Forecasting S&P 500 spikes: an SVM approach

In this study, we focus on forecasting long-tail events of the S&P 500 stock returns. The S&P 500 is widely considered as a bellwether for the overall US economy as it encompasses some of the largest—in terms of capitalization—corporations from both the NYSE and the NASDAQ stock exchanges. A timely and efficient forecast of such extreme changes is of great importance to market participants and policy makers, since they may trigger large scale selling or buying strategies that may significantly impact the specific market and the overall economy. We define as “spikes” the events where we have extreme upward or downward changes of the S&P 500 index; in our case, we use the returns that fall outside a two-standard deviations band. However, instead of simply using the unconditional overall standard deviation, in this paper we employ a GARCH (p,q) model to derive the conditional standard deviation of the returns. This is a more appropriate measure of immediate risk to market participants than the overall series’ unconditional standard deviation. Traditional forecasting models that rely on statistical analysis and traditional econometrics, assume that the returns follow some typical underlying distribution. These models usually fail to successfully and efficiently accommodate price spikes especially when it comes to forecasting. In our study, we use the atheoretical and data-driven Support Vector Machines methodology from the area of Machine Learning. This forecasting approach does not require any initial assumptions on the distribution of the data but rather exploits patterns that may be inhibited in the initial data space. These patterns may become more apparent and exploitable in the resulting feature space. We use 1860 daily observations from 01/01/2009 to 22/01/2017. Our overall optimum forecasting model achieved a 70.69% forecasting accuracy for the spikes and 73.25% for non-spikes.