Key Quantitative AI Papers Jan 3‑9 2026: Portfolio Optimization, Equity Correlation Forecasting, and Index Tracking Review

Smart Predict‑then‑Optimize Paradigm for Portfolio Optimization in Real Markets

Paper link: https://arxiv.org/pdf/2601.04062v2

Authors: Wang Yi, Takashi Hasuike

The study addresses the mismatch between higher return‑prediction accuracy and actual investment‑decision quality under realistic trading frictions and constraints. It adopts the Smart Predict‑then‑Optimize (SPO) paradigm, aligning the learning objective with downstream portfolio‑selection performance rather than point‑wise prediction error. A linear predictor built from return and technical‑indicator features is trained with an SPO‑based surrogate loss that directly reflects the quality of the resulting portfolio decisions. The predictor is coupled with a portfolio‑optimization model that incorporates transaction costs, turnover control, and regularization. Evaluation uses a rolling‑window backtest with monthly rebalancing on U.S. ETF data spanning 2015‑2025. Empirical results show that the decision‑oriented training consistently outperforms conventional prediction‑then‑optimize baselines and classic optimization benchmarks in risk‑adjusted returns, and demonstrates strong robustness during adverse market conditions such as the 2020 COVID‑19 crisis.

Forecasting Equity Correlations with a Hybrid Transformer Graph Neural Network

Paper link: http://arxiv.org/pdf/2601.04602v1

Authors: Jack Fanshawe, Rumi Masih, Alexander Cameron

The paper tackles forward‑looking stock‑stock correlation prediction for S&P 500 constituents and investigates whether improved correlation forecasts enhance graph‑based clustering for basket‑trading strategies. Correlations ten days ahead are predicted in Fisher‑z space, and a Temporal Heterogeneous Graph Neural Network (THGNN) is trained to predict residual deviations relative to a rolling historical baseline. THGNN combines a Transformer‑based temporal encoder that captures non‑stationary, complex time dependencies with an edge‑aware graph attention network that propagates cross‑asset information over the stock network. Input features include daily returns, technical indicators, industry structure, prior correlations, and macroeconomic signals, enabling institution‑aware predictions and attention‑based interpretability of feature and neighbor importance. Out‑of‑sample experiments covering 2019‑2024 demonstrate a statistically significant reduction in correlation‑prediction error compared with the rolling‑window estimator. When the predicted forward‑looking correlations are fed into a graph‑based clustering framework, the resulting baskets adapt to market stress and exhibit economically meaningful performance.

A Comprehensive Review and Analysis of Modeling Approaches for the Financial Index Tracking Problem

Paper link: https://arxiv.org/pdf/2601.03927v1

Authors: Vrinda Dhingra, Amita Sharma, Anubha Goel

The review categorizes index‑tracking models into three broad frameworks: (1) optimization‑based models, (2) statistical‑based models, and (3) machine‑learning‑driven data‑analysis methods. Extensive empirical studies on the S&P 500 dataset compare representative methods from each category. Within the optimization framework, the tracking‑error volatility model achieves the lowest tracking error, indicating the most accurate replication of the target index. In the statistical framework, a convex cointegration model delivers the strongest return‑risk balance. In the machine‑learning framework, a fixed‑noise deep neural network attains competitive tracking performance while maintaining notably low turnover and high computational efficiency. The analysis highlights the trade‑offs among accuracy, risk balance, turnover, and computational cost across the three modeling paradigms.