In briefRADAR: Retrieval-Augmented Diffusion for Portfolio Optimization with Stochastic Discount Factors
SummaryThis paper studies portfolio optimization within the stochastic discount factor framework by learning market state representations that capture underlying financial risk structures. The authors propose RADAR, which retrieves similar historical market regimes to build context-dependent noise distributions and applies conditional diffusion to denoise multimodal inputs. The authors report state-of-the-art results on key risk-adjusted metrics and economically meaningful signals for asset returns and correlations.
In briefOptimal Liquidation Near Support and Resistance under Multi-Skew Brownian Motion
SummaryThe paper solves a perpetual liquidation problem for a geometric multi-skew Brownian motion with local-time pushes at a support level and a resistance level, a technical-analysis-style model that depends only on price. A closed-form criterion yields three geometries: the continuation band lies below resistance, straddles it, or has its upper edge at resistance. Selling into resistance emerges as a conclusion rather than an assumption. Support parameters are identified exactly from exercise behaviour, while resistance strength is identifiable only within an interval with a closed-form endpoint.
In briefDeep kernel hedging: a paper combining neural features with kernel methods for derivative hedging
SummaryThis paper introduces a deep kernel hedging framework in which the hedging function lies in a reproducing kernel Hilbert space whose kernel is parameterized by a neural network embedding, and it handles path-dependent inputs via truncated time-augmented signature features. The authors derive a generalized representer theorem that reduces the empirical optimization to a finite-dimensional problem, and they use scalable random Fourier feature approximations with convergence guarantees. Experiments on synthetic and real data reportedly show competitive and robust hedging performance, especially in low-data regimes.
In briefFrom Word Counts to Context: Topic Models for Asset Pricing
SummaryThis paper compares a sentence transformer with k-means clustering against Latent Dirichlet Allocation (LDA) for generating topics from 394,661 news articles used in a financial portfolio construction pipeline. Topic coherence is measured by NPMI and financial performance by the Sharpe ratio. The sentence transformer approach shows higher observed scores on both measures, though the authors note the tests do not establish outperformance. They suggest stricter tests using only information available at each date, and broader datasets, are needed.
Kevin Foley, Jonathan Hartadi, Shivesh Prakash et al.
In briefRICE-Alpha: LLM Agent Stock Forecasting with Event Graphs and Reliability Correction
SummaryThis paper presents RICE-Alpha, a point-in-time stock-scoring framework that uses historical news events to forecast stock performance. It separates a multi-view base forecast from a reliability-calibrated residual correction based on how past events continued, using event graphs to preserve issuer-specific chronology. Tested on daily Nasdaq-100 and Hang Seng Index data from 2024 to 2026, it reportedly outperformed the evaluated LLM-based agents and momentum baselines. The authors conclude that historical event continuation adds information when it is temporally grounded, reliability-calibrated, and applied as a residual correction.
In briefOracle-parametrized AMMs: pricing rules that use external price feeds
SummaryThis paper introduces oracle-parametrized automated market makers (OP-AMMs), whose quoted prices depend on both pool reserves and an external oracle price, suited to settings like tokenized securities where price discovery happens off-chain. Under a strict oracle-contraction condition, the authors show the quoted price lies between the oracle price and an implicit autarkic price set by the reserves, and derive a loss-versus-rebalancing decomposition. The analysis extends to stale and discrete-update oracles and sandwich attacks, and a backtest on one-second SPY NBBO data illustrates the trade-offs.
In briefControlling Option Greeks: Portfolios with Built-in Hedging Bias
SummaryThis paper presents an end-to-end deep learning framework for systematic options trading that embeds hedging behavior by explicitly controlling portfolio-level risk exposures. It adds a differentiable risk-sensitivity penalty to the training objective to keep portfolios neutral on selected risk dimensions, trained on historical data. Tests on Nasdaq 100 equity options report that a well-calibrated penalty improves out-of-sample risk-adjusted performance and reduces realized directional exposure compared with an unregularized baseline.
