Uni-FinLLM: A Unified Multimodal Large Language Model with Modular Task Heads for Micro-Level Stock Prediction and Macro-Level Systemic Risk Assessment

Financial institutions and national regulatory bodies increasingly require decision-support systems capable of integrating heterogeneous data sources to evaluate risks ranging from micro-level stock fluctuations to macro-level systemic vulnerabilities. However, existing approaches often treat these tasks in isolation, limiting their ability to capture cross-scale financial dependencies. To address this challenge, we propose Uni-FinLLM, a unified multimodal large language model that employs a shared Transformer-based backbone and modular task heads to jointly process textual financial news, numerical market time series, corporate fundamentals, and visual representations of financial dynamics. This architecture enables collaborative modeling of multiple financial layers, supporting applications such as institutional investment decision-making, credit-risk supervision, and national systemic-risk early warning. Through cross-modal attention fusion and multi-task optimization, Uni-FinLLM learns a coherent financial representation space from which micro-, meso-, and macro-level predictions can be derived. We evaluate the model using three complementary datasets covering stock-level forecasting, corporate credit-risk assessment, and macroeconomic stress detection. Experimental evaluations across three major financial tasks confirm the effectiveness of the proposed unified architecture. On the micro-level stock prediction benchmark, Uni-FinLLM raises directional accuracy from 61.7% (Llama-Fin baseline) to 67.4%, while reducing MAPE to 10.9 and improving the hit ratio to 64.3%. For credit-risk prediction, the model boosts accuracy from 79.6% to 84.1% and increases ROC-AUC to 0.892, substantially outperforming traditional machine-learning and financial-language-model baselines. At the macro level, Uni-FinLLM achieves an early-warning accuracy of 82.3% and a crisis-F1 score of 79.8%, exceeding the leading GNN-based macro-risk model by a significant margin. Together, these results validate that a unified multimodal LLM equipped with modular task heads can jointly model micro-scale asset behavior and macro-scale systemic vulnerabilities, offering a practical and scalable decision-support engine for financial institutions and national regulators.