Deriving decisions from data typically involves a sequential process with two components, forecasting and optimization. Forecasting models learn by minimizing a loss function that stands as a proxy for task-specific costs (e.g. trading, scheduling) without considering the downstream optimization problem, which in practice creates a performance bottleneck and obscures the impact of data on decisions. This work suggests leveraging the structure of the optimization component and directly learning a policy conditioned on explanatory data, effectively proposing a single data-driven module. For this purpose, we describe an algorithm to train ensembles of decision trees by directly minimizing task-specific costs, and prescribe decisions via a weighted Sample Average Approximation of the original problem. We then develop a generic framework to assess the impact of explanatory data on optimization efficacy. The proposed method is validated on the case of trading renewable energy in a day-ahead electricity market, where we design hybrid policies that balance optimal trading decisions and predictive accuracy. The empirical results demonstrate improved performance compared to solutions derived under the standard stochastic optimization framework. Further, we provide valuable insights on how explanatory data impact optimization performance and how this impact evolves under different market designs.