Application of self-organizing maps to clustering of high-frequency Financial data

This paper analyzes the clustering of trades on the Australian Stock Exchange (ASX) with respect to the trade direction variable. The ASX is a limit order market operating an electronic limit order book. The order book consists of buy limit orders (bids) and sell limit orders (asks). A trade takes place if a new order arrives which matches an existing order in the limit order book. If the matched order is a bid (ask) then the trade is considered to be seller(buyer)initiated and the trade direction variable assumes a corresponding value. We employed self-organizing maps (SOMs) to perform unsupervised clustering and visualization of four dimensional trade level data for the ten stocks on the ASX with the largest market capitalization. Trade size, the best bid and ask volumes, and a variable capturing previous trade directions were used as input variables. The visualization of the data using the SOM transformation reveals that buyer-initiated and seller-initiated trades form two distinct clusters in correspondence with nonequilibrium market conditions and elicits the main structural features of the clusters.