Quantitative Finance Paper Digest: Nov 8‑14 2025 Highlights

Diffolio: A Diffusion Model for Multivariate Probabilistic Financial Time‑Series Forecasting and Portfolio Construction

Link: https://arxiv.org/pdf/2511.07014v1

Authors: So‑Yoon Cho, Jin‑Young Kim, Kayoung Ban, Hyeng Keun Koo, Hyun‑Gyoon Kim

Proposes a diffusion‑based model that combines a denoising network with a hierarchical attention architecture at asset and market levels. Introduces a correlation‑guided regularizer based on stable estimation of the target correlation matrix to capture cross‑asset dependencies. The model extracts features from historical returns and from asset‑specific and systematic covariates. Experiments show superior multivariate prediction accuracy and portfolio performance, achieving higher Sharpe ratios and higher growth‑optimal portfolio equivalents compared with baseline probabilistic forecasting models.

Equilibrium Portfolio Selection under Utility‑Variance Analysis of Log Returns in Incomplete Markets

Link: https://arxiv.org/pdf/2511.05861v1

Authors: Yue Cao, Zongxia Liang, Sheng Wang, Xiang Yu

Studies time‑inconsistent portfolio selection in incompletely efficient markets by balancing expected utility of log returns against variance. Characterizes equilibrium strategies via a coupled system of backward stochastic differential equations (BSDE). Existence theory is proved for (i) independent Brownian motions (ρ=0) and (ii) bounded trading strategies. For correlated Brownian motions (ρ≠0), constructs an approximate time‑consistent equilibrium by perturbing the ρ=0 solution, with error O(ρ²). Provides numerical examples using deep‑learning algorithms.

Forecast‑to‑Fill: Benchmark‑Neutral Alpha and Billion‑Dollar Capacity in Gold Futures (2015‑2025)

Link: http://arxiv.org/pdf/2511.08571v1

Authors: Mainak Singha, Jose Aguilera‑Toste, Vinayak Lahiri

Tests simple interpretable state variables—trend and momentum—on gold futures (2,793 trading days, 2015‑2025) using a rolling 10‑year training window and 6‑month forward‑looking test. Converts smooth trend‑momentum signals into volatility‑targeted, friction‑aware positions via Kelly pricing with fractional impact adjustment and ATR‑based exits. Out‑of‑sample results: Sharpe ratio 2.88, maximum drawdown 0.52 %, linear cost 0.7 bps, impact term γ=0.02. Regression on spot gold returns yields annualized return ≈43 % and alpha 37 % (Sharpe = 2.88, IR = 2.09) under a 15 % volatility target, β≈0.03. Bootstrap confidence interval [2.49, 3.27] and SPA test (p = 0.000) confirm statistical significance and robustness to delay, reversal, and cost pressure.

Forecasting Implied Volatility Surface with Generative Diffusion Models

Link: https://arxiv.org/pdf/2511.07571v1

Authors: Chen Jin, Ankush Agarwal

Introduces a conditional denoising diffusion probabilistic model (DDPM) to generate arbitrage‑free implied volatility (IV) surfaces. Conditions on a rich set of market variables: EWMA of historical surfaces, underlying asset returns and squared returns, and scalar risk indicators such as VIX. Empirical results show significant superiority over leading GAN‑based models in reproducing stylized facts of IV dynamics. Addresses small arbitrage opportunities in training data by adding a standard arbitrage penalty to the loss, weighted by a non‑parametric SNR‑based scheme that dynamically adjusts penalty strength during diffusion. Provides formal analysis of the trade‑off and a convergence proof showing the penalty introduces a small, controllable bias while keeping the generated distribution close to real data.

FCOC: A Fractal‑Chaotic Co‑driven Framework for Financial Volatility Forecasting

Link: https://arxiv.org/pdf/2511.10365v1

Authors: Yilong Zeng, Boyan Tang, Xuanhao Ren, Sherry Zhefang Zhou, Jianghua Wu, Raymond Lee

Presents the Fractal‑Chaotic Oscillation Co‑driven (FCOC) framework, integrating a Fractal Feature Corrector (FFC) to extract high‑fidelity fractal signals and a biologically‑inspired Chaotic Oscillation Component (COC) that replaces static activations with a dynamic processing system. Empirical validation on S&P 500 and DJI data demonstrates substantial improvements over state‑of‑the‑art models such as Mamba and previously underperforming architectures like Transformers on risk‑sensitive metrics.