Using real-time cluster configurations of streaming asynchronous features as online state descriptors in financial markets

We present a scheme for online, unsupervised state discovery and detection from streaming, multifeatured, asynchronous data in high-frequency financial markets. Online feature correlations are computed using an unbiased, lossless Fourier estimator. A high-speed maximum likelihood clustering algorithm is then used to find the feature cluster configuration which best explains the structure in the correlation matrix. We conjecture that this feature configuration is a candidate descriptor for the temporal state of the system. Using a simple cluster configuration similarity metric, we are able to enumerate the state space based on prevailing feature configurations. The proposed state representation removes the need for human-driven data pre-processing for state attribute specification, allowing a learning agent to find structure in streaming data, discern changes in the system, enumerate its perceived state space and learn suitable action-selection policies.