Accurate gold price forecasting is essential for financial decision making, particularly in volatile economic environments. Traditional time series models such as ARIMA and Transformers often struggle to capture the complex and dynamic dependencies between macroeconomic variables. To address these challenges, we propose KAN-GAT, a novel hybrid forecasting framework that integrates Kolmogorov-Arnold Networks (KAN) with Graph Attention Networks (GAT). The KAN module captures non-linear relationships among financial indicators, while GAT refines predictions by modeling temporal dependencies through graph-based learning. Experimental evaluations on real-world financial data—including gold prices, USD Index, oil prices, and inflation rates, demonstrate that the KAN-GAT significantly outperforms state-of-the-art models, achieving a MAPE of 2.75 ± 0.17, compared to 11.32 ± 15.19 by KAN, 8.22 ± 3.46 by GAT, 9.56 ± 1.77 by N-HiTS, 20.94 ± 9.20 by N-BEATS, and 31.00 ± 1.76 by Transformers. Additionally, KAN-GAT reduces RMSE to 2.94 ± 0.37 and MAE to 2.28 ± 0.15, achieving a statistically significant improvement over baseline methods. These findings highlight the effectiveness of our proposed hybrid approach in financial forecasting and establish KAN-GAT as a powerful tool for capturing complex market dynamics.