Part 2: Unlocking AI's Potential in Finance: Insights from QuantUniversity's Boston Fintech Week Workshop

QuantUniversity organized a half-day workshop on AI and Investing, bringing together an incredible panel of industry leaders as a part of Boston Fintech Week. In Part 1 of this series of articles, we covered insights from the panel on:

???Opportunities for FinTech Startups in AI

???AI Adoption in the FinTech Industry

???Bridging the Gap Between Startups and Enterprises

???The Future of AI in FinTech

In this part, we focus on Dr. Arun Verma from 彭博资讯 ’s lecture on Generative AI and Foundation Models for Quantitative Finance

Exploring the Synergy between Quantitative Finance and Machine Learning

Arun emphasized the importance of exploring the synergy between traditional quantitative finance (Quant) and machine learning (ML), with a focus on generative AI and foundation models. He highlighted some key use cases of generative ML in finance, such as:

• Time Series Forecasting
• Imputation of Missing Data
• Geographic Exposure Analysis

Differences between Quant and ML Communities

He noted the distinct characteristics of the Quant and ML communities:

• Quantitative Finance: Rooted in theoretical models and numerical methods, with an emphasis on causality, interpretability, and structured assumptions. It often deals with "small data" and a limited number of parameters—think classic option pricing models.
• Machine Learning: Relies on data-driven approaches to identify patterns, leveraging "big data" and models with extensive parameter spaces. ML often focuses on correlations, leading to black-box models—examples include image recognition and language models.

Evolution of Time Series Models

Arun traced the evolution of time series modeling, moving from traditional methods to the latest advancements in the field:

• Traditional Models: ARIMA, GARCH, Neural Networks (FNN, RNN, LSTM), Temporal CNNs (WaveNet).
• Recent Advancements: Transformers (LLMs), Diffusion Models, and Foundation Models.

Foundation Models for Time Series

Arun discussed foundation models and their applicability in time series analysis:

• Objective: Develop a versatile model capable of handling various time series tasks, such as forecasting (short and long horizon), anomaly detection, imputation, and classification.
• Approach: Pre-training a foundation model on extensive datasets, then fine-tuning or using it directly (zero-shot) for specific tasks.

Use of Generative ML for Imputation in Finance

The lecture also highlighted examples of using generative ML for imputation in finance, addressing the challenge of missing data:

• Importance: Missing data can arise from non-disclosure, inaccessibility, or future forecasting needs, making it crucial for accurate financial analysis.
• Applications:Predicting water use and GHG emissions for climate risk assessment.Estimating facility locations using manifold normalizing flows.Analyzing geographic exposure using collaborative filtering techniques.

Different Approaches to Imputation: Quant vs. ML

Finally, Arun compared how quantitative and ML approaches handle imputation:

• Quant Approach: Utilizes Bayesian methods like Kalman Filtering to estimate missing values based on observed data and process covariance.
• ML Approach: Leverages implicit factor models with instrument embeddings, incorporating asset characteristics and time series properties for imputation.

Dr. Arun Verma's lecture underscored how the Quant community is increasingly embracing AI and ML as powerful tools, providing numerous use cases and examples. It’s evident that the fusion of these disciplines presents exciting opportunities for financial innovation.

Stay tuned for Part 3 to be released tomorrow!

Thanks to Babson College for hosting this workshop!

If you would like to review the video and slides from the workshop, you can review them here!

Yours,

Sri