This study develops a correction model to improve the accuracy of horizontal spectral accelerations estimated by stochastic extended finite-fault simulation (EXSIM) in southeastern Korea. EXSIM predictions for five earthquakes (M4.3-5.5) recorded at eight stations reveal frequency-dependent residuals, with a tendency to underpredict spectral accelerations at frequencies ≥ 3 Hz. These discrepancies are correlated with eight variables: moment magnitude, stress drop, hypocentral distance, azimuth, average shear wave velocity up to 30 m in depth, relative elevation, and slope. To address these discrepancies, a multiple linear regression model is developed using eight variables that reflect earthquake source characteristics, wave-propagation paths, and site-specific conditions, including azimuth and topographic effects not fully accounted for in the original EXSIM. Application of this correction model substantially improves predictive performance, reducing root-mean-square error by 18.8% to 81.0% for the test sets. The corrected response spectra show good overall agreement with observations, including high-frequency spectral peaks. This approach enables the construction of reliable ground-motion databases. It enhances the accuracy of EXSIM predictions for scenario earthquakes, providing a practical tool for seismic hazard assessment in regions with sparse observational data.