Predictability in the Gold Market: Traditional Forecasting Models in a Period of Economic Uncertainty

This thesis investigates the out-of-sample predictability of the SPDR Gold Shares ETF (GLD), used as a tradable proxy for gold prices, over a period characterized by substantial market uncertainty. The analysis distinguishes between two related but conceptually different forecasting tasks: predicting monthly price levels and modelling the conditional volatility and downside risk of daily returns. Monthly GLD prices from January 2016 to January 2026 are analyzed using Exponential Smoothing, ARIMA and a hybrid ARIMA-ANN model. Their forecasting performance is evaluated against a naïve random-walk benchmark using a common out-of-sample procedure and consistent accuracy metrics. Daily log-returns over the same period are modelled using a Student-t GARCH(1,1) specification with alternative conditional-mean structures. Volatility estimates are then assessed through residual diagnostics, rolling estimation and one-step-ahead Value-at-Risk backtesting. The price-forecasting models provide an adequate description of the historical trend but do not outperform the naïve benchmark. The ARIMA-ANN hybrid does not improve on the parsimonious ARIMA model, indicating that greater model complexity does not necessarily provide additional predictive accuracy. By contrast, the volatility analysis identifies highly persistent conditional variance and heavy-tailed returns. The number and sequence of Value-at- Risk violations are not statistically inconsistent with the nominal coverage level, although the limited backtest length and the parameter instability detected by the Nyblom test require caution. Overall, the findings indicate limited predictability of monthly GLD price movements, while conditional volatility and short-horizon risk display a more structured, partially modellable behavior. The results emphasize the importance of naïve benchmarks, out-of-sample validation and parsimonious specifications in financial forecasting.