Discovering stock dynamics through multidimensional volatility phases

To investigate stock dynamics, we consider volatility as a temporal aggregation of semi-extreme events defined on three dimensions: return, volume and trading number. Onset and offset phases of volatility are computed by means of the hierarchical factor segmentation (HFS) algorithm based on high-frequency data. Through these computed volatility phases we search for dynamic patterns by resolving two questions: Is a return's volatility closely associated with significant price changes?, and can we derive an early prediction of the sign of the price change at the offset? Can volatility phases reveal which dimension—return, volume or trading number—is the driving force behind the others? Some computed new features of stock dynamics are counter-intuitive. Almost all significant price changes are marked by volatility within the three dimensions. We develop a data-driven potential-based model to make early predictions of the sign of significant price differences at the end of a volatile period. This model recognizes that when a stock's dynamic enters a volatility state, it typically settles into a subtle imbalance of oscillations between positive and negative returns, and leads to a significant price difference at the offset of volatility. We develop a new statistical analysis to show that a return's volatility onset is more likely to fall behind the onsets of volume and trading number, while the latter two dimensions are very well-correlated with each other. By incorporating this result with behavioral evidence extracted from scatterplots of the logarithm of volume versus the trading number, we postulate that stock dynamics are chiefly driven by a large group of participants, whose collective large-volume trading action is potentially responsible for stimulating volatility in both return and trading number.