Abstract

In foreign exchange (FX) markets, participants can execute transactions in a variety of different ways. One particularly prominent protocol is referred to as an ‘RFQ’ or Request for Quote.

In this scenario, a customer asks for a price in a currency pair and notional, and a dealer responds with a quote, which is either accepted (a trade happens) or declined (no trade happens) by a customer. The empirical frequency of acceptance is referred to as ‘RFQ hit ratio’, and is widely recognised as a key performance indicator of any electronic FX business operation. It is therefore essential to understand what factors may influence RFQ acceptance, to what extent and how they do so. Electronic markets generate large volumes of high frequency data which can be used to help tackle this problem, meaning that statistical leaning techniques naturally present themselves as promising tools. The nature of RFQ workflow in particular, with its binary outcome for a given customer request and dealer response, make it amenable for analysis in a classification setting. This setting has been extensively studied over many decades by scholars and practitioners alike. We look at theoretical and practical aspects of building a statistical classifier of FX RFQ acceptance, with a particular emphasis on out-of-sample predictive ability by feature type and fitting method.