A Direct Test of Methods for Inferring Trade Direction from Intra-Day Data
利用纽约证券交易所TORQ数据库，直接检验了多种从日内报价和交易价格推断买卖方向的方法的准确性，发现Lee和Ready算法与Tick检验表现相近但均低于预期，且使用这些方法会导致有效价差和符号成交量估计出现显著偏差。
This study directly tests the ability of several competing methods to identify market buy and sell orders using intra-day quote and trade prices, and identifies factors that affect the accuracy of the methods. Lee and Ready's (1991) algorithm performs about the same as the tick test, but the performance of both methods is worse than expected. The results show that the use of either algorithm to classify trades can lead to significantly biased estimates of effective spreads and signed volume, but the tick test provides better estimates of effective spreads and signed volume than Lee and Ready's method. I. Introduction The use of intra-day prices in empirical studies of securities markets is increasingly common and studies frequently require trades be identified as buyer or seller initiated. Unfortunately, most data sets do not identify trade direction. Methods have, however, been proposed that allow trade direction to be inferred from adjacent prices and quotes. The accuracy of these methods and the implica? tions for microstructure research are still unresolved issues in large part because trade direction is unobservable in most financial data sets. This study directly tests the ability of several competing methods to iden? tify market buy and sell orders using intra-day quote and trade prices. The tests are conducted using the TORQ database, a unique data set the New York Stock Exchange makes available to researchers that contains information on trades, quotes, and orders. Using the tick tests and Lee and Ready's (LR hereafter) (1991) method to classify trades as buys or sells, and comparing the results to the direction of the actual orders, I directly test the accuracy of the classification algorithms. The tests also identify the factors that affect the accuracy of the clas? sification methods. Additional tests demonstrate that using these algorithms can lead to biased inferences in two of their most common applications: estimating effective spreads and signed volume trading. Test results also show that the tick