Quantitative Finance Paper Digest: AI‑Driven Market Prediction Studies (Mar 7‑13 2026)

Stock Market Prediction Using Node Transformer Architecture Integrated with BERT Sentiment Analysis addresses the difficulty of forecasting stock prices in noisy, non‑stationary markets. The authors model the market as a graph where stocks are nodes and edges capture industry links, price co‑movements, and supply‑chain connections. A fine‑tuned BERT extracts sentiment from social‑media posts, which is fused with quantitative features via attention. A node‑transformer processes historical data to capture temporal evolution and cross‑sectional dependencies. Experiments on 20 S&P 500 stocks from Jan 1982 to Mar 2025 achieve a one‑day ahead MAPE of 0.80 % (ARIMA = 1.20 %, LSTM = 1.00 %). Sentiment analysis reduces overall error by 10 % and by 25 % during earnings announcements; graph modeling adds a further 15 % improvement. Directional accuracy reaches 65 %, and paired‑sample t‑tests confirm significance (all p < 0.05). In high‑volatility periods the model keeps MAPE < 1.5 % while baselines exceed 2 %.

A Hybrid Quantum‑Classical Framework for Financial Volatility Forecasting Based on Quantum Circuit Born Machines proposes a novel hybrid architecture that combines a classical LSTM with a Quantum Circuit Born Machine (QCBM). The LSTM extracts dynamic features from historical price series, while the QCBM serves as a learnable prior distribution guiding the prediction. The framework is evaluated on two real‑world high‑frequency datasets—the Shanghai Stock Exchange Composite Index and the CSI 300—using 5‑minute bars. Compared with a pure LSTM baseline, the hybrid model yields lower mean‑squared error, root‑mean‑squared error, and QLIKE loss, demonstrating the potential of quantum‑enhanced learning for financial time‑series forecasting.

Constructing a Portfolio Optimization Benchmark Framework for Evaluating Large Language Models introduces a benchmark that tests LLMs on mathematically solvable portfolio‑optimization problems rather than pure language tasks. The authors generate multiple‑choice questions by varying objectives, assets, and constraints, each with a unique correct solution and several plausible alternatives. Experiments with GPT‑4, Gemini 1.5 Pro, and Llama 3.1‑70B show distinct performance patterns: GPT‑4 attains the highest accuracy on risk‑based objectives and remains stable under constraints; Gemini excels on return‑based tasks but falters elsewhere; Llama records the lowest overall scores. The results highlight both the promise and current limits of LLMs for quantitative financial reasoning.

Factor Dimensionality and the Bias‑Variance Tradeoff in Diffusion Portfolio Models implements a conditional diffusion model that predicts the full distribution of asset returns conditioned on company‑level factors. By varying the number of factors, the authors observe a clear bias‑variance tradeoff: too few factors lead to under‑fitting and overly diversified portfolios, while too many cause over‑fitting, instability, and concentrated allocations with poor out‑of‑sample performance. An intermediate factor count achieves the best generalization and outperforms baseline portfolio strategies.