DTMF-Net: Diagnostics-Guided Temporal Multimodal Fusion Network for Financial Forecasting

Financial forecasting in real-world settings requires integrating heterogeneous and asynchronously aligned information sources, including structured market time-series and unstructured financial text. However, existing multimodal approaches often assume perfect temporal alignment or treat all textual inputs as equally reliable, leading to degraded performance under sparse and noisy conditions. This paper proposes a Diagnostics-Guided Temporal Multimodal Fusion Network (DTMF-Net), a causality-preserving framework that explicitly models alignment reliability during multimodal fusion. The proposed method incorporates (i) a market temporal encoder for sequential dynamics, (ii) a text encoder for semantic event representation, (iii) a leakage-aware alignment mechanism that enforces strict temporal causality, and (iv) a diagnostics-guided reliability module that quantifies event coverage, temporal lag, and semantic dispersion to modulate fusion. Experiments conducted on a real-world financial dataset demonstrate that DTMF-Net achieves consistent improvements over strong baselines, with an accuracy of 59.8% and improved performance across AUC, MCC, and regression metrics. Calibration analysis shows improved reliability through lower ECE and Brier score, while uncertainty estimates exhibit meaningful association with prediction error. A leakage validation study shows that relaxed or non-causal alignment can inflate apparent forecasting performance, highlighting the necessity of strict temporal constraints in financial multimodal evaluation. Furthermore, economic backtesting demonstrates improved risk-adjusted returns, with higher Sharpe ratio and reduced drawdown compared to benchmark strategies. These results establish that reliability-aware multimodal fusion, rather than increased model complexity, is critical for robust financial forecasting under asynchronous and low signal-to-noise conditions.