Abstract: Aggregating individual forecasts improves accuracy through the ''wisdom of the crowd'' effect. This paper examines how accuracy of the aggregate forecast can be further improved when prediction intervals are also elicited and available. We consider a simple theoretical model where forecasters observe an unbiased signal from a data-generating process. Each forecaster reports lower and upper estimates along with a most likely estimate for a continuous quantity. Such three-point estimates are commonly collected in practice, and provide information on the skewness of forecasts. Our theoretical analysis suggests that the simple average of most likely estimates overestimates small quantities, and underestimates large quantities. Evidence from experimental forecasting tasks supports these observations. We explore alternative approaches that use skewness in three-point estimates to alleviate the bias in average forecast.