The objective of this paper is to build an empirical model to predict the NBA teams' winning percentages from data collected in the past. The raw data has been standardized through Z-transform to remove any large variance bias effects. A multiple linear regression model was derived to predict the winning percentages. After trimming insignificant regression terms, the derived model can predict the teams' winning percentages with an R-Squared greater than 95%. The multicollinearity issues were addressed by minimizing the variance inflation factors. The redundant terms were removed to avoid the risk of over-fitting. The model has identified that three-point percentage, turnovers per game, and points per game were most critical to the team offensive efficiency. The nonlinearity terms have identified the complexity of basketball team behaviors. Defensive field goal percentage and points per game allowed were identified as the most significant interaction terms. The model accuracy was proved to be within +/-5% winning percentage of the predicted target across all thirty NBA teams.