Gold Price Forecasts in a Dynamic Model Averaging Framework – Have the Determinants Changed Over Time? (2014)
Baur et al (2014) adopt the Dynamic Model Averaging approach to forecast the price of gold. Results show that DMA approach outperforms alternatives such as the random walk and Bayesian Model Averaging. They also find that the factors that influence future gold prices vary significantly through time and can be distinguished from results that focus on the in-sample and non-predictive relationships within a classical regression model.
Dynamic Model Averaging
- 3pm London fixing price in US dollars
- the MSCI world stock price index
- the S&P500 composite price index as representative stock price indices
- a GARCH(1,1) process of the returns of the MSCI world index as a measure of stock price volatility
- the S&P GSCI commodity price index,
- the CRB commodity price index,
- the price of silver,
- the US consumer price index,
- a global composite price index,
- the US dollar trade-weighted index,
- the euro trade-weighted index,
- US 3-month Treasury bill
- US 10 year Treasury bond yields,
- the Barclays Capital US aggregate bond index,
- a global foreign currency reserves index.
The price of gold is influenced by a wide range of local and global factors such as commodity prices, interest rates, inflation expectations, exchange rate changes and stock market volatility among others. Hence, forecasting the price of gold is a notoriously difficult task and the main problem a researcher faces is to select the relevant regressors at each point in time. This combination of model and parameter uncertainty is explicitly accounted for by Dynamic Model Averaging which allows both the forecasting model and the coefficients to change over time. Based on this framework, we systematically evaluate a large set of possible gold price determinants and use both the predictive likelihood and the mean squared error as a measure of the forecasting performance. We carefully assess which predictors are relevant for forecasting at different points in time through the posterior probability. Our findings show that (1) DMA improves forecasts compared to other frameworks and (2) provides clear evidence for the time variation of gold price predictors.