Abstract
We propose a dynamical description of financial time series capable of making short-term prediction utilizing support vector regression on neighbourhood points. We include in our analysis estimation on the uncertainty by capturing the exogenous from historical prediction errors and adopting a probabilistic description of the prediction. Evidences from a series of backtesting using financial time series indicate that our model provides accurate description of real market data comparable with GARCH(1,1).
Notes
1 An embedding is a smooth, one-to-one transformation with a smooth inverse.
2 Supremum norm ||Xr−Xs||sup=max(|Xr1−Xs1|,…,|Xrd−Xsd|) represents the size of a hypercube enclosing Xr and Xs.
3 According to CitationGrassberger and Procaccia (1983), local scale ɛ can be defined through the region of small norm size where the correlation integral Cd(ɛ) satisfies a power law relation Cd(ɛ)αɛν with constant ν. In the reconstructed space, correlation integral Cd(ɛ) can be calculated as the probability of finding two d-histories that are less than ɛ apart. Intuitively, local scale must also be at least an order of magnitude below the average norm size over all reconstructed vectors.
4 See CitationSmola and Scholkopf (1998) for a complete discussion on SMO. We have developed and used here a modified version of the algorithm based on the discussion in CitationShevade et al (1999).
5 Another common choice of kernel function in SVR is the Gaussian function K(Xq,Xs)=exp(−||Xq−Xs||2/ω) on Euclidean norm with parameter ω. It tends to yield good performance under a general smoothness assumption on the data. The set of basis functions for Gaussian kernel is infinite and form a basis for the Hilbert space of bounded continuous functions. In our work, the performance of a Gaussian kernel is very similar to that of a polynomial kernel for financial time series.
6 In GARCH(1,1) with Gaussian innovation, we define xt=μ+νtɛt where ɛt is identically and independently distributed as a standard normal distribution. The variance measures are generated through the iteration equation νt+12=aνt2+b(xt−μ)2+c where a>0, b>0, and c>0 with a+b<1 are estimated based on maximizing the likelihood function given the sequence {xt}t=1n prior to the backtesting. For simplicity, the unconditional mean μ can be estimated by the historical mean as μ=E(xt)≅(1/n)∑t=1nxt.