In briefAlphaRLM: Training Language Models to Build Market Factors from Finance Literature
SummaryThis paper introduces AlphaRLM, a reinforcement-learning framework that trains LLMs to construct executable symbolic trading factors directly from financial literature, without requiring input factor expressions. It combines Literature-Guided Operator Assembly (LEGO) with rewards for executable validity and grouped-return separation. Experiments on CSI 300, CSI 1000, and S&P 500 report that AlphaRLM outperforms several factor-mining and LLM-based baselines in annualized excess return.
In briefExact phase-type point processes for measuring and predicting event streams
SummaryThe paper proposes a point-process framework built on a phase-type (Erlang-chain) state space, giving exact likelihoods for both measuring market reactions and predicting the next event. Applied to six markets around 670 scheduled macroeconomic releases from 2022 to 2026, the measurement model finds that the direct reaction to CPI, payrolls, PPI and the FOMC statement is absorbed in about a second, while the market's own echoes stretch half the response to 16-137 seconds. The prediction model, EPT-X, is reported to achieve the best test likelihood on several event-stream datasets and to match the best published model on StackOverflow.
WhyIntroduces a phase-type Hawkes model with concave likelihood, closed-form shock responses and stability criterion, validated on markets and multiple event-stream benchmarks.
In briefEx ante and scenario-conditional value of information in dynamic portfolio choice
SummaryThis paper studies portfolio decisions under uncertainty about expected returns, where investors update beliefs as new information arrives. It introduces scenario-conditional expected utility and value of information, comparing a fixed-belief investor, a price observer, and an attentive investor with limited attention capacity in a finite-horizon Gaussian model. Ex ante, learning and attention are valuable, but scenario-conditionally learning brings both a correction benefit and a fluctuation cost, so the ranking can reverse.
In briefTail instability and scale control in neural network models that generate Treasury yield scenarios
SummaryThis paper compares two conditional normalizing flow models that generate scenarios for daily changes in eleven Treasury yield maturities. On a new 60-date window, the original model produced extreme samples with an absolute standardized coordinate of 34.3 million, while a stabilized Gaussian-base version reached 14.72. The authors find that numerical stability did not translate into better prediction, as historical simulation scored better on distributional measures and coverage on the new window.
