Semi-Supervised Clustering for Financial Risk Analysis.

Area of Science:

• Computational finance
• Machine learning
• Data science

Background:

• Conventional financial risk analysis methods, including unsupervised and supervised approaches, have limitations in accuracy, semantic understanding, and data requirements.
• Existing semi-supervised methods struggle with the inherent low distinguishability of financial data, hindering accurate predictions.

Purpose of the Study:

• To develop an effective semi-supervised learning scheme for financial data prediction that requires minimal labeled data.
• To enhance the performance of semi-supervised approaches by addressing the challenge of data distinguishability in financial datasets.

Main Methods:

• A novel approach converts labeled data into global prior probabilities, propagating 'soft' probabilities instead of 'hard' labels.
• A label diffusion model is employed to adaptively integrate information from both feature and label spaces.
• This model ensures greater consistency between data affinity structures and labeling information.

Main Results:

• The proposed method demonstrates improved prediction accuracy on financial data compared to existing semi-supervised techniques.
• Experiments on two public financial datasets validate the effectiveness and robustness of the label diffusion model.
• The approach successfully handles the low distinguishability characteristic of financial data.

Conclusions:

• The developed semi-supervised scheme, utilizing soft probability propagation and a label diffusion model, offers a promising solution for financial data prediction.
• This method effectively overcomes limitations of traditional approaches, providing accurate predictions with limited labeled data.
• The findings suggest a significant advancement in applying machine learning to financial risk analysis.