<p>This study demonstrated the effectiveness of Discrete Wavelet Transform (DWT) as a denoising technique prior to regression modeling for loan prediction in Nigeria’s banking sector. Using monthly commercial bank loan data from the Central Bank of Nigeria spanning 1999 to 2022, the research addresses the common issue of noise in financial time series, which often leads to model misspecification, autocorrelated residuals, and unreliable forecasts. Four denoising methods were applied before model development: moving Average (MA), Savitzky-Golay (SG) filter, random forest regression, and DWT. Unlike many prior studies that adopt a single wavelet type, this work benchmarks over 90 DWT variants across multiple families (Haar, Daubechies, Symlets, and Coiflets) and decomposition levels to identify the optimal denoising configuration. The regression models were evaluated using standard performance metrics (R², MAE, MSE, RMSE, MAPE) and diagnostic statistics (Durbin–Watson, residual normality, and autocorrelation). Results show that while the raw data model achieved the highest R² (0.990), it exhibited strong residual autocorrelation (Durbin–Watson = 0.913). In contrast, the DWT-denoised model achieved superior residual independence and statistical robustness, with only a minimal reduction in R² (0.966). Although the random forest-denoised model offered slightly better predictive accuracy (R² = 0.9897), it required significant computational resources and lacked interpretability, limiting its applicability in regulatory or resource-constrained environments. This study contributes to the literature by positioning DWT as a practical, interpretable, and computationally efficient alternative to more complex machine learning approaches for financial forecasting—particularly in emerging economies with limited data and infrastructure. Future work may explore hybrid models that integrate DWT with deep learning techniques to further enhance forecasting accuracy while maintaining transparency and feasibility.</p>