Accurate in-situ stress prediction is critical to optimizing hydraulic fracturing design and enhancing reservoir stimulation in deep shale gas development. However, traditional methods often struggle with accuracy and adaptability due to the geological complexity and high-dimensional, nonlinear nature of well-logging data. To address this challenge, we propose AutoMix, a hierarchical hybrid prediction model built on the AutoGluon framework. AutoMix integrates leading machine learning algorithms—LightGBM, CatBoost, ExtraTrees, RandomForest, and XGBoost—along with deep learning architectures such as NeuralNetTorch and NeuralNetFastAI, using multi-level stacking and weighted ensembling to enhance prediction performance for maximum and minimum horizontal stress (σH and σh). The model was trained and evaluated using 278,003 sets of real-world logging data from 14 horizontal shale gas wells in the Zigong area of the Sichuan Basin. Eleven algorithms were benchmarked under varying sampling ratios (1/1000, 1/100, 1/10, and full dataset). Results show that AutoMix consistently achieves the lowest errors across all scales (MSE = 0.0046, MRE = 0.454, R2 = 0.9998), significantly outperforming any single model with strong generalization and robustness. Feature importance analysis identifies vertical stress, Poisson's ratio, feldspar content, and Young's modulus as key predictors, consistent with established geomechanical theory. Weight distribution analysis further reveals that deep learning models dominate in low-sample regimes, while tree-based models such as ExtraTrees and RandomForest take precedence as data volume increases—each excelling in σh and σH prediction, respectively. AutoMix demonstrates exceptional performance under data scarcity, nonlinearity, and multicollinearity, offering strong applicability and scalability for engineering-grade stress prediction in unconventional reservoirs.