Research Directions
Methods and pitfalls of machine learning in return prediction, feature engineering, and portfolio construction.
54 related activitiesLast 30 days 48 itemsFeatured among them 0 items
This paper proposes PALM (Point-in-time Adaptation for financial Language Models) as an alternative to annual pre-training of financial language models. It achieves adaptation by fitting low-rank adapters on permitted financial text without modifying any pre-trained weights.
Read originalBased on 4.57 million news articles covering approximately 3,000 US stocks from 2023 to 2026, the paper uses a compact classifier distilled from a large model teacher to label news with 17 event categories and 5 attributes, clustering them into 1.68 million stock-day events, of which 364,405 are neutral sentiment events serving as a placebo.
Read originalThe study proposes the Identifiable Variational Dynamic Factor Model (iVDFM) for multivariate time series modeling. The model combines variational inference to provide partial identifiability guarantees within a known fuzzy class range. The model's innovation terms are conditioned on observed auxiliary context and mechanism embeddings, follow a conditional exponential family prior, and are trained by maximizing the ELBO to estimate uncertainty in latent trajectories and predictions. This research has been published in the ProbML 2026 conference proceedings, and the code is open source.
Read originalThe paper proposes a branching signature model that generalizes the classical geometric rough path based on the branching rough path framework, proving its universal approximation theorem, which can approximate high-level signatures through iterative combinations of low-level signature mappings. The model can explicitly extend the original path to a high-dimensional space via the mapping Ψ, implementing it as the classical geometric signature of the extended path, and its effectiveness is verified in two applications: rough volatility calibration and optimal execution.
Read originalThe study proposes a conditional latent diffusion framework that jointly generates 30-step trajectories of the implied volatility surface and underlying returns. It learns surface representations through an arbitrage-aware autoencoder, with the diffusion process capturing the evolution dynamics. The framework is evaluated on the SPX surface dataset and can generate realistic probabilistic multi-step scenarios, with its forecast mean outperforming the persistence benchmark for horizons beyond the shortest maturity.
Read originalThe paper proposes using the half-width of conformal prediction intervals as a scaling measure for fractional Kelly position sizing with a leverage cap, and tests this rule under a pre-registered protocol. During the development phase, an autonomous LLM agent searched approximately 200 configurations on a 2016-2021 sample of 8 U.S. equity ETFs, with subsequent data sealed before the search, and the final configuration, 9 comparison portfolios, and conclusion mapping all registered before unsealing.
Read originalThis paper proposes an arbitrage-aware latent flow matching framework for unconditional generation of implied volatility surfaces. The framework first compresses high-dimensional surfaces into a low-dimensional latent space using a variational autoencoder with differentiable penalty regularization for calendar spreads, call spreads, and butterfly arbitrage. A flow matching model then learns to map a Gaussian prior to the empirical latent distribution, which is finally decoded back to a volatility surface. This method can generate 90.8% of surfaces that satisfy all tested static no-arbitrage conditions, performs best in the extreme Q99 range, and balances distributional similarity, tail preservation, and financial consistency without post-sampling reweighting.
Read originalA new arXiv paper uses Gaussian surrogate models to precisely calculate the finite-sample risk of candidate scoring decision models, analyzing the boundaries of the role of three training elements: teacher soft targets, frozen pre-trained backbones, and temperature calibration. The paper points out that after a student model inherits teacher bias, more student data cannot eliminate that bias; fixed representations only reduce risk below an explicit sample size threshold; under the probit model, exact recalibration maps to temperature scaling, and can only eliminate all excess risk when the direction of the teacher's coefficients is correct.
Read originalThe study addresses market-making challenges in Polymarket's thin order books, using the JaxMARL-HFT framework to train cooperative multi-agent market-making models on 417 real markets. The market-making agents are split into quoters and risk managers, trained using recurrent independent proximal policy optimization (IPPO) combined with RNN policies. Testing showed that the improved single-agent market-making model quoted more frequently and had smaller out-of-sample losses, but none of the models achieved positive average portfolio value in out-of-sample evaluation, with only single-seed training completed.
Read originalThis arXiv paper tests the degree to which 4 AI trading agents adhere to their own strategies and risk rules, based on 2627 closed trades from February 12 to August 27, 2026. 18% of the trades were live orders on Hyperliquid, while the rest were simulated at real-time prices, with a uniform round-trip fee of 0.09%.
Read originalAuDQ, an auditable data quality layer for high-risk tabular data pipelines, proposes an ontology-driven, warehouse-native design vision that treats data quality rules as first-class artifacts, supporting version control, governance text traceability, and immutable execution trajectory replay. Its agentic rule discovery core includes both fully autonomous and guided rule acquisition modes, along with human-machine collaborative review interfaces, and has completed preliminary validation experiments on real-world datasets from the finance and retail industries.
Read originalThis study introduces LOBGrad, a conditional DDPM framework for probabilistic order book forecasting. The framework combines an autoregressive RNN encoder with a diffusion-based generator, and establishes an excess-risk oracle inequality for the single-step DDPM noise prediction objective under exponential mixing and regularity assumptions.
Read originalResearchers propose a game graph equilibrium framework that defines network structure as a solidified record of strategic interactions, proving that two types of games correspond to four network topologies, and K-class interactions correspond to K^K classes of topologies. This framework enables mechanism inference on low-dimensional game-theoretic manifolds through constrained block modeling, can reconstruct generative dynamics from observed networks, and performs well in real-world networks such as international trade.
Read originalThis arXiv paper proposes a conditional weight of evidence method to measure the marginal impact of a single credit feature on the log odds of default, given the rest of an applicant's credit information. The same feature value can yield different evaluations for applicants with different credit backgrounds.
Read originalThe paper proposes RegimeRank, an interpretable industry ranking model that frames industry rotation decisions as a task of ranking industry actions by future risk-adjusted active returns, using the forward-looking industry benchmark Sharpe ratio as the measurement metric. The model represents fundamental and technical signals through expert-defined time-series features, uses cross-attention to condition on learned macro factors, then models cross-industry dependencies through self-attention, while using expected gradient attribution to identify the contribution of individual signal-transformation pairs.
Read originalThe study proposes an Alpha-stable Lévy diffusion predictor called SHADE, which achieves distribution-level shrinkage through a residual stable kernel by reading outputs at non-zero depths, addressing the problem of overly narrow predictive distributions and overconfidence when Gaussian diffusion models forecast heavy-tailed time series.
Read originalA new arXiv paper studies the predictive performance of foundation models for return prediction when historical observations are inconsistent with recent return distributions. Authored by Levent Bulut, the 10-page paper includes 5 figures and belongs to the statistical finance (q-fin.ST) field.
Read originalThe paper proposes a Neural Filter Estimator (NFE), trained on simulated paths of mean-reverting processes, that outputs the posterior distribution of the implied state for each spread in the cross-section in a single forward pass, incorporating estimation uncertainty into trading decisions. On simulated paths with known ground truth, the method achieves accuracy comparable to other practical estimators, with its advantage lying in correctly characterizing its own uncertainty.
Read originalThis study introduces topological data analysis (TDA) and persistent homology (PH) into forecasting models in the field of prediction markets, exploring whether the underlying structure of market belief states has predictive value. The research findings show that single-parameter PH can improve the predictive power and calibration accuracy of linear forecasting models, and that two-parameter multi-persistent homology can produce similar gains for nonlinear models. The paper is 41 pages long with 19 figures, and belongs to the field of Trading and Market Microstructure (q-fin.TR).
Read originalThis paper proposes AddMart, a few-shot implied volatility surface reconstruction framework that embeds no-arbitrage structure into the architecture to achieve global static no-arbitrage, enabling continuous and valid surface generation from a small number of option quotes. Experiments on 10 real-world option datasets show that the framework delivers competitive reconstruction accuracy with no significant static arbitrage violations, and can be directly applied to VaR, CVaR and crash probability estimation.
Read original
