Machine Learning from Schools about Energy Efficiency

We use high-frequency panel data on electricity consumption to study the effectiveness of energy efficiency upgrades in K–12 schools in California. Using a panel fixed effects approach, we find that these upgrades deliver between 12% and 86% of expected savings, depending on specification and treatment of outliers. Using machine learning to inform our specification choice, we estimate a narrower range: 52%–98%, with a central estimate of 60%. These results imply that upgrades are performing less well than ex ante predictions on average, although we can reject some of the very low realization rates found in prior work.