Unformatted text preview: tput range. This augurs well for a good performance in the future,
provided that care is taken to normalize the data so that a sudden increase in the index
value will not saturate the normalized value.
Fig. 11: Predicted and desired S&P 500 index
Max. % error = 4.044%
Avg. % error = 0.95% S&P 500 index 620
540 Desired value
P redicted value 520 Correct trend - 43 out of 50 times 500
0 10 20
Data number 40 50 13 It can be argued that the network can be trained every week rather than keeping it based
on the training, which will be very old near the end of the 50th week. Moving the training
window every week and retraining the network is a valid approach, which might be
necessary in practice. However, there is a danger of the network training on the noise,
inherent in the weekly changes and hence, performing worse than this network. In any
case, this procedure can be modified suitably and the prediction window can also be
reduced to suit the requirements.
Case 2: Crash of October, 19...
View Full Document
- Spring '12
- Neural Networks, Artificial neural network, Stock market index, neural network