This paper uses a method based on multivariate power-normal distribution for predicting future gold prices in Malaysia. First let r(t) be the vector consisting of the month-t values of m selected macroeconomic variables, and gold price. The month-(t+1) gold price is then modelled to be dependent on the present and l-1 on past values r(t), r(t-1), …, r(t-l+1) via a conditional distribution which is derived from a [(m+1)l+1] -dimensional power-normal distribution. The mean of the conditional distribution is an estimate of the month- (t+1) gold price. Meanwhile, the 100(a/2)% and 100(1-a/2)% points of the conditional distribution can be used to form an out-of-sample prediction interval for the month-(t+1) gold price. For a given value of l, we select various combinations of m variables from a pool of 17 selected macroeconomic variables in Malaysia, and obtain the combinations of which the corresponding mean absolute percentage errors (MAPE) are relatively smaller while the coverage probabilities and average lengths of the prediction interval are still satisfactory. It is found that the parsimonious model is one of which l = 2, m = 1 and involving the macroeconomic variable derived from the Gross Domestic Product, Kuala Lumpur Composite Index or Import Trade.