The motivation for the research presented in this thesis stems from the recentavailability of high frequency limit order book data, relative scarcity of studiesemploying such data, economic significance of transaction costs management, anda perceived potential of data mining for uncovering patterns and relationships notidentified by the traditional top-down modelling approach. We analyse and buildcomputational models for order submissions on the Australian Stock Exchange,an order-driven market with a public electronic limit order book. The focus ofthe thesis is on the trade implementation problem faced by a trader who wantsto transact a buy or sell order of a certain size.We use two approaches to build our models, top-down and bottom-up. Thetraditional, top-down approach is applied to develop an optimal order submission plan for an order which is too large to be traded immediately without aprohibitive price impact. We present an optimisation framework and some solutions for non-stationary and non-linear price impact and price impact risk. We find that our proposed transaction costs model produces fairly good forecasts ofthe variance of the execution shortfall. The second, bottom-up, or data mining,approach is employed for trade sign inference, where trade sign is defined as theside which initiates both a trade and the market order that triggered the trade.We are interested in an endogenous component of the order flow, as evidenced bythe predictable relationship between trade sign and the variables used to infer it.We want to discover the rules which govern the trade sign, and establish a connection between them and two empirically observed regularities in market ordersubmissions, competition for order execution and transaction cost minimisation.To achieve the above aims we first use exploratory analysis of trade and limitorder book data. In particular, we conduct unsupervised clustering with the self-organising map technique. The visualisation of the transformed data reveals that buyer-initiated and seller-initiated trades form two distinct clusters. We thenpropose a local non-parametric trade sign inference model based on the k-nearest-neighbour classifier. The best k-nearest-neighbour classifier constructed by us requires only three predictor variables and achieves an average out-of-sampleaccuracy of 71.40% (SD=4.01%)1, across all of the tested stocks. The best set ofpredictor variables found for the non-parametric model is subsequently used todevelop a piecewise linear trade sign model. That model proves superior to the k-nearest-neighbour classifier, and achieves an average out-of-sample classificationaccuracy of 74.38% (SD=4.25%). The result is statistically significant, afteradjusting for multiple comparisons.The overall classification performance of the piecewise linear model indicatesa strong dependence between trade sign and the three predictor variables, andprovides evidence for the endogenous component in the order flow. Moreover,the rules for trade sign classification derived from the structure of the piecewiselinear model reflect the two regularities observed in market order submissions,competition for order execution and transaction cost minimisation, and offer newinsights into the relationship between them. The obtained results confirm theapplicability and relevance of data mining for the analysis and modelling of stockmarket order submissions.
