Program & Abstracts

9:00–9:30 · Session I

Controlling Risk and Preventing Manipulation in Prediction Markets

John R. Birge

Prediction markets offer opportunities for the hedging of a variety of risks that depend on the outcomes of future events. They have recently grown rapidly in volume from under $5B per month in September 2025 to $25B in April 2026. They are now the subject of debates over their legal and regulatory status. Ideally, prediction markets aggregate diffuse information and reduce levels of uncertainty, but they are subject to manipulation by those with superior (e.g., “insider”) information. This talk will describe how such manipulation risk arises inevitably in certain market designs and how to construct mechanisms that prevent manipulation while allowing for effective information aggregation.

9:30–10:00 · Session I

Use and Misuse of Quadratic Penalties in Portfolio Optimization

Dan diBartolomeo

Many users of optimization tools add quadratic penalties to tailor portfolio attributes that are not directly related to return and risk (a target income yield, an ESG score). Many people think of such penalties as “soft constraints” that are less costly to pure return and risk objectives than a hard constraint. In this presentation we will illustrate how the numerical properties of quadratic penalties are often misunderstood leading to economically counterproductive uses. It will be shown that adding one or more quadratic penalties to the optimization is mathematically equivalent to adding an extra orthogonal risk factor to the risk model. The “scale” of the penalty magnitude is equivalent to the variance of factor returns within a factor model divided by the investor risk tolerance. The use of such penalties is effectively broadening the definition of risk from the definition in Markowitz and Levy (1979) to the more arbitrary concern of “the risk of not having a portfolio that looks like I want it to look”. It can also be shown that quadratic penalties are closely related to the concept of Lagrangian multipliers in convex optimization. Such multipliers describe the sensitivity of the optimization objective to relaxation of a hard constraint. It is easily understood one way to enforce a constraint is to target a “goal” value between minimum and maximum of the hard constraint and then vary the scale of the quadratic penalty until the constraint is satisfied. While the definition of quadratic penalties on portfolio attributes can be flexible, the numeric magnitudes associated with penalties must be bounded by economic rationality. Our last topics will be use of asymmetric quadratic penalties, and the parallels between quadratic penalties and non-parametric methods of portfolio construction methods such as the Analytic Hierarchy Process.

10:00–10:30 · Session I

Measuring Risks in Hedge Funds: Evaluation and Usefulness of Exposure Data in Form PF

Stathis Tompaidis

We use data from a new regulatory filling for large hedge funds, Form PF, to evaluate the accuracy of reported stress test exposure information and to construct measures of hedge funds' aggregate risk. At the individual fund level, we find the exposure information has weak power to explain individual fund returns. However, aggregating the reported stress test results by groups or for the entire set of reporting funds improves the situation significantly. We discuss which stress tests provide useful information and which are mostly uninformative. We build two aggregate measures of risk, value-at-risk and expected shortfall, and decompose the contributions to these measures by fund strategy and fund liquidity. Our results can be used to improve the form and to monitor aggregate risk across reporting funds.

11:00–11:30 · Session II

A Unified Framework for Measuring Financial Crash Risk

Boris Klebanov

The measurement of financial crash risk, whether at the market level or for individual stocks, is a major challenge in asset pricing, risk management, and regulation. Skewness has been used as a proxy for crash risk in the finance literature, but it summarizes the entire distribution through its third moment and does not isolate the left tail of the return distribution, which is where crash risk manifests. In contrast, Higher Moment Coherent Risk (HMCR) measures, introduced by Krokhmal (2007), are tail-sensitive because they penalize losses exceeding a threshold.

We present a unified framework for measuring financial crash risk based on HMCR measures. These measures generalize CVaR to higher-order moments. They satisfy the coherency axioms and are compatible with second-order stochastic dominance. The framework unifies HMCR measures, the Risk Quadrangle, and distributionally robust optimization.

Building on the connection between HMCR and the Risk Quadrangle, the associated error functional enables regression-based estimation of HMCR, while a distributionally robust interpretation offers protection against model misspecification. We also present results on the relationship between HMCR and skewness, bounds on HMCR, and a crash premium ordering theorem established under empirically verifiable assumptions, providing a theoretical foundation for using HMCR in crash risk measuring and capital regulation.

11:30–12:00 · Session II

Pricing Guarantees in Western Critical-Minerals Supply Chains: A Contingent-Claims and Newsvendor Approach

Guilherme Larangeira

Western governments and firms are rebuilding critical-mineral supply chains outside China with guarantees: price floors, tariffs, strategic reserves, and take-or-pay offtake. How should these be priced? Price guarantees can in principle be replicated in the forward market, so Black-76 gives a risk-neutral (preference-free) price. Demand guarantees cannot: with no replicating portfolio, any price embeds a choice of preferences, of which risk-neutral is but one. We model a supplier and a buyer in a take-or-pay contract as a newsvendor-type problem with Mean-CVaR preferences, give the buyer's best response in closed form, and show when a coordinating wholesale price exists. Early-stage work, with a Brazil–France offtake (Viridis and Solvay) as the case; the talk emphasizes the problem formulation and a sketch of the solution.

12:00–12:30 · Session II

3SAT: A Beautiful but Untrainable Neural Network

Andrew Mullhaupt

A small neural network capable of solving its task exactly can be extraordinarily difficult to train, and we construct a family of such nets. The parameters that minimize the loss of the network are the satisfying assignments of any given 3-CNF, so training the net inherits the NP-complete difficulty of 3-SAT. We construct a family of 3-CNFs which demonstrate that the difficulty of training is realized even at comparatively small parameter dimensions.

We use this family to test possible improvements to neural net training on a “known difficult” class of nets, and show that a bandit algorithm can mitigate the “constant factor” in the cost of solving the problem.

12:45–1:30 · Lunch talk

AI-First Permanent Capital: Free Cash Flow, Option Premiums, and Compounding Ownership

Manish Sainani

How can an investment system use short-term market uncertainty to build long-term ownership in productive businesses? This talk examines that question through the development of Hushh Fund A, its Alphabets 27 strategy, and the Sell the Wall Alpha machine. The investment thesis begins with business economics: substantial absolute free cash flow (FCF), durable FCF growth, capital efficiency, and a price that supports an attractive ownership return. The framework evaluates FCF scale and growth separately, with attention to accounting comparability, cyclicality, and forecast uncertainty. It also informs the concentrated Ultra Alpha Bets 10 and broader Ultimate Alpha Bets 69 portfolio concepts. Around selected holdings, the options overlay seeks to collect volatility risk premia and benefit from time decay through covered calls during strength and cash-secured puts during weakness. Relative Strength Index (RSI), trend, implied versus realized volatility, liquidity, and open-interest concentrations inform candidate trades. These signals do not establish mispricing or guarantee reversal. Net realized income, after losses, costs, applicable taxes, and required reserves, is intended for reinvestment into additional shares. The talk draws practical lessons from early versions: premium received is not economic profit; rolling an option does not erase a loss; covered calls can surrender shares and cap upside; and put assignments can increase exposure during falling markets. Share accumulation must therefore be assessed alongside total return, drawdown, and capital at risk. I will outline an AI-assisted research and risk architecture with traceable inputs, reproducible calculations, explicit exposure limits, and human oversight. The evaluation agenda includes walk-forward testing, realistic execution costs, comparison with the same holdings without options, and tail-risk measures such as expected shortfall. Tests must explicitly address survivorship bias, look-ahead bias, changing market regimes, and sensitivity to portfolio construction assumptions. I will close with opportunities for academic collaboration on determining when this approach can improve compounding ownership.

1:30–2:00 · Session III

Sequential Optimization of Dynamically Augmented CVaR

Eugene Feinberg

The presentation examines risk-sensitive control of finite-state, finite-action Markov decision processes with discounted cumulative costs. It develops Dynamically augmented CVaR (DCVaR), together with an optimization algorithm. A robust formulation incorporates a tail-risk parameter into the state. Earlier value-iteration methods for this augmented model, initially used to estimate optimal static CVaR, instead yield lower bounds. DCVaR provides a time-consistent risk criterion below static CVaR and also admits an interpretation related to nested CVaR. The algorithm is justified through analysis of a mass-transfer problem.

2:00–2:30 · Session III

Convergence Rate of a Functional Learning Method for Contextual Stochastic Optimization

Noel Smith, Andrzej Ruszczyński

We consider a stochastic optimization problem involving two random variables: a context variable $X$ and a dependent variable $Y$. The objective is to minimize the expected value of a nonlinear loss functional applied to the conditional expectation $\mathbb{E}[f(X, Y,\beta) \mid X]$, where $f$ is a nonlinear function and $\beta$ represents the decision variables. We focus on the practically important setting in which direct sampling from the conditional distribution of $Y \mid X$ is infeasible, and only a stream of i.i.d.\ observation pairs $\{(X^k, Y^k)\}_{k=0,1,2,\ldots}$ is available. In our approach, the conditional expectation is approximated within a prespecified parametric function class. We analyze a simultaneous learning-and-optimization algorithm that jointly estimates the conditional expectation and optimizes the outer objective. Using a specially designed measure of non-optimality, combining the squared norm of the objective function's gradient and the mean square error of the auxiliary parametric model, we establish that the method achieves a convergence rate of order$\mathcal{O}\big(1/\sqrt{N}\big)$, where $N$ denotes the number of observed pairs.

2:30–3:00 · Session III

Rethinking Systematic Risk. Drawdown-Based Risk Measures and the Low-Beta Anomaly

Piotr Szczepocki, Ewa Feder-Sempach, Stan Uryasev

University of Lodz, Faculty of Economics and Sociology, Department of Statistical Methods, Lodz, Poland University of Lodz, Faculty of Economics and Sociology, Department of International Economics, Lodz, Poland Stony Brook University, Department of Applied Mathematics and Statistics, Stony Brook, NY 11794-3600, USA

Purpose: The main objective of the study is to examine long-short investment strategies based on various beta measures. Stocks are sorted into beta deciles. A strategy that takes a long position in the lowest-beta portfolio (D1) and a short position in the highest-beta portfolio (D10) is then constructed. Its performance is evaluated using alternative risk measures, considering returns, volatility, and drawdowns. The strategy assumes unrestricted access to borrowing to leverage low-beta assets.

Methodology: To implement the strategy, we calculate five beta coefficients, two of which are based on market drawdowns. Specifically, we calculate betas based on (i) standard deviation, (ii) lower semi deviation, (iii) mean absolute deviation, (iv) Conditional Value-at-Risk, and (v) drawdowns, namely Conditional Drawdown-at-Risk (CDaR) and Expected Regret of Drawdown (ERoD). The empirical analysis is based on S&P 500 stocks over the 1995–2024 period. For each risk measure, stocks are sorted into deciles according to their estimated beta coefficients. The Security Market Line is estimated for the full 30-year sample as well as for 15, 10, 5, 3, and 1-year periods. The resulting portfolios are compared in terms of annualized returns, volatility, maximum drawdown, and the Sharpe ratio.

Findings: The low-beta anomaly is present across nearly all alternative beta measures. In most cases, portfolios of low-beta stocks outperform high-beta portfolios, resulting in profitable long–short (D1–D10) strategies. The time horizon used to evaluate betas plays a critical role. Over shorter investment horizons, CDaR and ERoD beta measures often provide better results than the standard beta. Moreover, the empirical Security Market Line exhibits a predominantly negative slope, indicating that lower-beta stocks earn higher average returns than higher-beta stocks, which indicates that the borrowing rate is higher than the expected market return. The overall results suggest that alternative risk measures, particularly those based on drawdowns, provide a useful framework for identifying portfolios with good risk-adjusted performance. Furthermore, portfolios based on CDaR and ERoD beta measures experience smaller maximum drawdowns, making them particularly suitable for capital preservation strategies by providing enhanced downside protection.

Originality: Unlike the standard CAPM beta, which measures systematic risk using standard deviation, the drawdown beta framework better captures asset sensitivity to market risk. CDaR and ERoD betas, result in profitable long–short (D1–D10) investment strategies and improved capital preservation over shorter investment horizons. The findings suggest that investors may benefit from incorporating drawdown-based beta measures into portfolio construction and risk management decisions.

Keywords: investment strategies, CAPM, drawdown beta, low-beta anomaly

3:00–3:30 · Session III

Dynamic Inverse Portfolio Optimization with Mean-Deviation Preferences

Volodymyr Volchenko

Inverse portfolio optimization aims to identify an investor’s risk preferences from observed trading decisions. The existing mean-deviation approach to this problem considers a single observed optimal portfolio. This talk examines a dynamic inverse portfolio problem based on a sequence of portfolio rebalancing decisions. At each time, the investor trades from the position inherited from the preceding date and incurs proportional and quadratic market-impact costs. The decisions are therefore linked across time: the optimality condition for the current position depends on both the current trade and the expected cost of the next rebalancing. Risk preferences are represented by time-varying mixtures of CVaR deviations at several tail-probability levels. For a finite scenario tree, the inverse problem reduces to a system of linear equalities and inequalities that characterizes the set of preference parameters and provides conditions for their uniqueness.

4:00–4:30 · Session IV

Efficient Multivariate Kelly Optimization Reveals Sigmoidal Scaling Laws

Ruslan Tepelyan

For a sequence of binary bets, the Kelly criterion provides a closed-form solution that maximizes the expected growth rate of wealth. In contrast, when multiple bets are placed simultaneously (e.g., in portfolio allocation or prediction markets), the optimal Kelly strategy generally requires numerical optimization over a joint outcome space. A naive formulation scales exponentially in the number of bets, requiring 2^N time and memory for N simultaneous wagers, which restricts existing methods to small problem sizes.

We present two complementary methods that dramatically extend the scale of multivariate Kelly problems that can be solved. First, in the case of independent bets, we introduce an integral transform formulation that eliminates explicit enumeration of outcomes, reducing the computational complexity of evaluating the objective from O(2^N) to O(N). Combined with numerically stable quadrature, this enables accurate solutions for problems involving hundreds of bets. Second, we develop a decomposition-based approach that constructs and solves carefully chosen subproblems, yielding feasible lower bounds and infeasible upper bounds on the optimal growth rate. This provides a practical mechanism for quantifying worst-case suboptimality as a function of subproblem size.

Together, these methods make it possible to study the large-N regime of the multivariate Kelly problem. Using synthetic data inspired by prediction markets, we show that the relationship between subproblem size and solution accuracy follows a simple and highly regular scaling law. In particular, the shortfall ratio between the lower and upper bounds is well-approximated by a sigmoid function of the relative subproblem size, with parameters that can be predicted from low-dimensional summary statistics of the problem.

4:30–5:00 · Session IV

Conditional Joint Riemannian Flow Matching and its application to time-varying dependence structure in Financial Returns

Jeongyeon Park, Aaron Young Shin Kim

Jeongyeon Park — Department of Applied Mathematics and Statistics, Stony Brook University. Aaron Young Shin Kim — College of Business, Stony Brook University

We propose a Conditional Joint Riemannian Flow Matching (CJRFM) model for multivariate financial returns. Standardized returns are decomposed into a directional component on the unit sphere and a log-radius component in Euclidean space, and the conditional distribution of future returns is learned from recent market history via a neural flow on Sᵈ⁻ × R. Conditioning lets the model adapt both the scale and the cross-asset dependence structure of future returns to recent conditions. In a rolling-window backtest on a 17-asset universe, CJRFM outperforms Gaussian, NTS, DCC-GARCH, NTS-GARCH, equal-weight, and unconditional (RFM) benchmarks in annualized return, Sharpe ratio, and reward-to-tail-risk measures, highlighting the value of modeling time-varying return distributions.

5:00–5:30 · Session IV

Loss-Versus-Rebalancing for Multi-Asset Geometric-Mean Market Makers

Joonhyuk Suh,, Stan Uryasev

Many decentralized exchanges use automated market makers (AMMs), which execute trades against liquidity pools according to deterministic reserve-based rules. When external prices move, arbitrageurs realign the pool with the market, transferring value from liquidity providers - an adverse-selection cost known as loss-versus-rebalancing (LVR). Milionis et al. establish a general LVR identity for such pools, expressed through the Hessian of the pool value function and the instantaneous covariance matrix of asset returns. Specializing this identity to Balancer-style multi-asset automated market makers with a weighted geometric-mean invariant, we show it collapses into an explicit pairwise form: instantaneous LVR equals half the pool value times a weighted sum, over every pair of assets, of the quadratic variation of that pair's return difference. Multi-asset LVR is thus driven by cross-asset return dispersion rather than marginal variance alone, and vanishes exactly when all assets move together. We derive an exact finite-grid benchmark gap for the Balancer invariant, connect it to covariance-based predictors via a second-order Taylor expansion, and show it converges to the continuous-time identity as the observation grid is refined. We validate these theoretical predictions empirically in synthetic multi-asset cryptocurrency pools.

5:30–6:00 · Session IV

Your Hypergradient is Skewed: Antithetic Neumann Estimation for Bilevel Optimization

Pawel Polak Joint work with Jason Bohne, Gary Kazantsev, and David S. Rosenberg (NeurIPS 2026)

Bilevel optimization is a core tool in practical machine learning workflows, including hyperparameter tuning, data reweighting, robust learning, reinforcement learning, and imbalanced classification. Hypergradient computation requires the inverse of the inner Hessian, typically estimated via a truncated Neumann series. While existing analyses focus on bias and variance, we show that the product structure of the Neumann series introduces a third source of error: skewness. Unlike ordinary gradients, whose asymmetry is suppressed by mini-batch averaging, the Neumann estimator amplifies skewness across factors, with the third-moment error growing with truncation depth, inner condition number, and per-factor noise asymmetry. This arises in many practical settings such as corrupted labels, outliers, heterogeneous, or class-imbalanced datasets.

To address this, we introduce the Antithetic Neumann Estimator (ANE), which pairs each Hessian sample with a mirrored counterpart around a reference Hessian and averages the resulting inverse-Hessian estimates. With an exact reference, the antithetic product eliminates all skewness; with an approximate reference, residual skewness scales linearly with reference error. ANE requires only a modification of the Hessian-vector step and integrates with existing stochastic or variance-reduced bilevel methods. Across robust regression, corrupted-label learning, offline reinforcement learning, and imbalanced classification, ANE-based methods achieve strictly better bias–variance–skewness tradeoffs, reducing loss by up to 24% and improving predictive metrics by up to 27%.

9:00 AM–9:40 AM · Students Day

Title to be announced

Michael Tiano

Schonfeld

We will discuss the current techniques and areas of research focused on by quantitative and tactical portfolio managers. This includes approaches to signal research, monetization methods, risk modeling, portfolio construction, transaction cost modeling and execution. We outline how these approaches can differ by strategy, asset class, and investment horizon. Where appropriate we will try and highlight for students which of these areas tends to be most fruitful for job opportunities as well as important skill sets to develop.

9:45 AM–10:25 AM · Students Day

Entering a Career in Quantitative Finance

Konstantin Kalinchenko

Millennium Management

We will discuss the landscape of options presented to graduate students seeking to enter the financial industry as quantitative professionals. We will also cover the technical and soft skills expected from candidates, the tactics for maximizing success, and some concrete steps a student pursuing a career in quantitative finance should take in order to stand out in the pool of candidates.

10:50 AM–11:30 AM · Students Day

Finding your first job in finance

Chandrasekhar Thammisetty

Flagstar Bank

In this seminar, Chandrasekhar Thammisetty will share his personal job search journey, offering practical advice for graduate students navigating the competitive world of finance. The session will focus on how to effectively network, utilize professors as mentors and connectors, and engage with financial societies in New York to build valuable relationships. Chandrasekhar will also discuss how cold emails can open doors to job opportunities, strategies for managing technical interviews and tips for making lasting impressions in conversations with HR. Additionally, the seminar will provide insights on preparing for technical interviews, with actionable strategies for securing your first role in the financial industry. Whether you're a domestic or international student, this talk will equip you with the tools and strategies to successfully transition from academia to the professional world.

11:35 AM–12:15 PM · Students Day

Retrieval, Agents, and Evaluation: Building LLM Systems Where Being Wrong Has Consequences

Peter Tsyurmasto

Axiom Intelligence Systems

Large language models have moved quickly from text generation to systems that retrieve evidence, reason over it, and produce documents. This talk surveys emerging best practice in building such systems and the challenges that remain open, using pharmaceutical regulatory affairs as a lens, because that setting is unforgiving: a response to a health authority is a formal commitment, every claim must trace to a source, an omission is as serious as an error, and an accountable human signs.

The first part covers retrieval-augmented generation: why grounding answers in documents became the dominant pattern, what is known about doing it well, and why retrieval quality bounds everything downstream and degrades as collections grow. The second part turns to agentic systems, where a task is broken into cooperating steps rather than answered in one pass, and to the trade-offs that decomposition introduces.

The third part concerns evaluation, which in regulated settings is infrastructure rather than reporting. I review current practice, from RAGAS-style reference metrics to LLM-as-judge, and the difficulties each faces: benchmarks must be curated by domain experts rather than scraped, completeness is harder to measure than correctness, failures must be attributed to a step rather than to the system as a whole, and the automated judge must itself be validated against human judgment before it can be trusted.

1:15 PM–1:55 PM · Students Day

Title to be announced

Robert Frey

Stony Brook University

Abstract to be announced.

1:30 PM–2:10 PM · Students Day

Can an AI Be Your Realtor? How AI Enters a Traditional Industry, and What Big Data Realty Is Building

Vladimir Bugera

Big Data Realty

Residential real estate is a $100 billion-a-year commission industry built on a single unit: the independent realtor, who does the whole job and is paid a percentage. Incumbent AI has gone where that realtor is paid, into search, content and lead conversion; none of it prices risk or is scored against the outcome. This talk uses real estate as a case study in how AI actually enters a traditional, licensed industry: not by replacing the professional, but by decomposing the job into the computable half (valuation as a distribution, list-price and bid strategy with the tail priced, agent selection, timing) and the licensed half (walking the property, the negotiation table, the closing), and rebuilding the business model around that split.

The first part covers how the US market works and what changed in 2024, when the NAR settlement made choosing an agent a priced decision. The second part shows the quant toolkit at work on real data from Miami-Dade and San Francisco: why “the value of a property” is four different numbers for four different decision makers, why the sale price is the only thing in the data and everything else must be inferred or chosen, and how a risk-constrained bid (maximize expected surplus subject to a CVaR constraint on overpaying) turns a valuation distribution into a decision a client can act on. The third part is what Big Data Realty, a licensed brokerage on two coasts, has built on top of that: public and MLS data in one schema, AI pipelines that analyze markets at scale, and a decision log that scores every recommendation against the outcome. The talk closes with what this looks like as a career: the skills that transfer from quantitative finance to a traditional industry, the ones that do not, and the open problems where students can contribute.