时 间:
【时空经济沙龙】Establishing Prior Beliefs for Innovations with Limited Precedents: The DME Approach
时间:2026年11月16日(周一)13:00-15:00
地点:思源东楼611
报告人简介:赵琳,上海财经大学教授,数字经济学院院长,上海市浦东新区数字化转型专家委员会委员,在清华大学数学系取得理学学士学位、理学博士学位。研究方向为行为管理、风险管理、运营管理,研究成果发表于管理学、应用经济学等国内外期刊,2019年至今任Management Science决策分析领域Associate Editor。
报告摘要:When firms evaluate (AI-driven) innovations with limited precedents, domain experts and professional forecasters can produce probability distributions over potential outcomes. Yet these expert-generated distributions are imperfect—subject to cognitive bias, limited information, and overconfidence. Adopting them directly as priors is unjustified; discarding them entirely wastes valuable judgmental input. We propose the DME approach (Divergence-Constrained Maximum Entropy), which resolves this tension by maximizing Shannon entropy subject to Kullback–Leibler divergence constraints between the elicited prior and expert-generated candidate distributions. The solution recovers log-linear opinion pooling (Genest, McConway & Schervish, 1986) as a special case, but endogenizes the pooling weights $\alpha_m$ through per-expert information budgets rather than requiring exogenous assignment. Each weight admits a direct behavioral interpretation—reflecting the decision-maker's willingness to deviate from a given expert—and can be elicited through a mixture-preference protocol analogous to swing-weighting in multi-attribute decision analysis. The framework accommodates multiple experts, permits $\sum \alpha_m \le 1$ to encode epistemic humility, and naturally compresses extreme probability masses, mitigating expert overconfidence. We formalize identification conditions, demonstrate comparative properties against subjective elicitation and unconstrained entropy baselines, and illustrate the method through an application to LLM adoption decisions.