ON SERIES LENGTH AND FORECAST PERFORMANCE OF SUPPORT VECTOR MACHINE

This article sought to study the effects of series length on forecast accuracy when applying support vector machines. Data were on rainfall volume of Abuja, Nigeria (January 1981-December 2021), CPI (January 1995-December 2022) and opening Bitcoin prices (1373 observations). Each series was divided into scenarios. Rainfall series had three scenarios while CPI and Bitcoin data had four scenarios each. Twelve last observations of each series were set aside for validation and 90% of observations were used for training. Modeling was done using support vector regression, a component of support vector machines. The study found that the shortest series, corresponding to Scenario 1 for each series typically produced the most accurate forecasts based on mean square error (MSE) for all three lead times considered. For each scenario, higher lead times typically recorded higher MSE. Comparison based on sum of ranks obtained on the basis of absolute error corroborated the position of MSE. Although the shortest series typically produced the most accurate results, findings have not convincingly established that the longer the series, the worse the forecasts. Combined forecasts were brought in to ascertain whether combining forecasts from different scenarios of a particular series could be more accurate than single forecasts. Although combined forecasts have justified their inclusion in the comparison by outperforming single forecasts on many occasions, the fact that they have not typically outperformed the shortest series does not justify their preference as their adoption means wasting resources and jeopardizing accuracy simultaneously. Generally, short series are better in forecast accuracy for short-term and long-term forecasting. Future research efforts can delve into other machine learning models. The inconsistent behavior of forecast accuracy associated with increasing series length calls for inclusion of higher number of scenarios that may require longer series in future studies.