10th Annual Bloomberg-Columbia Machine Learning in Finance Conference 2024
The workshop is organized by:



The 10th annual Columbia-Bloomberg Machine Learning in Finance conference will be held at Lerner Hall at Columbia University on Thursday, September 19th, 2024.
Register Here
Participating Researchers/Practitioners
A Machine Learning Framework for Stress Testing Financial Portfolios with Coverage Guarantees
Abstract:
A standard approach to financial risk management involves the use of scenario analysis to stress test portfolios. In the context of an S&P 500 options portfolio, for example, a scenario analysis might predict a loss of $1 million if the S&P 500 drops by 3% and its implied volatility surface rises by 5 percentage points. But how accurate is this predicted loss of $1 million? It is typically calculated under the implicit but flawed assumption that changes in the non-stressed risk factors are zero and therefore ignores the often significant statistical dependencies among risk factors. Additionally, even if the changes in unstressed factors are set to their conditional expected values, the predicted loss may still be inaccurate due to convexity effects, especially in derivatives portfolios. Another limitation of the standard approach is that the predicted loss figures are rarely if ever back-tested or statistically validated. In this paper, we address these issues by proposing a machine learning framework for scenario analysis. While we still produce point estimates of expected scenario losses, the key feature of our algorithms is the construction of confidence intervals (CIs) for the scenario losses with the widths of the CIs adapting in an online manner. We use ideas from the conformal prediction and non-parametric statistics literature to provide long-run coverage guarantees for the CIs. In a series of numerical experiments we demonstrate the superiority of our approach over the standard approach to scenario analysis. (Based on joint work with Zhongze Cai and Xiaocheng Li of Imperial College London)
Bio:
Martin Haugh is an Associate Professor of Analytics and Operations Research at Imperial College Business School (ICBS). Prior to joining Imperial College London, he spent more than 10 years in the Department of IE & OR at Columbia University as well as 4 years working as a quant in the hedge fund industry in New York and London. His research interests are in quantitative finance and risk management, data analytics, and dynamic programming.
Optimizing Portfolio Construction with JAX: Leveraging Evolutionary Algorithms for Enhanced Investment Strategies
Abstract:
The use of Machine Learning (ML) for portfolio construction and optimization has brought significant innovation to the asset management industry. Portfolio optimization methods span from the traditional mean-variance to full-scale optimization where assumptions on return distributions can be relaxed as well as investor preferences. Typically, black-box algorithms (BBO) such as Genetic Algorithm (GA), are one of the engines used for multi-asset class portfolio optimization based where utility function assumption are non-convex. However, this problem easily becomes computationally heavy and running times can increase exponentially. Recent developments in computation tools for large-scale ML, combined with evolutionary algorithms, have shown promising results in solving high-dimensional portfolio optimization problems. Our work consists of incorporating JAX-based evolution strategies into Vanguard’s asset allocation model to solve the expected utility maximization problem and find the optimal for our $35 billion AUM multi-asset model portfolio proposition. JAX is a state-of-the-art ML framework developed by Google that allows large-scale parallelization and orders of magnitude of acceleration. Our research finds that by leveraging on this tool, not only we find a better optimum in over 85% of the cases tested, but also up to 270-fold improvements in running time. With this presentation, we take a step back from typical ML applications and techniques used for alpha-generation and introduce an impactful tool that can be applied transversally to a multitude of ML methods.
Bio:
Giulio Renzi-Ricci is head of portfolio construction for Europe in the Investment Strategy Group at Vanguard. Giulio specializes in multi-asset portfolio construction, single fund solutions, and econometric forecasting; and oversees the asset allocation of Vanguard’s multi-asset products and model portfolios across Europe. His research has covered topics on active-passive blending, dynamic portfolio optimization, and ESG investing. His research has been published in The Journal of Investment Management, The Journal of Portfolio Management and The Journal of Investment Strategies, among others. Prior to joining Vanguard, Giulio worked at NERA Economic Consulting and Banca IMI. Giulio has a B.Sc. in economics from Bocconi University and a M.Sc. in finance and economics from the University of Warwick.
Slides have been distributed to registered delegates
Early Price Trajectory Data and Efficiency of Market Impact Estimation
Abstract:
Market impact is an important problem faced by large institutional investors and active market participants. In this paper, we rigorously investigate whether price trajectory data from the metaorder increases the efficiency of estimation, from the view of the Fisher information, which is directly related to the asymptotic efficiency of statistical estimation. We show that, for popular market impact models, estimation methods based on partial price trajectory data, especially those containing early trade prices, can outperform established estimation methods (e.g. VWAP-based) asymptotically.
Bio:
Fengpei Li is a Researcher at Morgan Stanley Machine Learning Research Team. Fengpei specializes in Monte Carlo methods, data-driven optimization, and their connection to quantitative finance in areas including market impact, statistical arbitrage and options pricing. Prior to joining Morgan Stanley, Fengpei completed his Ph.D. in Operations Research from Columbia University and a B.S. in Mathematics from the University of California San Diego.
Non-adversarial Training of Neural SDEs with Signature Kernel Scores
Abstract:
Neural SDEs are continuous-time generative models for sequential data. State-of-the-art performance for irregular time series generation has been previously obtained by training these models adversarially as GANs. However, as typical for GAN architectures, training is notoriously unstable, often suffers from mode collapse, and requires specialised techniques such as weight clipping and gradient penalty to mitigate these issues. In this paper, we introduce a novel class of scoring rules on path-space based on signature kernels and use them as objective for training Neural SDEs non-adversarially. By showing strict properness of such kernel scores and consistency of the corresponding estimators, we provide existence and uniqueness guarantees for the minimiser. With this formulation, evaluating the generator-discriminator pair amounts to solving a system of linear path-dependent PDEs which allows for memory-efficient adjoint-based backpropagation. Moreover, because the proposed kernel scores are well-defined for paths with values in infinite dimensional spaces of functions, our framework can be easily extended to generate spatiotemporal data. Our procedure permits conditioning on a rich variety of market conditions and significantly outperforms alternative ways of training Neural SDEs on a variety of tasks including the simulation of rough volatility models, the conditional probabilistic forecasts of real-world forex pairs where the conditioning variable is an observed past trajectory, and the mesh-free generation of limit order book dynamics.
Bio:
Zacharia Issa is a quantitative researcher at the Bank of New York. His research interests include generative modelling for financial time series, non-parametric regime detection, regime clustering, and statistical testing for stochastic processes. Zacharia completed his PhD in Applied Mathematics at King's College, London, and holds a Master of Philosophy (Mathematics) degree from the University of Sydney.
Multi-Adapter joint fine-tuning of Diffusion Models with LoRA for Visual Illusions
Approximate Risk Parity with Return Adjustment and Approximation Bounds
Penalized Reflected Stochastic Gradient Langevin Dynamics for Deep Learning
A Novel Algorithm For Online Bilevel Optimization For Price Forecasting Around Market Moving News
Credit Card Risk Models: A Survey on Scoring and Limits
Diffusion Model for Financial Time Series
House Price Index Analysis - Affordability and Expectation
Generating Financial Losses From Extreme Events via GANs
Graph Neural Network Models for Retail Basket Recommendation Prediction
Clustering and Graph Representation Learning in Fraud Detection
Anomaly Detection for Time Series Data
Share Repurchase Strategy Optimization
Explainable Asset Allocation and Portfolio Construction
Explainable AI: Improving the Stability of LIME
eXplainable AI for finance
Abstract:
Artificial intelligence is creating one of the biggest revolution across technology driven application fields. For the finance sector, it offers many opportunities for significant market innovation and yet broad adoption of AI systems heavily relies on our trust in their outputs. Trust in technology is enabled by understanding the rationale behind the predictions made. To this end, the concept of eXplainable AI emerged introducing a suite of techniques attempting to explain to users how complex models arrived at a certain decision. For cross-sectional data classical XAI approaches can lead to valuable insights about the models' inner workings, but these techniques generally cannot cope well with longitudinal data in the presence of dependence structure and non-stationarity. Furthermore, the literature lacks a comprehensive framework for testing stability in explanations provided by state-of-art XAI methods. We here propose novel XAI techniques for AI systems which preserves and exploits the natural time ordering of the data.
Bio:
Prof. Dr. Branka Hadji Misheva is a Professor in Applied Data Science and Finance at BFH, working on AI applications in finance, XAI methods, network models and fintech risk management. At her role, she leads/co-leads many research and innovation projects among which the notable examples are: two large EU funded projects, the Fintech-ho2020: A Financial supervision and Technology compliance training programme and a COST Action on Fintech & Artificial Intelligence which is an interdisciplinary research network comprised of over 300 researchers from 51 countries globally. Hadji Misheva is also the co-lead of a Marie Sklodowska-Curie Action Industrial Doctoral Network on Digital Finance. The Action aims to foster the development of the next-generation researchers and professionals who can contribute significantly to the digital finance sector. The network is comprised of 18 partners from both academia and industry throughout Europe, among which Deloitte, Swedbank, Bank for International Settlements, and Raiffeisen Bank. Hadji Misheva has furthermore participated in the acquisition of over 20 SNF, Innosuisse and EU projects and she is the research author of over 30 papers in the field of machine learning applications to credit risk modelling, graph theory, predictive performance of scoring models, lead behavior in crypto markets and explainable AI models for credit risk management.
Leveraging Deep Graph Learning for Enhanced Decision-Making in Financial Trading Applications
Abstract:
This presentation explores the adoption of Deep Graph Learning (DGL) as a critical tool for enhancing decision-making in financial trading. DGL offers a sophisticated method for embedding complex relational data about market dynamics and trader behaviors into a graph-based analytical framework. By representing traders and market activities as nodes and edges, DGL facilitates the extraction of nuanced relational patterns and preferences, essential for developing dynamic, context-aware trade optimization systems. This talk will include a comparative analysis of traditional decision-making tools versus graph-based approaches, highlighting the distinct advantages and potential limitations of DGL in real-time trading environments. Additionally, we will examine the technical architecture and the challenges associated with implementing DGL in real-time applications. The session aims to provide an academic and practical perspective on how advanced deep learning techniques can be effectively applied in trading scenarios, showcasing both the methodology and its application in real-world settings.
Bio:
Mutisya Ndunda is the Head of Data Strategy and AI at Trumid. In his role, he is responsible for the development of Trumid’s data and analytics services, optimizing the end-to-end platform experience through the application of artificial intelligence and machine learning capabilities. Mutisya has over 20 years of capital markets experience and has done extensive work in quantitative, data-driven research and analysis. Before joining Trumid in 2020, Mutisya served as the CEO of Alpha Vertex, a fintech company specializing in AI-driven analytical solutions. Prior to this, he was the Global Head of Strategy and Business Development, Enterprise Solutions at Bloomberg, and spent three years at Susquehanna as Head of Business Development. Mutisya started his career in corporate strategy at Citi and then moved to Merrill Lynch. Mutisya holds a bachelor’s degree in electrical engineering and a master’s degree in financial engineering, both from Cornell University.
Giles Shaw is the Machine Learning Lead at Trumid. In this role, Giles is responsible for driving the identification, implementation, and use of machine learning solutions within the company. Prior to Trumid, Giles has worked in a number of different roles as an AI Researcher and Engineer, including building AI solutions for government agencies, and working with ex-Amazon Alexa creator on deep learning solutions. He holds a Master’s degree in mathematics from the University of Oxford and a PhD in mathematical analysis from the University of Cambridge.
Learning Not to Spoof: Normative Guidance for Autonomous Trading Agents
Abstract:
As intelligent trading agents based on reinforcement learning (RL) gain prevalence, it is important to ensure that RL agents obey laws, regulations, and human behavioral expectations. Existing RL Finance literature contemplates topics like variance penalization and catastrophic failure but is mostly silent concerning subtle non-normative behavior like market manipulation. Such behavior may violate legal or regulatory, rather than physical or monetary, constraints. For AI learning systems, it is difficult to conceptualize intent-based rules like those concerning disruptive trading practices. My current work explores methods to dissuade RL trading agents, tasked with profit maximization, from adopting spoofing as an “optimal” strategy. For both practical and ethical reasons, the investigations take place in a simulated stock market.
Today’s talk will begin with a brief overview of the ABIDES simulation, designed to support high-fidelity agent-based research in market applications, and a brief introduction to reinforcement learning, a branch of machine learning that seeks to optimize the total utility of a sequence of decisions. With this background in place, the bulk of the talk will follow a series of experiments in which an intelligent stock trading agent inadvertently learns to spoof the market in which it participates. The experiments include: hand-crafting a spoofing trader, observing its behavior to construct a spoofing detector, designing the policy space for an autonomous AI trading agent, noting that this agent discovers spoofing as the “optimal” strategy quite by accident, and exploring affirmative interventions to mitigate that inadvertent discovery. The talk will conclude with some recent advances since the initial presentation of this work at the ACM ICAIF meeting in November 2022, at which it was awarded Best Paper.
Bio:
Dr. David Byrd is the Marvin H. Green Jr. Assistant Professor of Computer Science at Bowdoin College, where he has been named a J.P. Morgan AI Research Faculty Fellow, and helps organize the annual ACM International Conference on AI in Finance (ICAIF). He earned his Ph.D. in Computer Science from Georgia Tech, for which he co-authored the popular ABIDES simulation platform. Prior to grad school, he had a fifteen year career in the internet and telecom startup space. His research areas of interest include responsible AI autonomy, market dynamics and microstructure, and simulation of complex systems. He is currently focused on the inadvertent emergence of disruptive behaviors in autonomous trading agents.
Building an AI/ML Pipeline to Nowcast Intraday Bond Prices
Abstract:
In the fast-paced world of financial markets, precise and consistent intraday bond pricing is essential. This talk will explore the development of AI/ML pipelines for nowcasting bond prices, ensuring comprehensive coverage of a specified set of bonds at a given frequency. We will address the complexities posed by real-time, noisy data, and the need for rapid, reliable predictions across vast amounts of time series data, involving billions of ticks per day. We will focus on strategic approaches to managing and processing this data at scale, including data preprocessing, feature engineering, and model validation techniques.
Bio:
Camilo Oritz is Head of AI Finance Engineering at Bloomberg. In this role, he manages all pricing-related efforts in Bloomberg’s AI Engineering group. His team of machine learning and quant engineers applies AI-driven techniques to provide high quality prices on a large universe of bonds that perform well in the absence or presence of recent observations, liquidity, or volatility. Since joining Bloomberg in 2014, Camilo has led several AI teams focused on Communications, News, Editorial, and AI Platforms. Camilo earned his Ph.D. working on large-scale convex optimization at Georgia Tech, which he attended as a Fulbright Science & Technology Awardee.
Slides are not being distributed per speaker/company request
Registration
Early registration is available until Friday, August 30, 2024, after which regular registration rates will apply. The early (regular) registration rates are:
Corporate delegates: $200 ($250)
Academics, Alumni, & Non-Columbia students*: $75 ($100)
Current Columbia students*: $40 ($50)
No refund after August 30, 2024
*Those availing of student rates will be required to show a valid student ID at the event.