Uncovering Hidden Structure in Bond Futures Trading

Abstract

This study uncovers trading styles in the transaction records of US Treasury bond futures.It uses transaction-by-transaction data from the Commodity Futures Trading Commissions' (CFTC)Computerized Trade Reconstruction (CTR) records. The data set consists of 30 million transaction - thecomplete US T-bond futures market for 3 years. Each transaction record consists of time (by theminute), price, volume, buy/sell, and an identifier of the specific account.We use statistical clustering techniques to group together trades that are similar. Two sets ofassumptions have to be made: (1) What is a trade? We define a trade to begin when an account opensa position, and to end when its position size returns to zero. We describe each trade by several trade-specificvariables (e.g., length of trade, maximum position size, opening move, long or short) and severalexogenous, market-specific variables (e.g., price, volatility, trading volume). (2) What process generatedthe data? We assume a mixture of Gaussians. An observed trade is interpreted as a noisy realization ofone of the mixture components. This paper assumes identity covariance matrices. Furthermore, eachtrade is fully assigned to a single cluster. We compare this approach to diagonal and to full covariancestructure with probabilistic assignments.Trade profit was held back in the clustering process. It turns out that the clusters differ significantlyin their profit and risk characteristics. Using conditional distributions, we summarize featuresof profitable trading styles and contrast them with losing strategies. We find that profitable styles tendto hold trades longer, trade at higher volatility, and trade earlier in the contracts. We also show howsome clusters uncover "technical" traders. Using the information about the individual accounts, theassignments of accounts to clusters are described by entropy, and the transitions of a given accountthrough clusters is modeled by a first order Markov model.