Dynamics of Hierarchical Clustering in Stocks Market During Financial Crises

AbstractWe examined the behaviors of stocks in the Stock Exchange of Thailand (SET) during financial crises from 2008 to 2020 by using a variety of indicators such as average entropy, average correlation, and average Fisher distance. All of these indicators increased dramatically during a financial crisis, which agreed with the Ulcer index indicating the root mean square of percentage price drawdown. We construct dynamics of the hierarchical tree structure of 37 stocks in SET, using the correlation distance and the Fisher distance between stocks in different time windows and compared the results. As the hierarchical clustering evolves over time by varying the threshold of inconsistency, the variation of information shows that during a period of a financial crisis, the structure of hierarchical clustering has a small tendency to change their cluster than usual, as evidenced by their anticorrelation. The dynamical Fisher distance is more robust, can provide earlier information, and yields a much-slower decay of signal-to-noise ratio as the averaging window size increases. The dynamic hierarchical using the Fisher distance gives less-variant clustering and can shed more insights into stock behaviors near or during a financial crisis.Highlights• Although the Thailand stock market is relatively small, it still exhibits the stylized facts discovered in larger financial markets.• We demonstrate that the Fisher distance allows ones to investigate dynamic hierarchical clustering among stocks, from which ones can probe the market behavior more distinctly. The visibility signals from the Fisher information does not decay as fast as that from the correlation or entropy.• The Fisher distance can provide potentially earlier signal for changes in stock volatility in Thai Stock Exchange.• During the financial crises, the hierarchical clustering tends to keep its structure more intact than during calm periods. The hierarchical clustering using the Fisher distance yields more robust and well-defined clusters than that using the correlation distance