NEW YORK—Algorithms used by researchers to identify transaction-level activity in the federal funds market are ill-suited for this task, according to a new study from the Federal Reserve Bank of New York.
The U.S. federal funds markets is an interbank market for unsecured, mostly overnight loans of reserves held by banks at the Federal Reserve Banks. The traditional source of data on the fed funds market is based on fed funds trades reported by the major fed funds brokers to the New York Fed. However, an alternative source of funds data, used exclusively to conduct academic research, utilizes algorithms that search all transactions sent over Fedwire to identify pairs of payments that look like fed funds loans.
In “Challenges in Identifying Interbank Loans,” New York Fed economists Olivier Armantier and Adam Copeland formally test how accurate these algorithms are in identifying individual fed fund transactions. To accomplish this objective, the authors compare a set of payments known to be fed funds transactions from 2007-2011 with the set of payments pegged as such by an algorithm. The authors find that, for the period, an average of 81 percent of all pairs of payments identified by the algorithm are not, in fact, fed funds transactions conducted by the two banks, while an average of 23 percent of the banks’ actual fed funds transactions go unrecognized by the algorithm.
These results, the authors conclude, raise serious concerns about the appropriateness of using the algorithm’s output to study the fed funds market. As a consequence, the findings raise questions about the validity of empirical results previously obtained using the algorithm’s output. Finally, the analysis underscores the need to demonstrate formally, prior to any analysis, that the indirect inferences produced by an algorithm are accurate.
Olivier Armantier is an assistant vice president and Adam Copeland a research officer at the Federal Reserve Bank of New York.