attendees
Dec 17-18, 2024	Project synthesis meeting	Yiting Liu, Jeffrey Chu	Beijing	Final methodology consolidation
Feb 20-22, 2025	ICMS 2025 Conference	Jeffrey Chu, Stephen Chan	AUS Dubai	Paper presentation, collaboration development
Jul 11-13, 2025	JCSDS Conference	Stephen Chan, Yuanyuan Zhang	China	Statistical methodology exchange
Aug 21-23, 2025	EcoSta 2025	Jeffrey Chu, Stephen Chan	Waseda University, Tokyo	Organized invited session on digital finance
Oct 20-23, 2025	CIEP 2025 - 23rd China International Talent Exchange Conference	Joerg Osterrieder	Shanghai	Keynote presentation on Digital Finance, LLM Narratives, HFT; Economic Resilience Roundtable; Meeting with Editors (JASA, Digital Finance, FRL, JIFMIM); Young Scholars Forum

1.2 Methodological Development

The core scientific contribution lies in the development of graph-based credit risk models that leverage network topology to improve default prediction. Traditional credit scoring methods treat loan applications as independent observations, ignoring potential correlations among borrowers with similar characteristics. Our methodology addresses this limitation through:

Graph Construction Innovation: Unlike existing graph neural network approaches that assume pre-existing network structures, we develop homophily-guided graph construction that creates meaningful borrower networks from tabular loan data. The approach exploits the principle that borrowers with similar default behavior tend to share observable characteristics.

Theoretical Foundation: The methodology builds upon foundational work in semi-supervised learning (Zhu et al., 2003), message passing neural networks (Gilmer et al., 2017), and the sociological concept of homophily (McPherson et al., 2001). By constructing graphs where edges connect borrowers with similar characteristics and outcomes during training, we ensure high label homophily that enables effective information propagation through the network.

Architecture Design: The Graph Attention Network (GAT) architecture provides both discriminative power and interpretability. The attention mechanism learns to weight neighbor contributions dynamically, producing predictions that can be explained in terms of similar historical borrowers, a requirement for regulatory compliance in credit decisions.

1.3 Empirical Validation

Comprehensive benchmarking against 15 baseline methods across 5 international datasets demonstrates consistent superiority of the proposed approach:

Dataset	Region	Loans	Baseline AUC	Our AUC	Improvement
Bondora	EU	134,529	0.778	0.812	+4.4%
LendingClub	US	2,260,668	0.762	0.798	+4.7%
German Credit	DE	1,000	0.752	0.781	+3.9%
Prosper	US	113,937	0.771	0.803	+4.1%
Home Credit	Global	307,511	0.775	0.809	+4.4%

Statistical significance confirmed via Friedman test (p < 0.001) and Nemenyi post-hoc comparisons. The improvements translate to substantial economic value: for a CHF 100M lending portfolio, improved default prediction reduces expected losses by CHF 500,000-900,000 annually.

---

2. Milestone Achievement

2.1 Database Development

Original Plan: Access Ant Group consumer lending data for Chinese market analysis.

Outcome: Adapted to public P2P lending datasets due to regulatory constraints (Personal Information Protection Law of China, 2021).

Scientific Value: The adaptation enhanced reproducibility and broader applicability. Public datasets enable independent verification of results and facilitate methodological comparison across the academic community. The datasets span European (Bondora), US (LendingClub, Prosper), and global (Home Credit) markets, providing validation across diverse regulatory and economic contexts.

2.2 Literature Review

Status: Completed

A systematic literature review following PRISMA guidelines identified 847 initial records, with 127 papers included in final synthesis. The review establishes:

1. Research Gap: Existing GNN applications in finance assume pre-existing network structures; credit risk requires graph construction from tabular data
2. Methodological Opportunities: Homophily principles from social network analysis provide theoretical foundation for graph construction
3. Regulatory Considerations: Interpretability requirements necessitate attention-based architectures over opaque alternatives

2.3 Methodology Paper

Status: Completed

The manuscript “Credit Risk Prediction via Graph Neural Networks with Homophily-Guided Graph Construction” presents:

• Novel graph construction methodology for tabular financial data
• Comprehensive benchmarking against 15 methods on 5 datasets
• Economic impact quantification framework
• Interpretability analysis for regulatory compliance

Target journal: Journal of Management Information Systems (Impact Factor: 7.2)

2.4 Software Development

Status: Completed

Production-ready implementation includes:

Component	Technology	Status
Graph Construction	NetworkX, Faiss	Complete
GNN Models	PyTorch Geometric	Complete
Training Pipeline	PyTorch Lightning	Complete
Evaluation Suite	scikit-learn	Complete
R Shiny Dashboard	Shiny, reticulate	Complete

---

3. Publication Portfolio

3.1 Papers Prepared for Submission

3.2 Papers Under Review

3.3 Published Papers

---

4. Adaptive Project Management

The project demonstrated effective adaptation while expanding scientific impact:

Initiative	Strategy	Outcome
Data access constraints	Pivoted to public P2P datasets	Enhanced reproducibility, broader applicability
Network expansion	Leveraged COST Action and MSCA connections	Access to 300+ researchers across 51 countries
Focused mentorship	Intensive PhD training model	Deeper engagement, higher quality outputs
Multi-institutional collaboration	Joint supervision with Chinese partners	Cross-cultural research synergies

---

5. Institutional Collaboration Framework

5.1 New Partnerships Established

MSCA Digital Finance Network: Prof. Joerg Osterrieder serves as Coordinator of the Marie Sklodowska-Curie Industrial Doctoral Network on Digital Finance (2024-2027). Prof. Jeffrey Chu joined as associated partner (2024). This integration enables PhD student exchanges and joint supervision opportunities across 15 institutions in 12 countries.

NSFC Collaboration: Swiss partners joined the Beijing Natural Science Foundation project IS23126 on network-based fraud detection (2023-2025), establishing reciprocal collaboration frameworks.

5.2 Network Integration

The project strengthened integration with major research networks:

Network	Role	Members	Countries
COST Action CA19130 (2019-2024, concluded)	Former Chair (Osterrieder)	300+	51
MSCA Digital Finance	Coordinator	15 institutions	12
SNSF Leading House Asia	Grantee	-	-

5.3 Sustainability Framework

Continued collaboration assured through:

1. Joint Grant Applications: SNSF BRIDGE (submitted Dec 2024), Innosuisse Bilateral (planned), NSFC Joint Projects (discussed)
2. Conference Organization: EcoSta 2025 invited session, future joint conferences
3. Student Exchange: Framework established for PhD visits in both directions
4. Joint Publications: Pipeline of collaborative papers in progress

---

6. Impact Quantification

6.1 Quantitative Outcomes

Metric	Target	Achieved	Assessment
Peer-reviewed publications	4	8+	Exceeded
Conference presentations	3	5+	Exceeded
PhD researchers trained	2	2	Met
Partner institutions	2	4	Exceeded
Follow-up funding applications	1	4	Exceeded
Software releases	1	1	Met

6.2 Qualitative Achievements

1. Methodological Innovation: Novel homophily-guided graph construction for credit risk assessment
2. Empirical Contribution: Comprehensive validation across 5 international datasets with 15 baseline comparisons
3. Economic Framework: Quantification methodology translating ML improvements to business value
4. Capacity Building: PhD researchers trained in cutting-edge graph neural network methodologies
5. Network Development: Strengthened Swiss-Chinese ties with clear pathways for continued collaboration

---

7. Future Directions

7.1 Immediate Next Steps (2025)

1. Publication Completion: Submit prepared manuscripts to target journals
2. Dissemination: Present at planned conferences (JCSDS, EcoSta 2025)
3. Software Release: Publish Python/R packages to PyPI and CRAN
4. Grant Decisions: Await SNSF BRIDGE decision (Q2 2026)

7.2 Medium-Term Plans (2025-2027)

Program	Amount	Status	Timeline
SNSF BRIDGE	CHF 130K	Submitted	Decision Q2 2026
Innosuisse Bilateral	CHF 300K	Planned	Application 2026
SNSF Sino-Swiss	CHF 500K	Planned	Monitoring calls
NSFC Joint Project	CNY 1M	Discussed	Partner application
SNSF Leading House Asia 2025	CHF 50K	Planned	New call expected

---

8. Appendix Materials

Supporting documents available in the Appendix:

Document	Description
JMIS Submission	Main methodology paper
P2P Lending Paper	Submitted to Financial Innovation
Ethereum Fraud Detection	JRSS-A accepted paper
Conference Photos	Evidence of collaboration

---

Download Reports

Report Versions

Version	Pages	Description	Template
ARP Official	11	Official submission format	ARP Reporting Form
Extended Academic	12	Detailed methodology	Custom academic

---

(c) Joerg Osterrieder 2025