A Dual-Output Deep Neural Network Model for Stock Risk and Potential-Loss Prediction in Organisational Finance

Organisational stock risk management continues to rely on parametric models whose core assumptions break down during periods of market stress, leaving a gap between forecast accuracy and interpretable risk transparency. This study addresses that gap with a dual-output Deep Neural Network (DNN) that fuses a correlation feature module with an LSTM-based sequence feature module to jointly produce a continuous Value-at-Risk (VaR)/Conditional Value-at-Risk (CVaR)- style risk score and a categorical (Low/Medium/High) risk label. The model was trained via backpropagation with the Adam optimiser on financial, sentiment, macroeconomic, and microstructure indicators drawn from Nigerian Exchange (NGX) tickers, and implemented as a full-stack platform with an integrated model-comparison and external-benchmarking module. Internally, the DNN outperformed LSTM, Random Forest, and Support Vector Machine (SVM) baselines, achieving the highest classification accuracy of 86 percent and an R-squared value of 0.8426. Benchmarked externally against two published systems, the DNN achieved the lowest mean absolute error and the highest accuracy, outperforming a reinforcement-learning-based Deep Portfolio Optimisation framework (79 percent) and a federated-learning framework (72 percent); however, both existing systems achieved a materially higher Sharpe ratio, reflecting their design around portfolio-level optimisation rather than pointwise risk prediction. These results position the proposed model as a complementary, transparency-focused alternative to reward-shaped reinforcement-learning approaches rather than a wholesale replacement for them.