Improving Exponential Smoothing Forecasts with Multi-Step and Adaptive Fitting
Summary
The article develops modifications to a damped linear growth exponential smoothing model for financial time series. First, it changes parameter fitting from minimizing one-step forecast error to minimizing the combined squared errors for one-, two-, and three-step forecasts. On twelve currency and dollar-index test series, averaged across overlapping sections, this change slightly lowers the reported two- and three-step MAPE and the combined error, while the one-step error rises slightly. The evidence is limited to the specified historical datasets and short forecast horizons.
The next modification introduces time-varying smoothing coefficients using smooth transition exponential smoothing, aiming to adapt to changes in the input series. The article evaluates variants using forecast errors and discusses confidence intervals, emphasizing that long-horizon forecasts are unreliable and that nearer horizons deserve more attention. It presents the work as incremental indicator improvement, not a solution to currency forecasting. The supplied text is truncated during the adaptive-method discussion, so its final detailed results and procedures are not fully available.
Key ideas
- Fitting smoothing parameters to multiple forecast horizons can shift error performance across horizons.
- The modified objective combines squared one-, two-, and three-step forecast errors.
- On the stated test sequences, two- and three-step MAPE improve slightly while one-step MAPE increases slightly.
- Smooth transition exponential smoothing varies coefficients to adapt to changes in the series.
- Longer forecast horizons remain difficult, and the reported results are limited to the selected historical data.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.