Treasury Bond Yields: A Neural Net Analysis
by John Kean
In this article, STOCKS & COMMODITIES contributor John Kean uses a neural network to look for
predictable patterns between Treasury bond yields and two of the driving forces behind the bond market,
inflation and government deficits.
Treasury bond yields are one of the most closely watched financial statistics. Aside from bond traders,
those in the stock, commodities and foreign exchange markets know that moves in the Treasury bonds
often set off intermarket waves that spread out like a stone thrown into a pond. Besides the financial
markets, the health of the economy as a whole is heavily affected by the level of long-term interest rates,
all of which closely track Treasury bond yields. Most people think of mortgage rates when they think
about long-term rates and the economy. But this is a prosaic consideration and pales certainly when
compared with the long-term competitive repercussions of hindering industrial investment with high
borrowing rates relative to other nations.
In this article, I will examine using a form of artificial intelligence, called neural net analysis, in Treasury
bond yield prediction. I will look at forecasting bonds from two standpoints: first, predicting month-end
Treasury bond yields for trading purposes, one month ahead, and second, predicting changes in Treasury
bond real yields (meaning yield minus inflation), 10 months ahead.
The concept behind neural networks was originally brought forth almost 50 years ago and had its origin in studies of the mechanisms by which the brain learns. Over the years, the neural network concept went
through periods of favor and disfavor with researchers, based on both perceptions of its promise and
available hardware power. There has been a resurgence of interest in this field over the last decade, and
neural nets are now being applied in many fields including medicine, banking, general business,
engineering and science. This approach gives results that are superior to conventional methods of
statistical analysis and pattern recognition.
The type of neural network discussed here is a software imitation of the brain's learning process. The
basic learning unit of the brain is the neuron, of which there are many varieties. When neurons are
activated, they interact with each other by releasing electrochemical packets across the synaptic gaps
between them, altering the internal activity level of the neurons. The learning process comes about
through the modification of these connections between neurons.