Smart Stock Prediction

Chanjuan Wen

1 Goal

• Based on stock price data of the previous week, we use supervised machine learning algorithms to predict the price of stocks for the next day.
• After defining a set of features as the measurement for good performance of stocks, we use unsupervised machine learning algorithms to identify a set of stocks with good performance.

2 Methods

2.1 Supervised Learning Algorithm

• Linear kernel: K(x⁽ⁱ⁾, x^(j)) = (x⁽ⁱ⁾)^T(x⁽ⁱ⁾) + c
• Quadratic kernel: K(x⁽ⁱ⁾, x^(j)) = 1 − (||x⁽ⁱ⁾ − x^(j)||)/(||x⁽ⁱ⁾ − x^(j)|| + c)
• Polynomial kernel: K(x⁽ⁱ⁾, x^(j)) = (a(x⁽ⁱ⁾)^T(x^(j)))^d
• Gaussian kernel: K(x⁽ⁱ⁾, x^(j)) = exp( − (||x⁽ⁱ⁾ − x^(j)||²)/(2σ²))

2.2 Unsupervised Learning Algorithm

3 Dataset and Features

• Historic stock price. We collected the historic price data of seven stocks in information technology sector of Yahoo Finance. They are Apple, Autodesk, Cisco, Intel, Microsoft, Oracle, Yahoo. The features of these stocks includes: Date, Symbol, Open Price, High Price, Low Price, Low Price, Volume. The size of this data set is 1650. Figure 1 displays part of the data.
  
  
  Figure 1 Historic Stock Price
• General stock information. We also collected some general information about S&P500 stock index, including current price, sum of trading volume, P/E, 200-day Moving Average, Change percent, Market Capitalization. This size of this data set is 500. Figure 2 displays part of the data.
  
  
  Figure 2 General Stock Information

4 Preliminary Results

• Price prediction of next day base on previous day’s price data. As illustrated in the figure 4, we can see that the rate of wrong prediction is pretty high for all the four kernels, at about 30% ~ 50%.
  
  
  Figure 3 price prediction of next day base on previous day’s price data
• Price prediction of next day base on previous week’s price data. As illustrated in figure 5, when we use the data of previous week to predict the next day’s price, compared to the previous case, the number of errors is much smaller. Especially, the quadratic kernel achieved the best performance. The rate of wrong prediction is at about 13%.
  
  
  Figure 4 price prediction of next day base on previous week’s price data
• Price prediction of next week based on one day’s price data. Basically, this prediction doesn’t have much value. And our experiment results also show that the number of errors are pretty high, nearly 50%.
  
  
  Figure 5 price prediction of next week based on one day’s price data

5 Remaining Tasks

• Optimize the prediction for stock price. On one hand, I will try to collect more data with more valueable features which will result in more precise prediction. On the other hand, I will try to implement other classification algorithms, such as decision tree, to compare the results with what the SVM produces.
• Identify a set of stocks with good performance. I will analyze the data and define some features for stock with good performance. Then I will implement at least two unsupervised learning algorithms, such as K-means and Mixture of Gaussians, compare the results.

6 Timeline

References

[1] http://en.wikipedia.org/wiki/Support_vector_machine

[2] http://en.wikipedia.org/wiki/K-means_clustering

[3] Christopher King, Christophe Vandrot, John Weng. A SVM Approach to Stock Trading, 2009.

[4] Hung Pham, Andrew Chien, Youngwhan Lim. A Framework for Stock Prediction, 2009.