Forecasting T-Bill Rates With A Neural Net by Milam Aiken
Have you ever wondered how accurate the consensus estimates of interest rates were from the top US economists? How about how accurate those economists are in comparison with neural net technology? A recent Forbes magazine article noted that three-month Treasury bill futures more accurately predict interest rates a year in advance than do forecasts made through annual surveys of approximately 50 of the nation's top economists. In addition to being more accurate, these forecasts were more readily available because data on T-bill futures are published daily in many financial periodicals. The comparison between a market forecast and the expert forecast intrigued me, so I decided to use a neural network to forecast T-bill rates for a comparison of all three forecasts. Here's how a neural network to forecast these rates can be developed using two commercially available programs.
HOW CAN RATES BE PREDICTED?
Using the Business Cycle Indicators (BCI) software, I was able to view how different data series compared with three-month T-bill rates (BCI has more than 250 data series from which to choose). Many of the series appeared to have no correlation at all, were coincident or even lagging. Most of the financial data series (bonds, interest rates and so forth) were highly correlated but coincident (changing along with the T-bill yield) and had no predictive ability. However, the three leading economic indicator series and money supply (M1) often turned before T-bill yields, giving a high degree of predictive value. The leading economic indicators are the Department of Commerce Leading
Economic Indicator Composite Index (LEI), the Center for International Business Cycle Research Short Leading Composite Index, and the CIBCR Long Leading Composite Index.