Abstract: To address the limited stress components constrained by hydraulic fracturing measurements, the difficulty of uniform-compression unit load cases in characterizing the burial-depth-dependent gradients of deep horizontal stresses, and the multicollinearity that may arise after introducing gradient-compression unit load cases, 20 valid test intervals from three boreholes below the -600 m level of the Luohe Iron Mine were investigated. An inversion observation vector was constructed from the measured data. A four-load-case combination consisting of self-weight, bidirectional uniform compression, and horizontal in-plane shear was established, and bidirectional gradient-compression unit load cases were further introduced to form a six-load-case combination for initial in-situ stress inversion. Pearson correlation coefficients and variance inflation factors were used to analyze the correlations among unit-load-case responses, ridge regression was employed to improve the stability of combination-coefficient estimation, and leave-one-out cross-validation (LOOCV) was adopted to evaluate the reliability of the inversion results. The results show that the four-load-case combination can reasonably reproduce the overall magnitudes of the horizontal principal stresses, but the burial-depth-dependent trends of the inverted stresses are inconsistent with the measured results. In contrast, the horizontal principal stresses obtained using the six-load-case combination increase with burial depth; however, the responses of the uniform- and gradient-compression load cases in the same direction are highly correlated, with a maximum variance inflation factor of 531.75, resulting in unstable estimation of the corresponding combination coefficients. After ridge-regression estimation, the comprehensive horizontal-principal-stress error and the comprehensive LOOCV error of the six-load-case combination are 11.17% and 12.09%, respectively. Both the maximum and minimum horizontal principal stresses increase with burial depth, showing the same variation direction as the measured results, although the inverted stress-gradient values remain lower than the measured values. This indicates that the six-load-case combination improves the characterization of the burial-depth-dependent variation of horizontal principal stresses while maintaining good inversion accuracy and predictive stability. The results demonstrate that introducing gradient-compression unit load cases into deep initial in-situ stress inversion can improve the characterization of the depth-dependent variation of horizontal stresses; however, the response correlations between the gradient- and uniform-compression load cases should be taken into account to improve the stability of combination-coefficient estimation and the reliability of the inversion results. These findings provide a reference for load-case configuration and stable solution of initial in-situ stress inversion in similar mines.