Algorithm Research On Financial Time Series Data Generation

The development of big data era requests researchers paying much attention to the value of data itself.However,Data shortage due to the limitations of environment,practice,cost or algorithm seen everywhere.Data Augment is one of the most generation method stimulated an idea of conducting algorithm optimization from handed dataset.Existing generation algorithms have many limitations for characterizing them complex features,make them often have no good generalization for financial time series especially.Moreover,an effective evaluation method for generating data which is separated from subjective judgment needs to be put forward.Focus on the generation algorithm and its efficiency of financial time series data,this paper developed two aspects of work:(1)Using Analysis-Synthesis model to fit window distribution and generation from distribution sampling firstly,which effectively uses cluster features as the guideline market characteristics.Considering the potential effectiveness of frequency-domain features,this paper proposes an optimized FFT-AS generation algorithm through an adaptive FFT feature extraction and DC algorithm cluster discovery process.For effect identification,Maximum Mean Discrepancy and econometrics characteristics are introduced for preliminarily exploring effectiveness,and a horizontal comparison with the autoregressive model conducted.It is indicated FFT-AS fitted from frequency features is much better than AS or autoregression classes.(2)This paper presents a targeted GAN or CGAN driven by kinds of network optimizations and an innovative algorithm for Maximum Mean Discrepancy(MMD)parameters searching to promote the long-term dependency and targeted generator's distribution learning.A generalization evaluation method named TSTR proposed and some ML classification methods chosen to algorithms comparison.Experiments show that CGAN and GAN can learn the potential distribution of data better than other methods.And the generated data is not a simple restoration of real data by the reconstruction loss training algorithm and K-S test for generated data viability.In summary,this paper firstly improves traditional AS model through introducing domain features from its basic though and extension.Then presents a better GAN generation framework for financial time series than AS or other traditional econometric models depend advantages on loose conditional assumptions and directive generator distribution constraint training,also its reflection on whole market characteristics;And two better generalization evaluation methods suitable for generation algorithm proposed which could avoid subjective decision-making,identify generation data survivability and take into account downstream research tasks.