Enhanced Financial Data Generation using Recurrent Neural Networks within Variational Autoencoders: a Sector-Based Analysis

EasyChair Preprint 12798

Abstract

This study proposes a new generative model for quarterly financial data time series that is based on a variational autoencoder (VAE). The program provides financial data for a synthetic firm in a certain industry throughout a year, comprising four quarters of data. This technology employs a recurrent neural network-based VAE to successfully capture both multivariate distributions and temporal dependencies in the data. Compared to classic models such as the Multivariate Normal Monte Carlo Model and the Multivariate Gaussian State Space Model, our model's synthetic samples are more realistic, with higher visual fidelity and lower discriminative scores. In addition to the basic model, the derivative model has a conditional channel that generates samples with preset future performance. The user-friendly interface of our product makes it simple to utilize. Analyzing the patterns and features of synthetic samples can reveal significant information on alpha factor trading and risk management measures. Furthermore, applying the model to various datasets has the potential to improve these findings even more.

Keyphrases: Financial data simulation, Recurrent Neural Network (RNN), generative model, time series analysis, variational autoencoder

@booklet{EasyChair:12798,
  author    = {Parth Garg and Pulkit Sharma and U M Prakash},
  title     = {Enhanced Financial Data Generation using Recurrent Neural Networks within Variational Autoencoders: a Sector-Based Analysis},
  howpublished = {EasyChair Preprint 12798},
  year      = {EasyChair, 2024}}