LEVERAGING GENERATIVE ADVERSARIAL NETWORKS (GANS) FOR REALISTIC SYNTHETIC DATA GENERATION

Abstract

The intricate process of creating synthetic data requires precise mathematical and statistical replication of the original data parts. There are significant privacy concerns associated with using and sharing real data for research or model building in industries like banking because of the sensitive information that is often included. Also, real data could be hard to come by, especially in niche fields where it's expensive or difficult to collect a wide variety of high-quality records. Due to data scarcity or availability issues, machine learning model training and testing may be hindered. We tackle this problem in this article. To be more specific, we need to create a new dataset that shares characteristics with an existing stock market dataset. The anonymized input dataset has a number of issues, including an imbalance, missing rows, duplicates, and improper data formatting (no columns or rows), as well as no normalized, scaled, or balanced values. Here, we take a look at generative adversarial networks as a deep-learning strategy, assess its ability to produce synthetic data, and compare it to the original stock dataset. Making fake datasets that hide some information while imitating the input portions' statistical features is our innovation's meat and potatoes. To illustrate the point, synthetic datasets can replicate the actual dataset's stock price distribution, trading volume distribution, and market trend distribution. As a result of the increased variety in the produced datasets, academics and industry professionals are better equipped to investigate various market circumstances and investment approaches. This variety has the potential to make machine-learning models more resilient and better at generalising. The average, similarities, and correlations are the metrics we use to assess our artificial data.