Combining Wavelet Decomposition with Machine Learning to Forecast Gold Returns

I combine the discrete wavelet transform with support vector regression to forecast gold-pricedynamics. I investigate the advantages of this approach using a relatively small set of economic and financial predictors. In order to measure model performance, I differentiate between statistical and economic forecast evaluation, where the economic valued-added of forecasts is simulated using a trading strategy. I show that disentangling the predictors with respect to their time and frequency domain leads to improved forecast performance. Results are robust to alternative forecasting approaches. Findings on the relative importances of such wavelet decompositions suggest that the influence of short-term and long-term trends is not stable over the full evaluation period