Comment: Meantime the later mentioned network
runs an excellent performance outperforming RexP in almost every year (see
performance).
Artificial Intelligence
Interest Rate Forecasts with a Neural Network
Neural Networks for Interest Rate Forecasts of German Bonds now Operational
Applications of artificial intelligence make more and more inroads in areas that have been previously reserved for highly experienced specialists. Particularly dramatic are the developments in neuro computing. Much progress has been made so that above average results are beeing achieved in interest rate forecasts and portfolio management.
Traditionally financial forecasts have been calculated on the basis of econometric models. Such models frequently have thousands of equations whose parameters are estimated on the basis of historic data. The quality of the forecasts delivered by such models depends to a large extent on the knowledge built in by the architects.
A breakthrough in interest rate forecasts is achieved by learning systems that find out by themselves - Backed by a huge databasism - which data are important and how to process such data. The greatest progress is provided by neural networks.
Neural networks have first been developed in medicine to describe the funktions of the human brain. The 'perceptron' suggested by Rosenblatt in 1958 fascinated the medical and the computer world alike as it was hoped to solve many problems.
Todays neural networks are generally much further developed. They still consist of many interlinked neurons. Those are cells, that receive, process and broadcast impulses.
Like the human brain those cells process the impulses in a specific
way. Outgoing signals are fired only if incoming weighted signals surpass
a certain threshold. This way all kinds of logic connections for example
"and" or "or" can be realized.
By this "threshold-logig" neural networks are far superior to the forecast techniques presently used. This becomes evident as traditional econometric models can be seen as a special case of neural networks. Therefore neural networks offer more possibilities for structuring.
The speciality is the far reaching ability to learn logic relations of data. One example is the influence of monetary growth and capacity utilization on the level of interest rates. After training with historical data the neural network can make a judgement on their significance and logic relationship. It will for example recognize that lower interest rates will occur if money supply growth is high and capacity utilization is low, whereas interest rates will rise when money supply grows strongly and capacity utilization is high.
Learning is thereby represented by an extensive mathematical search, whereby the network is beeing adapted accordingly.
This procedure is well suited for interest rate forecasts and in contrary to presently usesd expert systems the knowledge does not have to be known a priori. The neural network is confronted with historic data in a training phase. If a sufficient number of patterns and their interest rate consequences are fed into the system the network can recognize such patterns with the assistance of the built-in learning process after some time and indicate which changes in interest rates are likely to occur.
The patterns can be very complex. For interest rate forecasts they contain mainly economic data such as rates of inflation, commodity prices, balances of payments, orderbook levels, exchange rates, data of other countries and the movements of yield curves. During the learning process the less important parts of the pattern are suppressed.
With the help of such data a considerable level of future interest fluctuations can be forecast. Even the reactions of market participants to sudden and unforeseen events such as war, carry a timelag and in the meantime these developements reflect themselves in calculable data such as oil prices and can be taken into account.
On basis of the stored experience the network is in a position to forecast quite accurately. To this end the network has to go through a second phase. It is confronted with the actual pattern, whose consequences it does not know. The various impulses such as rates of inflation and money supply are stored and the neurons fire according to their stored knowledge. If the new pattern is similar to the previously experienced situations, the result will be similar to past results. Even in case of a pattern not seen before the neural network can make reliable forecasts.
While working with a neural network it becomes evident, that the learning process must be shortend by a preselection of relevant data. Thereby already known relations can be implemented. In the following learning process these relations will either survive or will be displaced by better relations. In the search of such knowledge one can see that scientific theories are often valid, but that the kind of measuring and processing is to be specified. The data for the training process are selected from several international statisitical databases which all together contain more than 300.000 time series. As the qualitiy of forecasting increases with the amount of experience, the selected database contains the most important economic and financial data from several countries of the last 30 years. It is obvious that a system with such experience in its memory is on principle able to outperform human decision makers.
With the aim of forecasting interest rates a special learning process was developed in which proved statistical and econometric procedures are incorporated. These procedures guaranty, that once recognized relations cannot be eliminated by controversial new training data. All data gain their importance equivalent to their influence seen in the past.
The ex-post interest rate forecasts of a neural network were connected with a decision making system, which invests an amount of DM 1000 either in one month time deposit or in ten year government bonds.
From 1970 to 1990 the periods of rising bond prices were mostly recognized in the previous month and the price increase could be used to achive capital gains in the bond position. When bond prices were declining losses could usualy be avoided by investing the money in fixed time deposits. Finally the originally invested amount of DM 1.000 grew to more than DM 20.000 in the period of 20 years which corresponds to an annual yield of about 16% a year taking into consideration transaction fees of 0,25%. By investing exclusively in one month time deposits the amount would have grown to DM 3.300.
In 1992 the real portfoliomanagement with the neural network resulted in a profit of 14,08% including all fees and transaktion costs - beating the index which gained 13,4 %.