Least-Squares SVM Forecasting with Gaussian Kernels and SOM Clustering
Summary
The article explains least-squares support-vector machines for time-series regression and describes how lagged observations form predictor vectors for forecasting. LS-SVM replaces a constrained nonlinear optimization with a linear system whose kernel matrix measures similarity between observations. The example focuses on a Gaussian radial-basis kernel, with its width and a regularization parameter tuned against test-set error. It also proposes using a self-organizing map to cluster input vectors and reduce the number of kernel representatives, addressing the quadratic scaling of the full matrix.
For lag selection, the author discusses partial autocorrelation and favors a consecutive window for broader applicability, while noting that the useful lags and tuned parameters can change with the sample. The method is implemented in MQL and illustrated with trading-oriented forecasts, but the excerpt offers no robust performance evidence. The author cautions that results depend on the instrument and timeframe, that historical-price-only forecasts may be weak for currencies, and that any trading use still requires risk controls and attention to news.
Key ideas
- LS-SVM forecasts a time series by regressing future values on vectors of lagged observations.
- A Gaussian kernel allows nonlinear relationships, while regularization and kernel width require empirical tuning.
- Partial autocorrelation can inform lag depth, but its pattern may change when the sample changes.
- A self-organizing map can replace many training vectors with cluster representatives to reduce matrix size.
- Forecast usefulness depends on the market and timeframe, and the article does not establish robust trading profitability.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.