The foreign exchange market, characterized by high volatility and economic significance, requires accurate predictive models. This study investigates the application of the Temporal Fusion Transformer (TFT), enhanced with Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), for forecasting major foreign exchange (forex) currency pairs: USD/EUR, USD/JPY, USD/CNY, USD/AUD, and USD/INR. The proposed framework integrates a wide range of economic indicators, which include interest rate differentials, GDP growth, and trade balances, alongside investor sentiment derived from Twitter and ESG-related news sentiment. By addressing the non-linear, multiscale nature of forex time series, the CEEMDAN-TFT model facilitates improved signal decomposition and interpretability. Empirical results indicate that the model demonstrates competitive performance across all five currency pairs, with the USD/EUR pair exhibiting the highest predictive accuracy. Other pairs, exhibiting good predictive accuracy, include USD/JPY and USD/CNY, underscoring the model’s adaptability to varying economic contexts. Performance is assessed using multiple error metrics, and the model is benchmarked against standard neural network approaches (MLP, RNN, LSTM, CNN). Variable importance analysis highlights the dynamic influence of interest rates, investor sentiment, and ESG factors across different market regimes. This study provides empirical evidence that including ESG and investor sentiment can improve the accuracy of currency forecasting models. This study provides guidance and a framework for informed decision-making for traders, analysts, and policymakers.