Weekly Quantitative Finance Paper Digest (Nov 22‑28, 2025)

Reinforcement Learning for Portfolio Optimization with a Financial Goal and Defined Time Horizons (arXiv:2511.18076v1)

Method : Combines the G‑Learning reinforcement‑learning algorithm with GIRL (G‑Learning Inverse Reinforcement Learning) to tune reward‑function parameters. The objective is to maximize portfolio value on a predefined target date while minimizing periodic investor contributions.

Experimental setting : Tested in high‑volatility market simulations with diversified assets.

Results : Sharpe ratio improves from 0.42 (baseline from prior work) to 0.483. Parameter comparison shows λ = 0.0012 (optimized by GIRL) versus λ = 0.002 (default); the performance gain from GIRL is modest, indicating that G‑Learning alone is already robust.

Paper link: https://arxiv.org/pdf/2511.18076v1

Portfolio Optimization via Transfer Learning (arXiv:2511.21221v1)

Problem : Asset markets often share underlying informational features that can be exploited across markets.

Approach : Develops a transfer‑learning‑based portfolio strategy that leverages cross‑market data. Forward validation is used to assess the usefulness of information before incorporation. The method iteratively adds information‑rich datasets and discards misleading ones.

Outcome : The strategy asymptotically approaches maximal Sharpe ratios as more reliable information is incorporated.

Paper link: https://arxiv.org/pdf/2511.21221v1

Hybrid LSTM and PPO Networks for Dynamic Portfolio Optimization (arXiv:2511.17963v1)

Architecture : A hybrid system where a Long Short‑Term Memory ( LSTM) network forecasts asset price dynamics and a Proximal Policy Optimization ( PPO) agent refines portfolio allocations in a continuous action space.

Data : Multi‑asset dataset covering U.S. and Indonesian equities, U.S. Treasury bonds, and major cryptocurrencies from Jan 2018 to Dec 2024.

Benchmarks : Equal‑weight portfolio, index‑based portfolio, LSTM‑only, and PPO‑only baselines.

Evaluation metrics : Annualized return, volatility, Sharpe ratio, maximum drawdown, all adjusted for transaction costs.

Findings : The hybrid LSTM‑PPO model yields higher returns and greater resilience under non‑stationary market regimes compared with all baselines.

Paper link: https://arxiv.org/pdf/2511.17963v1

Re(Visiting) Time Series Foundation Models in Finance (arXiv:2511.18578v1)

Scope : First large‑scale empirical study of Time‑Series Foundation Models (TSFMs) on global financial markets.

Dataset : Daily excess‑return data spanning multiple markets, used for zero‑shot inference, fine‑tuning, and training from scratch.

Comparisons : Pre‑trained TSFMs vs. models trained from scratch on financial data, evaluated against strong baseline time‑series models.

Key results : Pre‑trained TSFMs perform poorly in zero‑shot and fine‑tuning settings. Models trained from scratch achieve substantial predictive accuracy improvements and economic gains. Further gains are realized by scaling dataset size, adding synthetic data augmentation, and performing hyper‑parameter tuning.

Paper link: https://arxiv.org/pdf/2511.18578v1

KAN vs LSTM Performance in Time Series Forecasting (arXiv:2511.18613v1)

Models compared : Kernel‑Based Adaptive Networks ( KAN) and Long Short‑Term Memory ( LSTM) networks for stochastic stock‑price forecasting.

Metric : Root‑Mean‑Square Error (RMSE) across multiple forecasting horizons.

Findings : LSTM consistently yields lower RMSE than KAN, confirming superior accuracy for sequence modeling. KAN provides theoretical interpretability via the Kolmogorov‑Arnold representation theorem but incurs higher error rates and limited practical applicability. The primary advantage of KAN lies in computational efficiency for resource‑constrained scenarios where modest accuracy is acceptable.

Paper link: https://arxiv.org/pdf/2511.18613v1