Objective To develop a data-driven model for estimating tangential ground reaction force (GRFt) from lower limb kinematic data and to select the most suitable input based on a balance between input quantity and estimation accuracy, with the aim of measuring GRFt in outdoor gait experiments. Methods Gait data from ten subjects walking at five different inclines (-10°, -5°, 0°, 5°, 10°) were used to train two data-driven models: a Backpropagation Neural Network (BPNN) and Polynomial Sparse Identification (PSI). The performance of these models was evaluated using eight kinematic data combinations and the normal ground reaction force (GRFn) as inputs to determine the optimal model and input combination. Results Under the same input dimensionality, the combination of hip-knee joint angles proved more accurate in estimating GRFt than the knee-ankle joint angle combination. Specifically, the BPNN and PSI models based on the former combination predicted GRFt with errors of 1.61% and 1.84% body weight (BW), respectively, while the latter combination resulted in errors of 2.82% and 3.15% BW. Moreover, when GRFn and all joint angles were used as inputs, the model’s prediction error was only 1.46% BW. Conclusion The combination of GRFn and hip-knee joint angles achieves an optimal balance between computational complexity and estimation accuracy. These findings support the accurate estimation of GRFt in outdoor gait testing.