FRIDAY SEMINAR
SPEAKER:
Professor Thomas Philips
NYU Tandon School of Engineering, Finance and Risk Engineering.
TITLE: “A Simple Real-Time Algorithm to Identify Turning Points in U.S. Business Cycles”
DAY & DATE & TIME:
Friday, 23rd August 2024
2:30 PM -3:30PM
VENUE: Seminar Hall, Department of Management Studies, IISc.
Abstract:
We describe a simple online (i.e. real-time) algorithm that identifies both recessions and expansions in the U.S. remarkably well, and with only two false alarm from 1960–2024. More often than not, the transitions it identifies are no more than a month removed from the officially determined dates that are published after a long and variable delay by the NBER’s Business Cycle Dating Committee. Its data requirements are minimal, and can be obtained from the Federal Reserve’s FRED database within the first week of each month. As of early August 2024, the algorithm suggests that the U.S. economy entered a recession in July 2024. In spite of its good performance at detecting business cycles, tests show that it has minimal implications for market timing. The generally good performance of the algorithm with minimal data raises a number of interesting questions about business cycle timing:
I have a number of questions about the application of Machine Learning to the identification of business cycles and list them here for exploration:
1. What is the most effective decision variable for the detection of business cycles, and why?
2. Can the identification of business cycles be cast as a problem in classification? If so, can it be addressed using logistic regression or classification trees?
3. How can different measures of unemployment be combined without exacerbating quantization noise?
4. Can a non–linear filter remove quantization noise without impacting detection speed and the rate of false alarms?
5. Can machine learning algorithms recognize the need for a robust measure of location such as the Hodges-Lehmann Mean?
6. How can appropriate conditioning information (e.g. the slope of the yield curve ) be identified and automatically incorporated into the decision rule?
7. Can some other aspect of the decision variable, or some other measure of unemployment, modulate the threshold as effectively as the slope of the yield curve?
8. Can the need for threshold modulation be eliminated? Can the algorithm be adapted to work with a constant threshold?