This research presents DLensRisk, a novel and integrative framework that combines deep learning models with Value-at-Risk (VaR) methodologies to forecast steel price volatility and assess financial risk in offshore oil and gas projects. Steel is a critical input in subsea infrastructure, such as flexible risers, manifolds, and Christmas trees, whose procurement is often affected by global price instability, long lead times, and contract inflexibility. DLensRisk evaluates seven forecasting models across regional datasets (USA, Europe, Asia, and China), including classic methods, ARIMA, Exponential Smoothing, and Linear Regression, and deep learning architectures: Recurrent Neural Networks (RNNs), Transformer, Neural Hierarchical Interpolation for Time Series (N-HiTS), and Temporal Convolutional Network (TCN). Model performance is assessed using SMAPE, MAPE, MASE, and MARRE, with statistical validation via the Friedman and Nemenyi tests. The best-performing model for each region is integrated with four VaR methodologies, Historical Simulation, eGARCH, tGARCH, and Filtered Historical Simulation (FHS), to construct a Financial Risk Index. Results show that deep learning models consistently outperform traditional approaches, with N-HiTS leading in China, RNN in Asia and Europe, and Transformer in the United States. For risk modeling, FHS proved most suitable for stable markets like Asia and China, while tGARCH delivered more conservative estimates for volatile markets like the USA and Europe. By aligning predictive modeling with formal risk metrics, DLensRisk offers a robust, data-driven strategy for mitigating procurement risk in capital-intensive energy projects. This framework enhances scenario planning and supports more accurate, regionally adaptive financial decision-making in the oil and gas industry.