Speaker
Giovanni Cerulli, CNR IrcRes
Abstract:
In standard optimal policy learning (OPL), the policymaker is assumed to be risk- neutral and maximizes expected welfare under a treatment assignment rule. This paper develops a risk-adjusted alternative motivated by Roy’s (1952) safety-first principle. We show that maximizing the probability of exceeding a socially required minimum outcome naturally leads to a pointwise optimal treatment rule based on the ratio of conditional means to conditional standard deviations of potential outcomes. Importantly, this mean/variance rule does not arise from mean–variance preferences, but from a probabilistic safety criterion. We provide theoretical foundations and an empirical application to assess optimal policy learning for the CAP (Common Agricultural Policy) using data from the Farm Accountancy Data Network (FADN), known in Italian as "Rete di Informazione Contabile Agricola" (RICA).