Unformatted text preview: the 1110th data point (January, 21st, 1994).
It would be interesting to see how the network performs for these worst case examples,
since, it would be expected to perform extremely well, when the training and testing data
is in the same range.
Fig. 1: S&P 500 index from 1972 through 1996
1986 600 S&P 500 index 500
200 Bull market
since, Jan 1994 100
0 200 400 600
Data number 1000 1200 To test the network in these critical regions, it is trained using data from about 4 years
prior to the crash of 1986 or the start of the 1994 bull run. For both these test cases, the
network used has one hidden layer with 7 hidden neurons and 19 input neurons (input data
plus a unity input for the offset). Weights are initialized to random values in the range [1,1]. The hidden and output layers use a sigmoidal activation function. While the offset for
the hidden layer is provided by the fixed unity input of the 19th input neuron, there is no 11 offset for the sigmoids in the output layer. These test cases and t...
View Full Document
This note was uploaded on 07/20/2012 for the course ECON 203 taught by Professor Girishdev during the Spring '12 term at Indian Institute of Technology, Kharagpur.
- Spring '12