An increase in the number of asset pricing models intensifies model uncertainties in asset pricing. While a pure “model selection” (singling out a best model) can result in a loss of useful information, a full “model pooling” may increase the risk of including noisy information. We make a trade-off between the two methods and develop a new two-step trimming-then-pooling method to forecast the joint distributions of asset returns using a large pool of asset pricing models. Our method allows investors to focus on certain regions of the distributions. In the first step, we trim the uninformative models from a pool of candidates, and in the second step, we pool the forecasts of the surviving models. We find that our method significantly enhances portfolio performance and predicts downside risk precisely, and the improvements are mainly due to trimming. The pool of sensible models becomes larger when focusing on extreme events, responds rapidly to rising uncertainty, and reflects the magnitude of factor premiums. These findings provide new insights into asset pricing model evaluation.

资产定价模型数量的增加加剧了资产定价模型不确定性。尽管单纯的“模型选择”（挑选出最佳模型）可能会导致有用信息的丢失，但是完整的“模型库”可能会增加包含嘈杂信息的风险。我们在这两种方法之间进行了权衡，并开发了一种新的两步修整再合并方法，以使用大量资产定价模型来预测资产收益率的联合分布。我们的方法使投资者可以专注于分布的某些区域。第一步，我们从一组候选者中修剪出无信息的模型，而第二步，我们对尚存模型的预测进行汇总。我们发现，我们的方法显着提高了投资组合的绩效并准确预测了下行风险，而这种改进主要是由于调整。当关注极端事件时，明智的模型池变得更大，可以对不确定性迅速做出反应，并反映出要素保费的大小。这些发现为资产定价模型评估提供了新的见解。