Model selection for one‐day‐ahead AUD/USD, AUD/EUR forecasts

Literature on exchange rate forecasting often focuses on varied algorithms' prediction performances and is comparatively silent regarding lag length and model structure (linear or nonlinear) selection. However, model selection and lag lengths are important decision criteria for practitioners when undertaking technical analysis. This research explores the impact of varied lags and different model structures on one‐day‐ahead forecasts of AUD/USD and AUD/EUR exchange rates. Two linear (MLR and SVR with linear kernel) and three nonlinear (SVR with RBF kernel and three gamma values) data mining models are adopted, along with lag lengths of 1, 5 (1 trading week), 10 (2 trading weeks), 20 (1 month) or 120 (6 months) days for both non‐differenced and first‐order differenced time series. The investigation highlights that, irrespective of estimating the magnitude of future exchange rate or the future inter‐day changes in exchange rate, linear models outperform nonlinear models at lower lag levels (1 or 5 or 10 days). At higher lags (20 or 120 days) and for the first‐order differenced time series, nonlinear models show promising outcomes. The MLR model, which emphasises training error reduction, generally outperforms the SVR model, which emphasises generalisability. Nonlinear models may show some prediction biases, while linear models appear to show no specific bias. Thus, for technical analysis upon AUD/USD and AUD/EUR exchange rates, instead of using a large lag of historical prices, a linear model with a lag length of 1 can produce good one‐day‐ahead forecasts.