This paper defines and studies a variable selection procedure called Testing-Based Forward Model Selection. The procedure inductively selects covariates which increase predictive accuracy into a working statistical regression model until a stopping criterion is met. The stopping criteria and selection criteria are defined using statistical hypothesis tests. The paper explicitly describes a testing procedure in the context of high-dimensional linear regression with heteroskedastic disturbances. Finally, a simulation study examines finite sample performance of the proposed procedure and shows that it behaves favorably in high-dimensional sparse settings in terms of prediction error and size of selected model.