Evaluating the Forecasting Performance of the TVP-VARModel for the Swedish Real Exchange Rate : Forecast Evaluation Against a Random Walk Benchmark over a Two-Year Horizon

This thesis applies Bayesian statistics with a time-varying parameter vector autoregression (TVP-VAR) model to forecast Sweden’s real exchange rate (RER). The model is estimated using quarterly macroeconomic data of six variables between 1995 and 2019. Estimation is conducted via Gibbs sampling, and posterior inference is used to evaluate stability and forecast uncertainty. The forecast setup is designed to both a rolling window and a recursive procedure that does not incorporate updated actual data after each forecast step. Forecast accuracy is assessed using RMSE, MAE, and their corresponding Theil’s U ratio. The TVP-VAR forecasts are evaluated against a random walk (RW) benchmark. While the TVP-VAR model captures time-varying dynamics and generally provides strong one-step-ahead forecasts, its performance does not outperform the benchmark. Limitations include computational cost and sensitivity to prior settings.