In the United States, consumers invest billions of dollars annually in energy efficiency, often on the assumption that these investments will pay for themselves via future energy cost reductions. Measuring the returns to energy efficiency investments requires estimates of counterfactual energy consumption, and recent research suggests that industry standard approaches to measuring savings may be overstating the gains from energy efficiency considerably. We develop and implement a machine learning approach for estimating treatment effects using high-frequency panel data, which are now widely available from smart meters. We study the effectiveness of energy efficiency upgrades in K-12 schools in California, and demonstrate that the machine learning method outperforms standard panel fixed effects approaches. We find that energy efficiency upgrades deliver only 53 percent of ex ante expected savings on average, and find a similarly low correlation between school-specific predictions of energy savings and realized savings. We see suggestive evidence that HVAC and lighting upgrades perform closer to ex ante expectations, as do smaller upgrades. However, we are unable to predict high realization rates using readily available demographic information, making targeting-based improvements challenging.
JEL Codes: Q4, Q5, C4
Keywords: energy efficiency; machine learning; schools; panel data