In briefNon-perturbative option pricing for rough volatility using quantum-field-theory effective action
SummaryThe paper proposes a non-perturbative framework for stochastic-volatility option pricing based on the two-particle-irreducible effective action and Dyson-Schwinger gap equations. It approximates the joint law of state variables with a self-consistent Gaussian and handles both Markovian and rough (Volterra) models, including nonzero correlation. Reported benchmarks show close agreement with Monte Carlo references and comparisons against Hagan's expansion and rough SABR.
In briefEfficient frontier computed with a standard LASSO solver
SummaryThis note numerically checks Schmelzer and Hastie's result that Markowitz's Critical Line Algorithm and the LASSO path trace the same curve, using scikit-learn's lars_path with no portfolio software. It tests long-short portfolios under a leverage cap, fixed leverage with varying risk appetite, and the classical long-only fully invested frontier, and reports agreement with independent quadratic-programming solvers. The script, checks, Fama-French returns and an R cross-check are published so the results can be reproduced.
In briefSeparating liquidity-driven from strategic loan defaults using deep learning
SummaryThis paper applies a deep neural network to a proprietary commercial mortgage panel from 1998 to 2021 to distinguish liquidity-constrained default from strategic default, two explanations for delinquency that are hard to tell apart with standard models. Among delinquent loans, property net operating income is found to matter more for default rates than loan-to-value. The estimated thresholds remain stable across the 2008 financial crisis and the COVID-19 shock in 2020 and 2021.
In briefA spread-gated Hawkes model of best bid and ask dynamics, applied to limit order sizing
SummaryThe paper models the joint dynamics of the best bid and ask prices with a spread-gated Hawkes-flocking process that distinguishes four types of best-quote movements and includes a spread-dependent cross-side excitation term. The authors prove non-explosion on finite horizons, provide an O(N) recursive likelihood, and validate maximum likelihood estimation by simulation. On intraday data for INTC and MSFT, removing the cross-side term is rejected and the full model fits substantially better by AIC and BIC. As an application, they derive a closed-form optimal size for a single-period limit order placed at the best or second-best quote.
In briefWhen Hedging Changes the Payoff: Option Replication Under Price Impact and Trading Costs
SummaryThe paper studies how hedging a derivative by trading the underlying asset can change the payoff it was meant to replicate, once trading causes price impact and execution costs. In a binomial model replication is characterized by a fixed-point equation, and in continuous time a nonlinear pricing PDE is derived. For monotone convex Lipschitz payoffs such as calls and puts, exact replication is established under midpoint execution costs, and numerical experiments illustrate the effects.
In briefComparing two ways to split tail risk among components
SummaryThe paper compares two proportional measures of how much each risk component contributes to aggregate tail risk: the share of tail risk capital allocated under Conditional Tail Expectation (CTE), and the expected realized risk share under Geometric Tail Expectation (GTE). It establishes when the two allocations agree, how they are ordered when they differ, and how they separate asymptotically in the extreme tail. The difference is shown to be a normalized tail covariance.</summary>
In briefCZAR Loss: A Loss Function Designed for Financial Log-Return Prediction
SummaryThe paper argues that standard symmetric regression losses, such as mean squared and mean absolute error, are poorly aligned with financial return prediction, since they make a constant zero forecast nearly optimal and penalize models with weak directional skill. It introduces the CZAR (Composite Zero-Agnostic Return) loss, a piecewise quadratic function that is convex in the prediction and has closed-form gradient and Hessian for use in gradient-boosted libraries. In idealized tests and a LightGBM experiment on intraday BTC log-returns, the authors report reduced zero-returns bias and better directional accuracy on large-magnitude returns.
Joel Pfeffer (1), J. M. Diederik Kruijssen (1), Florian Stecker (1) et al.
In briefFinAutoRubric: Expert-guided automatic rubric generation for evaluating financial research agents
SummaryThis paper presents FinAutoRubric, a framework for evaluating finance research agents. Experts specify reusable evaluation guidance, and agents and code use it to generate, review, and validate query-specific rubrics, with reusable criteria stored in a Task Bank. On three expert-authored finance benchmarks, its rubrics match expert scoring about as closely as the strongest compared generator, agree with human grading, and were preferred by in-house analysts in blind review. The authors also release the 100-query FinAutoRubric Benchmark.
In briefRepresentation risk: pretrained image encoders can change predictive conclusions
SummaryThe paper argues that the pretrained image encoder used to turn pictures into features is a source of model uncertainty, which the authors call representation risk. Ten frozen encoders were compared across tasks including house prices, racehorse performance, histology, chest radiographs, facial age, and rice disease, with the encoder chosen on locked validation data. The results show no single encoder is best everywhere, and the authors recommend benchmarking candidates, selecting on validation data, and reporting paired and split-level uncertainty. The workflow is implemented in the LOOKAGAIN-ML software package.
WhyIntroduces representation risk and compares ten frozen encoders across six applications, showing large validation-selected performance gaps.
Ardyn Nordstrom, Morgan Nordstrom, Vamuyan Sesay et al.
In briefTesting whether a Pepsi–Coca-Cola pairs trade still works after its original fit period.
SummaryThis paper tests whether a pairs trading strategy between PepsiCo and The Coca-Cola Company, based on cointegration, is statistically robust and economically exploitable. The authors estimate the spread's mean-reversion parameters over 2013–2018, fix them, and optimise a threshold trading strategy in-sample over 2018–2023. The strategy is assessed with transaction-cost and parameter sensitivity tests, walk-forward validation, and adjusted and deflated Sharpe ratios, then evaluated out-of-sample from 2023 onward. The results show that weakening mean-reversion in the spread undermines the strategy's out-of-sample effectiveness.
