In briefCorrFlow: Exact, permutation-equivariant generation of valid correlation matrices via flow matching
SummaryThis paper presents CorrFlow, a method for generating correlation matrices that separates keeping outputs valid from learning the data distribution. A generalized Fisher transform maps correlation matrices to unconstrained coordinates, where a graph flow-matching network learns the empirical distribution and is permutation equivariant. On 80×80 Taiwan equity matrices, CorrFlow produced 100% valid draws without repair and a lower feature sliced-Wasserstein distance than CorrGAN and image diffusion. The authors note the results do not establish superiority over fully tuned image models or isolate each component's contribution.
In briefSelecting Sparse Features for Financial Regression Using Monte Carlo Tree Search
SummaryThis paper proposes a feature selection method for high-dimensional financial regression that uses Monte Carlo Tree Search with Generalized Nested Rollout Policy Adaptation. Candidate feature subsets are scored by out-of-sample validation performance rather than training error. On a task explaining weekly NASDAQ-100 returns from 1,072 engineered features, the method reaches an out-of-sample R² of 0.6961, compared with 0.6313 for an optimized Lasso baseline.
In briefRealized Tail-Risk Expectile Models for Stable, Accurate VaR and ES Forecasts
SummaryThis paper proposes a realized tail-risk conditional autoregressive expectile (RTCARE) model for jointly forecasting value-at-risk (VaR) and expected shortfall (ES). It adds realized tail measures, such as realized quantile and realized ES built from intraday high-frequency returns, to the dynamics and measurement equations of the conditional expectile. Parameters are estimated by Markov chain Monte Carlo, and forecasts are assessed in numerical examples and an empirical study using U.S. stock returns. The RTCARE models perform competitively and often better than existing alternatives across the tail levels considered, with stable results across Monte Carlo replications and rolling forecast evaluations.
In briefEarly warning of pension sponsor distress using public filings and dynamic investment limits
SummaryThe paper presents a transparent early-warning pipeline built on public data, centred on a Sponsor Vulnerability Index (SVI) that uses SEC EDGAR balance-sheet fundamentals and two macro series. On 36 U.S. sponsors (1,418 firm-quarters), it reaches an out-of-fold AUC-ROC of 0.887 under leave-one-firm-out validation. It also adds a stress-augmented Monte Carlo VaR and a self-investment cap that falls as the SVI signals distress, and reports the method's failure modes.
In briefAdding transaction-text behavioural features to credit risk scorecards
SummaryThis paper examines whether payment narratives in transaction data can serve as an alternative source for credit risk modelling. Using natural language processing and regular expressions, the authors convert unstructured transaction descriptions into interpretable spending categories with a newly developed classification algorithm. Adding these features to a scorecard significantly increased its Gini coefficient compared with a scorecard without them.
In briefEntropy-regularized dynamic portfolio selection under no-shorting constraints
SummaryThis paper studies entropy regularization for discrete-time dynamic mean-variance portfolio selection with no-shorting constraints. It introduces a state-dependent temperature matched to the problem's homogeneity, which preserves the piecewise-quadratic value function and finite-dimensional Bellman recursion. The authors prove that a sufficiently small common temperature keeps the Bellman structure and convexity over the whole finite horizon, and the original solution is recovered as the temperature vanishes.
In briefUsing a supervised neural network to initialize reinforcement learning for timing exits from SPX short-straddle options
SummaryThis paper studies when to exit SPX short-straddle option portfolios using two stages. A multilayer perceptron is first trained on ex-post optimal hold or exit labels from historical data, then used to initialize reinforcement-learning actors fine-tuned with a risk-aware trading reward. Out-of-sample tests over 2024 to 2026 show the supervised MLP substantially outperforms passive holding and fixed-time or profit-and-loss exit rules.
Jianning Wang, Robert \'Slepaczuk, Pawe{\l} Sakowski
In briefFLEX Options and Market Outcomes: How Customized Exchange-Traded Options Affect Regular Options and Stocks
SummaryThis paper examines whether activity in FLEX options, exchange-cleared contracts with customizable strikes, expirations, exercise styles and settlement terms, leaves measurable effects on regular listed options and underlying equities. Using data from March 2025 to March 2026 covering 41 underlyings and 53 event dates, the authors find a significant post-event rise in listed-option open interest and show FLEX activity concentrates in high implied-volatility environments. They conclude FLEX activity mainly serves as an observable marker of institutional volatility demand and open-interest positioning rather than a predictor of stock returns.
Shivesh Prakash, Frank Ferstler, Nicholas Pezolano
In briefHow AI shopping assistants affect retail revenue and product range choices
SummaryThis paper studies how a retailer should choose its product range when AI shopping assistants shape which items customers consider. Ignoring AI mediation is shown to cause average revenue losses of 2.52% to 11.37% relative to the Plackett-Luce optimum, while a simple deterministic approximation based on fitted AI utility orderings loses at most 0.53% on average. The authors also prove that, for a fixed number of customer segments, the optimization problem admits a fully polynomial-time approximation scheme.
In briefFunding Debt Relief: How Intermediary Liquidity Shapes Mortgage Buyouts
SummaryThis paper studies how intermediary liquidity affects mortgage resolution, using 1.9 million Ginnie Mae buyout options from 2019 to 2022. When funding markets froze in 2020, nonbank issuers repurchased far fewer delinquent loans than depositories servicing loans in the same pools, and shifted away from large balances. About $19 billion of repurchases were delayed, and most were completed later.
In briefA distributional framework for characterizing and managing risk in large language models by measuring how their repeated responses vary.
SummaryThe paper notes that large language models are inherently stochastic, so repeated runs under identical conditions can yield different responses. It proposes a framework applicable across the model lifecycle, estimating empirical response distributions via Monte Carlo simulation and studying how they shift under semantic interventions. The authors illustrate it with borrower classification and music album recommendation tasks, linking it to model risk management by measuring robustness and control effectiveness. Because it relies only on observed responses, it can also be applied to black-box models accessed through commercial APIs.
