There is a large literature focused on exchange rate forecasting. A common finding is that it is difficult to systematically predict exchange rate movements, consistent with the results of Meese and Rogoff (1983), who showed that it is difficult to beat a random walk forecast. This note heads down the same well-worn path, but assesses the usefulness of the Commitments of Traders (COT) positioning data for understanding and predicting exchange rate movements.
COT data has been used extensively in research assessing the impact of speculation in commodity markets, but there has been comparatively little recent research into the usefulness of this data for thinking about foreign exchange market developments. COT data also provides classifications of trading entities that allow the trading behaviour of different groups of traders (such as hedgers and speculators), and their respective market impact, to be examined. This note focuses on the positioning of speculative traders who are thought to express their beliefs of future currency movements through futures positions. The COT data is available at weekly frequency and aggregates holdings of futures in key US markets. The lag between the collection of the COT data and its publication may in fact limit its information content. For example, assuming markets are informationally efficient, new information is expected to be quickly incorporated into spot exchange rates and futures positions at the same time. Therefore, the data cannot provide insight as to whether exchange rates change contemporaneously as futures positions are opened or closed (in real-time).
The key question examined in this note is what the information content of COT positioning data is for major currencies, and at what horizon is positioning data best for forecasting exchange rate changes. As has been found by earlier studies, our results suggest that futures positioning data can help with interpretation of historical exchange rate changes, although its use as a predictor of exchange rates is limited. However, higher frequency positioning data, such as hourly or daily data, may in fact have predictive power.