When it comes to buying and selling foreign currency, companies want to seek out the best possible execution prices. They can achieve this by combining a number of liquidity providers (LPs) into what is known as an ‘aggregator’ (which is somewhat similar to looking for house insurance on a website and aggregating all available providers instead of considering just one). This should in principle lead to improved liquidity, a narrower spread (i.e. the difference between buy and sell prices, usually referred to as bid and ask) and, therefore, a better price.
In FX, however, a larger number of LPs does not always lead to a reduction in execution costs, due to various specificities of the market, such as ‘last look’ (i.e. when LPs have the final say on whether to accept or reject the order within a set period of time) and the recycling of liquidity (i.e. when LP A uses prices supplied by LP B in an attempt to profit from widening their spread, without supplying additional liquidity of its own). Hence, the question of which LPs should make up an aggregator ought to be one of quality over quantity, with the typical criteria to consider being: spread, fill ratio, costs of rejections, last look hold time and market impact.
We argue that an additional metric, namely ‘correlation in liquidity’, is useful for determining whether an LP should be part of an aggregator. In particular, by minimising the average correlation between LPs’ mid prices (i.e. the midpoint between bid and ask), companies can diversify their liquidity, in a similar manner to diversifying risk of a portfolio by including uncorrelated assets, and achieve, with a smaller number of LPs, a narrower inside spread (i.e. the average difference between minimum ask and maximum bid, available in an aggregator at the same time).
Our analysis of mid prices for G7 currency pairs from 11 venues shows that including LPs with the smallest average correlation into an aggregator one after another, leads to a faster reduction in the inside spread than when they are added randomly.