This time, let's try to apply the neural network on Shanghai Stock Exchange Index(000001.SS on Yahoo! Finance). The network has 2 hidden layers, the 1st one has 5 nodes, the 2nd one has 2 nodes. Output layer has 1 node.

After 5000 epoches training with 2400 sample data, the result is showing below.

It seems the pattern of chart is better than the turning time. Particularly, the backward generalization gets a very interesting result.

When I do the prediction, I do not use the price in the prediction period to adjust the network or as input any more. That is, I only use the price data in the training period to recognize the pattern, then apply the pattern backward or forward directly. So the prediction is fairly a long term prediction.

Click on the picture to enlarge it

After more tuning work on my very primary neural network to predict the Price of stock market, I got some interesting result on ^HSI. 3600 sample data were used to train the network this time, from Mar 9, 2005, backward 3600 days. Generalization was applied forward from Mar 10, 2005 to Aug 23, 2006.

It's a simple MLP network with 3-layers, the hidden layer has 10 nodes, and 1 node for output layer.

Here's the result:

Click on the picture to enlarge it

And the zoom-in view, the result is also applied to the future (From Mar 10, 2005 to Aug 23, 2006 on this picture)

This is the 2nd version of my very primary neural network to predict the Price of stock market on BlogTrader Platform. 800 sample data were used to train the network, and generalization was applied forward and backward.

Here's the result:

Click on the picture to enlarge it

I've designed the 1st version of my very primary neural network to predict the Trends and Turns (not the Price) of Stock market on BlogTrader Platform. 400 sample data were used to train the network, and generalization was applied forward and backward.

Here's predicted result(very primary):

Click on the picture to enlarge it

I'm writing the Neural Network module for BlogTrader Platform. The first implementation is a Multi-Layer Perceptron. I'm glad that the MLP is going to compute something out for me today.

Training a neural network is a really tricky work, it seems we need a lot of interaction with the results and parameters.

A typical training with 200 training data, 3-10-1 topological MLP network runs about 4 minutes for 5000 epochs at this time. I do not apply any performance optimizing on it till now.