Buildings constitute a large portion of energy consumption in the United States. Accurate forecasting of building energy consumption is integral to the implementation of energy efficiency initiatives and intermittent renewable energy supplies. The availability of high-resolution energy consumption data has allowed researchers to utilize machine learning techniques that forego domain-specific knowledge (e.g., building construction materials, geometric properties) to forecast energy consumption. While there is a growing body of literature surrounding the use of machines learning to forecast building energy consumption, previous research has yet to explore the use of feature selection to determine the most important subset of variables and produce interpretable predictive models. In this paper, we explore the use of Lasso, a shrinkage and selection method for linear regression that estimates sparse coefficients, to select the most important feature subset of a residential energy forecasting model. We evaluate the selected subset on an empirical data set from a multi-family residential building in New York City and compare the results to previous forecasting models without feature selection. Results of this work has implications on the data acquisition and sensing systems required to yield accurate predictions of residential energy consumption.