Rolling ARIMA-GARCH Forecasts for EUR/USD Directional Trading

Summary

This article explains ARIMA models for forecasting a time series’ mean and GARCH models for its changing conditional variance, then combines them in a directional EUR/USD strategy. It fits models to a rolling window of daily log returns, selects ARIMA orders by Akaike information criterion, fits a GARCH(1,1) model to the residuals, and takes a one-day long or short position according to the forecast’s sign. The author also explores filtering trades by forecast magnitude and by a residual autocorrelation diagnostic.

The article describes a walk-forward backtest and reports that the strategy modestly outperformed buy and hold over the tested period. Filtering small forecasts appeared useful, while the Ljung-Box diagnostic added limited value in this dataset. These findings are tentative: transaction costs are excluded, model fitting can fail, and the selected window and model choices may affect results. The article suggests testing alternative volatility models, confidence-based filters, and ensembles, while emphasizing that the evidence does not establish a durable trading edge.

Key ideas

• ARIMA models the conditional mean using lagged observations and errors, while GARCH models time-varying conditional variance.
• The strategy selects ARIMA orders by AIC on a rolling window, then fits a GARCH model to the residuals.
• The sign of the next-day return forecast determines whether the strategy holds a long or short EUR/USD position.
• The article reports modest historical outperformance over buy and hold, without including transaction costs.
• Forecast magnitude filtering showed promise in the sample, whereas residual-fit filtering offered limited benefit.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.