Abstract
We use a real-time boosting approach to study the time-varying out-of-sample informational content of various predictor variables (inflation rate, exchange-rate fluctuations, stock market returns and interest rates) for forecasting gold-price fluctuations. While the predictor variables have predictive power, the economic value added of forecasts does not suffice to leverage the performance of a simple trading rule above the performance of a buy-and-hold strategy.
Notes
1 For earlier research on the informational efficiency of the gold market, see Tschoegl (Citation1980) and Ho (Citation1985). Fuertes et al. (Citation2010) study the profitability of trading rules in commodity markets.
2 See also Vrugt et al. (Citation2007). For a neural-network approach to forecasting gold-price fluctuations, see Parisi et al. (Citation2008).
3 The boosting approach is not restricted to linear models. Friedman et al. (Citation2000) discuss various boosting approaches from a statistical point of view.
4 As recommended by Bühlmann and Hothorn (Citation2007), we estimate the boosting approach on demeaned data, where we compute the mean values of excess returns and the predictor variables in real time. For computing forecasts, we simply add the period-t real-time mean excess returns of gold-price fluctuations to the forecast computed from the period-t forecasting model. The forecasting model, thus, contains the real-time mean excess returns and one or more of the predictor variables.
5 The datasource is the Federal Reserve Bank of St. Louis. The only exception is the exchange rate, which is from the Bank for International Settlements. For further details on the data, see Pierdzioch et al. (Citation2014a).
6 Results of market-timing tests (not reported) confirm that forecasts have significant explanatory power with respect to the sign of subsequent gold-price fluctuations.
7 For a formal description of the trading rule, see Pesaran and Timmermann (Citation1995) and Pierdzioch et al. (Citation2014a). We rule out the possibility of short selling and leveraging.
8 In order to retain the temporal and cross-sectional dependence of the data, we use the block bootstrap (Politis and Romano, Citation1994) to resample from the data. All simulations use the same seed of random numbers.