Forecasting with the maximum of correlated components having bivariate t-distributed errors

The value of the maximum of two forecasts as a component ina linear combination forecast is assessed when the errors havea bivariate t distribution. It is found that, while the maximummay provide information additional to that of the mean forecastif the components are not equally biased and have heteroscedasticerrors, the best linear combination forecast does not give anyweight to the maximum.     

Property appraisal with a log - t predictive distribution

A multiplicative model with log-normal errors is used to appraisethe market value of a property and a natural conjugate prior,describing the beliefs of an expert appraiser about the parameters,is elicited. Loss is an asymmetric function of the predictedprice as a proportion of the actual price, and the optimal predictionis a multiplicative adjustment to the exponent of the predictivemean of the logarithm of price. The prediction of the priceof a property is made using both the elicited prior and tworeference priors. The robustness of the procedure is assessedby means of the evaluation of a number of elasticities of priceprediction with regard to changes in a parameter value, eitherof the prior distribution or the loss function.