In briefSpouse-Protected Tontines: A Retirement Pool That Keeps Paying the Surviving Spouse
SummaryThis paper proposes a tontine design in which the first death in a household does not transfer money to the pool; instead, the account continues to pay the surviving spouse. The authors set up a multidimensional withdrawal and rebalancing problem that trades off expected real payments against the CVaR of terminal shortfalls, solved with a neural-network policy parameterization. Numerical experiments calibrated to Australia show the spouse-only phase is a first-order, persistent household event, and that book-level diversification removes most of the excess upper-tail cost of spouse continuation.
In briefQuantCode Model: Language Models for Executable Trading Code
SummaryThis paper studies how to adapt large language models to turn natural-language trading strategy descriptions into executable Backtrader code. The authors compare continued pretraining on trading-framework code with supervised fine-tuning on agent-validated request-to-code pairs, evaluated on the 400-task QuantCode-Bench. Supervised fine-tuning gave the larger gain, raising Judge Pass for Qwen3.6-35B-A3B to 58.2%. The authors also report that domain specialization can degrade structured tool-calling ability, which targeted recovery training partly restores.|
In briefRank Collapse of the Predictor Jacobian in Decision-Focused Learning
SummaryThis paper studies how a rank restriction in the predictor Jacobian limits the parameter-update directions available in decision-focused learning (DFL). Rank-one Jacobians make per-example gradients nearly collinear, and the authors show that such local statements do not imply common minimizers or collinear batch updates. In experiments, DFL gains over MSE stayed below 1.8% in 38 equity configurations, while full-capacity SPO+ reduced regret in shortest-path and knapsack tasks, with only the knapsack result surviving multiple-comparison correction.
In briefStochastic Knothe-Rosenblatt: fast calibration of stochastic local volatility models
SummaryThe paper notes that European option smiles pin down an asset's risk-neutral marginal distributions but not how they are linked across maturities. It proposes a modular calibration overlay for existing stochastic and path-dependent volatility models, building a martingale that exactly matches all prescribed marginals while staying as close as possible to the reference dynamics in an adapted Knothe-Rosenblatt sense. The authors cover numerical implementation, convergence to the SLV model, benchmarks on Heston and Bergomi dynamics, and a multi-asset extension.
Mathias Beiglb\"ock, Manuel Hasenbichler, Gudmund Pammer
In briefMulti-period bond portfolio optimization with transaction costs via a Markov Decision Process
SummaryThe paper addresses how bank treasury portfolios should allocate across bonds of different maturities under interest-rate risk, noting that static allocations can suffer large losses when rates rise. It proposes a simulation-based multi-period framework that models yield curves with a Dynamic Nelson-Siegel parameterization and vector autoregressive factor dynamics, approximated by a discrete-state Markov chain. Within a finite-horizon Markov Decision Process with proportional rebalancing costs, the investor maximizes expected terminal wealth, with the optimal policy found by backward induction. The authors also quantify the error from truncating the transition kernel, finding that mean terminal wealth is barely affected while drawdown and tail statistics are substantially distorted.
In briefSay, Echo, Do: Strategic Narratives and Revealed Positioning in Financial Markets
SummaryThis paper studies how institutional statements (Say), media repetition (Echo), and actual trading positions (Do) relate in financial markets, asking when text signals should be followed or faded. A linear-quadratic model identifies when an informed institution talks an asset down while buying it. The authors propose methods to detect echo articles, learn return-aligned embeddings, and measure who moved first. In simulated markets with known ground truth, echo sentiment predicted returns with a significantly negative sign in all 29 markets.
In briefFrom Intraday Orderbook to Imbalance Price: How Cross-Market Trading Shapes Balancing Prices
SummaryAs variable renewables make power scheduling more uncertain, this paper uses probabilistic modeling to study how intraday orderbook information reflects later imbalance price formation in Germany and Austria. It compares orderbook representations such as VWAP and last mid price, including neighboring products and products from the neighboring country. The authors find VWAP with neighboring products performs best in both countries, while adding fundamental features reduces test loss in Germany but increases it in Austria.
In briefA Generalized Langevin Model of Latent Liquidity and Concave Price Impact
SummaryThe paper models market impact as the response to submitted order flow net of counterflow from latent traders, who become active when price moves exceed their individual thresholds. Order flow depletes this latent pool, whose recovery follows a generalized Langevin equation with memory over several time scales. For an undepleted pool, the model yields an intermediate square-root regime between linear small- and large-order limits without assuming a square-root law directly. Numerical experiments show that depletion narrows this regime, and calibration to market data is deferred to a companion paper.
In briefInformation Games: How Rivalry Reshapes Firm Positioning in Language-Model Space
SummaryThe paper introduces ESCAPE, a rational-share game in which firms with different capability costs position themselves in a language-model representation space, and proves a unique equilibrium. It derives an exact reallocation formula separating opportunity effects from crowding effects, showing that competition need not push firms apart. Rivalry mainly changes how differently firms respond to the same opportunity shift. Corporate news data show that peer structure is persistent but changes over time.
