Semiparametric modeling of competing risks in a limit order market

This dissertation is based on three papers that have come out of the research conducted at Yale University during 1999–2001 and finished in Indiana University during 2002.; Chapter two introduces the competing risks methodology as an empirical tool for modeling high-frequency financial data in continuous time. The competing risks are applied to the analysis of the timing and interaction between the Deutsche Mark/U.S. dollar quotes and transactions in the Reuters D2000-2 electronic brokerage system. Estimation of the model mostly supports the empirical evidence from previous research on electronic limit order markets. In particular, the composition of order flow is found to be sensitive to the state of the limit order book and the trading history. The direction of past trade is found to have strong predictive power for the future market activity. The model detects an adverse information effect due to non-trading as the traders submit and cancel their orders most aggressively immediately after the limit order book events.; Chapter three studies the problem of semiparametric hazard rate estimation in the competing risks environment. Special attention is paid to the situation where the sample of observed durations is highly skewed, which is fairly common for high-frequency financial data. The chapter provides a review of large sample properties of alternative k-nearest neighbor estimators and local linear smoothers. The asymptotic theory is applied to the problem of baseline hazard rate estimation for a large number of limit order book events.; Chapter four extends the results of the previous chapters. The set of covariates is expanded to include a broad range of limit order trading and liquidity characteristics. The cross-sectional and serial correlation of Cox regression residuals is captured by the past order flow and the counts of recent transactions. The principal component analysis applied to the covariate indices identifies five pervasive factors that explain a major portion of trading activity. The multifactor modification leads to substantial data compression, improves the goodness-of-fit, and boosts the short-term predictive power of the model relative to popular moving average-type forecasting rules. The competing risks methodology provides a valuable framework for understanding and forecasting the behavior of heterogeneous agents in a competitive market environment.