Banks have made substantial investments in data centers over the past decade, positioning them well to leverage the advancements in AI technology. However, a significant challenge arises from the proprietary and complex nature of much of the data at their disposal. To address this issue, banks are exploring the option of creating synthetic data to train their AI models effectively. A recent report from JPMorgan’s AI research team, spearheaded by Vamsi Krishna Potluru, the research director for synthetic data, delves into the practical applications of synthetic data within the banking sector.
The research team, comprising esteemed individuals such as AI research MDs Tucker Balch and Manuela Veloso, along with Deepak Parmarand, a director of AI specializing in generative AI products, has identified financial markets simulation as a particularly fruitful application of synthetic data. While synthetic data does not possess the mystical predictive powers some may hope for in the stock market realm, it does offer a valuable platform for refining and evaluating investment strategies. By introducing more variability and mitigating time-period biases, synthetic data enables the creation of diverse market scenarios for testing trading algorithms, including counterfactual scenarios that challenge the robustness of existing strategies.
Nevertheless, the adoption of synthetic data is not devoid of challenges. For instance, when employed in generating document layouts, deep neural models exhibit limitations such as the need for initial annotations to initiate the process, incapacity to create new primitives, and issues related to image quality. In contrast, Bayesian neural networks show promise in addressing these shortcomings, as evidenced in the synthetic example provided in the report.
Another area where synthetic data shows promise is in text generation for financial documents. Similar to its application in market simulations, synthetic data can enhance models by introducing new data perspectives. For instance, it can generate modified versions of existing documents to reflect bearish sentiments as opposed to bullish ones, thereby enriching the dataset for training purposes.
The report concludes by acknowledging that synthetic data utilization is still nascent, emphasizing the critical need to distinguish between synthetic and real data as its significance grows. While JPMorgan has yet to implement synthetic data practices, it highlights the involvement of leading universities like Cornell and Stanford in leveraging JPMorgan’s datasets to advance algorithm development in areas such as anti-money laundering and market execution.
