This paper presents a Scaled Sequential Thresholded Least Squares (S2TLS) algorithm to construct sparse regression models for flight load prediction. The combined use of a sparsification parameter λ and a magnification factor χ is proposed to tune both the model complexity and the regressor complexity. A bumpiness function is introduced to preferentially select simple regressors to improve model prediction. A cost function J consisting of the estimation residual and the bumpiness is then proposed to determine parameters (χ, λ) satisfying balanced performance. Parametric analysis is undertaken to investigate the effect of χ and λ on sparse regression performance. It is found that the optimal solution is hidden within a trench-like region of the (χ, λ) domain. Two methods using different optimization variables and algorithms are then presented to search for optimal combinations of λ and χ. Case studies are performed, and model results are compared against the test and CFD data of flight loads. Excellent agreement is observed, and the data (even with significant complications) is well bounded by the 95% confidence interval. Importantly, the underlying load-driving factors can be successfully identified. The new S2TLS algorithm represents a robust, efficient, and accurate method for flight load modeling and prediction.