Price impact estimation of bond market :|a machine learning approach

Abstract

We conduct an empirical study to analyze the aggregate price impact behavior of Turkish government bonds over 17 years. Moreover, we want to find out the determinants of price impact for each trade. The primary motivation is to provide better guidance to high-frequency traders, fund managers, and institutional traders since generating alpha mostly depends on transaction costs when assets tend to perform similarly. We characterize the order and transaction records of Turkish government bonds in detail. We find that non-residents’ positive portfolio flow and AUM in the Turkish private pension market are associated with lower aggregate price impact. In contrast, divergence from other emerging markets drives costs up. We present how price impact varies across trade size, market conditions, bond characteristics, and the state of the limit order book to compare our model with various liquidity measures proposed in the literature. Except for Kyle (1985)’s price impact lambda, our model subsumes the information that other liquidity measures provide. We confirm the square-root law by figuring a concave relationship between trade size and price impact. Finally, we examine how price impact forecasts can be improved via recent Machine Learning Techniques. We show that boosted trees with XGBoost outperforms other alternative models, i.e., random forest and elastic net methods, in out-of-sample estimations. We also find that the prevailing slope of the limit order book, maturity of bonds, and market volatility appear to be the top three factors affecting the slippage. Given the study’s novel features, we aim to contribute price impact literature in employing machine learning techniques to long-term order book data.