DYNAMIC PREDICTOR SELECTION AND ORDER SPLITTING IN A LIMIT ORDER MARKET

Abstract

Recent empirical research has documented the clustered volatility and fat tails of return distribution in stock markets, yet returns are uncorrelated over time. Certain agent-based theoretical models attempt to explain the empirical features in terms of investors' order-splitting or dynamic switching strategies, both of which are frequently used by actual stock investors. However, little theoretical research has discriminated among the behavioral assumptions within a model and compared the impacts of the assumptions on the empirical features. Nor has the research simultaneously replicated the return features and empirical features on market microstructure, such as patterns of order choice. This study constructs an artificial limit order market in which investors split orders into small pieces or use fundamental and trend-following predictors interchangeably over time. We demonstrate that, on one hand, the market that features strategies with order splitting and dynamic predictor selection can independently replicate clustered volatility and fat tails with near-zero return autocorrelations. However, we also show that patterns of order choice do not match those found in certain previous empirical studies in both types of economies. Thus, we conclude that, in reality, the two strategies can work to generate the empirical return features but that investors may also use other strategies in actual stock markets. We also demonstrate that the impact of both strategies on the volatility persistence tends to be greater as the number of traders increases in the market; this finding implies that the order-splitting strategy and dynamic predictor selection are more crucial for the empirical phenomena pertaining to larger capital stocks.