Debiasing Alternative Data for Credit Underwriting Using Causal Inference

Abstract:

Alternative data provides valuable insights for lenders to evaluate a borrower's creditworthiness, which could help expand credit access to underserved groups and lower costs for borrowers. But some forms of alternative data have historically been excluded from credit underwriting because it could act as an illegal proxy for a protected class like race or gender, causing redlining. We propose a method for applying causal inference to a supervised machine learning model to debias alternative data so that it might be used for credit underwriting. We demonstrate how our algorithm can be used against a public credit dataset to improve model accuracy across different racial groups, while providing theoretically robust nondiscrimination guarantees.

Bio:

Chris Lam is the founder and CEO of Epistamai, an early-stage AI startup focused on using causal AI for high stakes decision making. He previously worked as a data scientist at the Federal Reserve Bank of Chicago, where he did research on algorithmic bias in credit decisions. He was also a technology strategist at HP and a project leader at Consumer Reports. He has a BS in computer science from the University of Pennsylvania, an MS in electrical engineering from Columbia University, and an MBA from Northwestern University (Kellogg).

Slides

Robust financial calibration: A Bayesian approach for Neural SDEs

Abstract:

We present a Bayesian framework for the calibration of financial models using neural stochastic differential equations (neural SDEs), for which we also formulate a global universal approximation theorem based on Barron-type estimates. The method is based on the specification of a prior distribution on the neural network weights and an adequately chosen likelihood function. The resulting posterior distribution can be seen as a mixture of different classical neural SDE models yielding robust bounds on the implied volatility surface. Both, historical financial time series data and option price data are taken into consideration, which necessitates a methodology to learn the change of measure between the risk-neutral and the historical measure. The key ingredient for a robust numerical optimization of the neural networks is to apply a Langevin-type algorithm, commonly used in the Bayesian approaches to draw posterior samples.

Bio:

Eva Flonner is a research associate at University of Applied Sciences Wiener Neustadt specializing in computational and mathematical finance, with a strong focus on theoretical and practical aspects of machine learning. She holds a PhD in Mathematics from Vienna University of Economics and Business and has gained practical experience in the areas of Digital Finance Transformation and Equity Research at leading Austrian companies. Outside of work, she enjoys running and classical music.

Slides are not being distributed per speaker request

Stories that (are) Move(d by) Markets: A Causal Exploration of Market Shocks and Semantic Shifts across Different Partisan Groups

Abstract:

Macroeconomic fluctuations and the narratives that shape them form a mutually reinforcing cycle: public discourse can spur behavioural changes leading to economic shifts, which then result in changes in the stories that propagate. We show that shifts in semantic embedding space can be causally linked to financial market shocks -- deviations from the expected market behaviour. Furthermore, we show how partisanship can influence the predictive power of text for market fluctuations and shape reactions to those same shocks. We also provide some evidence that text-based signals are particularly salient during unexpected events such as COVID-19, highlighting the value of language data as an exogenous variable in economic forecasting. Our findings underscore the bidirectional relationship between news outlets and market shocks, offering a novel empirical approach to studying their effect on each other.

Bio:

Felix is a final-year PhD student in Machine Learning at the University of Oxford’s Department of Engineering Science, co-supervised by Janet B. Pierrehumbert and Stefan Zohren. His research is focused on incorporating Natural Language Processing into forecasting settings. His research has two main themes: first, he explores how to extract and encode text to help improve economic, financial, or epidemiological forecasts; second, he tests whether modern LLMs are suitable in this setting. Felix is interested in understanding how the temporal bias implicit within statically trained LLMs affects predictions. He has probed LLMs for temporal leakage, developed point-in-time training regimes that remove look-ahead bias, and is currently researching methods to scale temporal bias removal to larger models through targeted knowledge editing.

Slides

Rolling-robust (R2) Uniform Manifold Approximation and Projection for Dimension Reduction

Score-Regularized GAN for Simulating Private Equity Contribution Patterns

Integrated Gradients in Finance: Baseline Effects and Stability across CNNs and Transformers

Automating Franchise Document Analysis with Large Language Models

Satellite Images for Physical Risk Modelling and Economic Activity

Failure Modes in Generative AI through Optical Illusion Interpretation

Graph-Driven Portfolio Optimization: EI-Minimizing GAT, Pattern Recognition with FFT, and HMM Volatility Regimes

Taiwan Real Estate Market Analysis: Comprehensive Data Infrastructure Development and Price Index Forecasting

Enhancing Financial Diffusion of Tabular Data

Rolling Robust Regime Detection (R2-RD) and Explaining Transitions through Decision Trees

Portfolio Rebalance with Tax Optimization

Graph Neural Network Models for Retail Basket Recommendation Prediction

LLM and Graphs for Financial Texts

Transfer Learning using quasi-randomized networks

Anomaly Detection of Time Series Data

Three-Factor OLS Model for Prediction

Attention Based High Dimension Representative for R2-UMAP

Predicting NASDAQ Equity Auction Volume Using Convolutional Neural Networks

MIXTURE OF EXPERTS-DIRECT PREFERENCE OPTIMIZATION

SCALING CONDITIONAL AUTOENCODERS FOR PORTFOLIO OPTIMIZATION VIA UNCERTAINTY-AWARE FACTOR SELECTION

Efficient CVaR Estimation for Mortgage Pools via Static Importance Sampling

Generating Market Color with Large Language Models

Deep Generative AI for Portfolio Management

Extending Macroeconomic Regime Detection with News Sentiment & Topic Modeling

RAG-Based Synthetic Scenario Modeling

An NLP Model Focused on Removing the Necessity of Prompt Engineering

Explaining Anomalous Equity and Credit Market Movements via Textual Data

GraphRAG for ECB AI Assistant

TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are stronger Zero-Shot Forecaster

Knowledge Graphs for financial text extraction

REGIME-AWARE REVENUE PREDICTION

Forecasting S&P 500 Volatility with Machine Learning Models

Mean-variance and hierarchical risk parity: An empirical study of large-cap stock portfolios

Exploring LLM Biases in Hotel Booking

Stock Predictability around S&P 500 Rebalancing Events

Deep Learning-Driven Portfolio Construction under PolyModel Theory

BIAS-DRIVEN GRAPH TRANSFORMERS WITH MEMORY

Fair Classification with Tail-Distribution Constraints

Enhance Stability Dynamic Graphs Using Bias-Driven Graph Transformers and Memory Mechanism

LEARNING TO TRADE WITH PREFERENCES: INTERPRETABLE EXECUTION

Rolling-robust (R2) Uniform Manifold Approximation and Projection for Dimension Reduction

Recurrent Lagrangian PPO with Decision Tree in Trading

Modeling A Systematic Withdrawal Plan: A Stochastic Algorithm to Estimate Initial Fund Requirement

Portfolio Rebalance with Tax Optimization

Sales Prediction

Applying Fairness in Multi-Label ML for Financial Transaction Modelling

Abstract:

This talk explores a practical strategy for mitigating bias in multi-label machine learning models used in financial transaction analysis. Drawing from a time-constrained collaborative research setting, I will share how I implemented a fairness-aware method based on adversarial training to reduce demographic leakage from model outputs.

The session will begin with a high-level overview of how fairness can be defined, measured, and addressed in financial ML systems. It will then focus on the implementation of an adversarial debiasing architecture, where a main model learns to predict transaction labels while an auxiliary adversary attempts to infer protected attributes from its outputs, encouraging the system to make predictions that are both accurate and less biased. Examples shown will be based on synthetic data. 

Bio:

Deniz is a London-based data science manager at PwC UK, specializing in guiding organizations across retail, financial services, and insurance to develop impactful data-driven strategies through analytics and AI. She holds a MSc in Computer Engineering, as well as a BSc in Statistics and Operations Research, combining strong analytical rigor with technical knowledge.

Her expertise spans time series forecasting, machine learning, deep learning, and generative models, with experience designing and deploying supervised, unsupervised, ensemble, and generative AI solutions. Deniz is also passionate about ethical AI, having contributed to research on mitigating bias in financial machine learning models through the Alan Turing Institute.

As a Women in Data ambassador at the firm, Deniz actively promotes diversity and inclusion within the data science community.

Slides

Beyond the Reported Cutoff: Temporal and Cross-Sectional Knowledge Biases of LLMs in Financial Domains 

Abstract:

Large Language Models (LLMs) are frequently utilized as sources of knowledge for question-answering. While it is known that LLMs may lack access to real-time data or newer data produced after the model's cutoff date, it is less clear how their knowledge spans across historical information. In this study, we assess the breadth of LLMs' knowledge using financial data of U.S. publicly traded companies by evaluating more than 197k questions and comparing model responses to factual data. We further explore the impact of company characteristics, such as size, retail investment, institutional attention, and readability of financial filings, on the accuracy of knowledge represented in LLMs. Our results reveal that LLMs are less informed about past financial performance, but they display a stronger awareness of larger companies and more recent information. Interestingly, at the same time, our analysis also reveals that LLMs are more likely to hallucinate for larger companies, especially for data from more recent years. We will make the code, prompts, and model outputs public upon the publication of the work. 

Bio:

Sudheer Chava is the Alton M. Costley Chair and Professor of Finance at Scheller College of Business at Georgia Institute of Technology, Atlanta. He serves as the director of the Masters in Quantitative and Computational Finance (MS-QCF) program, Financial Services Innovation Lab and the Center for Finance and Technology at Georgia Tech. Sudheer received his Ph.D. from Cornell University and an MBA from Indian Institute of Management – Bangalore. His expertise is in Banking, FinTech and, Applications of AI in Finance. More than 50 undergrad, MS and Ph.D. students in Finance and Machine Learning conduct research in his Financial Services Innovation Lab. Dr. Chava has consulted for many major financial institutions. He teaches executive education programs on FinTech and is an advisor to FinTech startups.

Slides

Deep Learning for Limit Order Book Forecasting

Abstract:

Limit Order Book (LOB) markets drive modern electronic trading, where buy and sell orders are dynamically matched to determine asset prices. Forecasting price movements in this high-frequency setting is inherently complex due to the high-dimensional and noisy nature of LOB data. This presents both a challenge and an opportunity for deep learning models, which can extract hidden patterns but must be carefully evaluated for real-world applicability.

By analysing a heterogeneous set of stocks traded on the NASDAQ exchange in this talk we will I) assess stocks’ predictability in relation to their microstructural features, such as their so-called tick size, ii) extend model’s evaluation beyond standard machine learning metrics to consider the feasibility of actual trading strategies carried out using the forecast. 

Finally, we will discuss the integration of deep learning architectures with network-theoretic representations that capture complex and non-trivial dependency structures among volume levels. This approach offers new insights into the spatial distribution and temporal degradation of information in LOBs, bridging the gap between microstructural modelling and deep learning-based forecasting in high-frequency financial markets.

Bio: 

Dr. Silvia Bartolucci is an Associate Professor in the Department of Computer Science at University College London (UCL), where she is part of the Financial Computing and Analytics Group. 

Her research leverages network, statistical physics and data-driven modelling tools to investigate critical behaviour in socio-economic systems and market microstructure, monitoring emerging trends and systemic risks in traditional and decentralized financial markets. 

Before joining UCL, Dr. Bartolucci was a Research Associate in the Department of Finance at Imperial College Business School, working within the Centre for Financial Technology. She holds a Ph.D. in Applied Mathematics from King's College London and a background in Theoretical Physics from Sapienza University in Rome. 

Slides

Privacy Amplification by Structured Subsampling for Deep Differentially Private Time Series Forecasting

Abstract:

Many forms of sensitive data, such as web traffic, mobility data, or hospital occupancy, are inherently sequential. The standard method for training machine learning models while ensuring privacy for units of sensitive information, such as individual hospital visits, is differentially private stochastic gradient descent (DP-SGD). However, we observe in this work that the formal guarantees of DP-SGD are incompatible with time series specific tasks like forecasting, since they rely on the privacy amplification attained by training on small, unstructured batches sampled from an unstructured dataset. In contrast, batches for forecasting are generated by (1) sampling sequentially structured time series from a dataset, (2) sampling contiguous subsequences from these series, and (3) partitioning them into context and ground-truth forecast windows. We theoretically analyze the privacy amplification attained by this structured subsampling to enable the training of forecasting models with sound and tight event- and user-level privacy guarantees. Towards more private models, we additionally prove how data augmentation amplifies privacy in self-supervised training of sequence models. Our empirical evaluation demonstrates that amplification by structured subsampling enables the training of forecasting models with strong formal privacy guarantees.

Bio:

Anderson is an Executive Director at the Morgan Stanley Machine Learning Research Department. He joined Morgan Stanley in 2019. Anderson was previously a quantitative researcher/trader at Infinium Capital and Tower Research Capital and a senior quantitative researcher at Graham Capital. Anderson has authored and co-authored conference and journal papers. He received his Ph.D. in Economics from University of Minnesota.

Slides