Enhancing Financial Synthetic Data Generation Through Local Historical Pattern Retrieval

Kurugundra, Vasudha (2025) Enhancing Financial Synthetic Data Generation Through Local Historical Pattern Retrieval. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (559kB) | Preview

Abstract

Financial markets exhibit complex, non-stationary behaviors with regime transitions that challenge traditional synthetic data generation methods. This research investigates the enhancement of generative models through local historical pattern retrieval for improved financial synthetic data generation. A context-enhanced framework was developed that integrates Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) with a self-contained historical pattern retrieval system utilizing FAISS vector similarity search. The implementation employs a wrapper-based architecture that preserves trained model integrity while providing inference-time context integration through attention mechanisms. Historical S&P 500 data spanning 2006-2024 was processed using comprehensive feature engineering (109+ technical indicators) to create a sequence database of 60-day market windows. The system utilizes regex-based natural language processing for market regime queries and PCA-reduced embeddings for efficient similarity search. Experimental evaluation demonstrates significant performance improvements over baseline models: context-enhanced VAE achieved 90.7% performance compared to75.4% baseline (+15.3%), while context-enhanced GAN achieved 88.4%compared to72.3% baseline (+16.1%). The average improvement of 15.7%across both architectures validates the effectiveness of local historical pattern retrieval for financial data generation. The self-contained implementation addresses practical deployment requirements for financial institutions, eliminating external dependencies while maintaining superior performance. These findings contribute empirical evidence for wrapper-based enhancement architectures and provide a production-ready framework for context-enhanced financial synthetic data generation.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Subhnil, Shubham UNSPECIFIED
Subjects:	H Social Sciences > HG Finance Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	01 Jul 2026 11:24
Last Modified:	01 Jul 2026 11:24
URI:	https://norma.ncirl.ie/id/eprint/9434

Actions (login required)

View Item