Exchange rate determination and inter-market order flow effects

Abstract

The dependence of foreign exchange rates on order flow is investigated for four major exchange rate pairs, EUR/USD, EUR/GBP, GBP/USD and USD/JPY, across sampling frequencies ranging from 5 min to 1 week. Strong explanatory power is discovered for all sampling frequencies. We also uncover cross-market order flow effects, e.g. GBP exchange rates are very strongly influenced by EUR/USD order flow. We proceed to investigate the predictive power of order flow for exchange rate changes, and it is shown that the order flow specifications reduce RMSEs relative to a random walk for all exchange rates at high-frequencies and for EUR/USD and USD/JPY at lower sampling frequencies.

JEL Classifications: :

• F0
• F3
• G1

Keywords:

• exchange rate determination
• order flow

Acknowledgements

Thanks to Charles Goodhart, Richard Lyons, the editors, an anonymous referee and participants in the 2002 Venice Summer Institute for excellent comments. Thanks also to the Bank of England for providing the transactions data used in this study. All remaining errors are our own.

Notes

See e.g. Clark (1973), Epps and Epps (1976), Tauchen and Pitts (1983) and Karpoff (1987).

Note that in defining order flow one must distinguish between buyer and seller initiated transactions. Of course every trade consummated in a market has both a buyer and a seller, but from the current perspective the important member of this pair is the aggressive trader, the individual actively wishing to transact at another agent's prices.

See also Hasbrouck (1991) and Madhavan and Smidt (1991) who study equity markets, and Cohen and Shin (2002) who study fixed income markets.

Similarly, Chordia, Roll, and Subrahmanyam (2001) show that daily changes in US equity market levels are strongly related to market wide order flow measures.

To compare the size of these markets, according to the Bank for International Settlements (2002) in April 2001 the EUR/USD represented 30% of all spot FX trading, the JPY/USD 21%, GBP/USD 7% and GBP/EUR 3%. The first three of these are the three largest currency pairs while GBP/EUR is only the eighth.

For a full description of the segments of the spot FX market and the data available from each see the excellent descriptions contained in Lyons (2001).

In this article we define the overnight period to be 18:00 to 06:00 the following day. It should be noted that this definition is only proper for traders in London and New York, but not for traders in Asian markets. It corresponds to the portion of the day when trade on D2000-2 is least intensive, even for USD/JPY.

We have experimented with denser time aggregation levels and the results do not alter the pattern we report in this article.

Since the normality of our return data is rejected by the Jarque-Bera test (not reported), we also experimented with a LAD estimator for these regressions, but the results were not qualitatively affected.

Note that our definition of the aggregation time interval is slightly different from that in Evans and Lyons (2002b). Whilst their ‘daily’ aggregation interval is defined as a period from 16:00 to 16:00 next day our definition is a period from 06:00 to 18:00 excluding overnight period. We also experimented with an interval definition that includes the overnight period and find results that do not differ qualitatively from those reported here.

We have performed multi-step forecast analysis but it added little new information to the results we present here. Results are available on request.