Machine learning and short positions in stock trading strategies

1
Machine learning and short positions in stock
trading strategies

• D.E Allen, R. Powell and A. K. Singh
• Edith Cowan University

2
Reading questions

1. What is short selling and why is it
   controversial?
2. What are Support Vector Machines (SVM) and why
   are they a useful technique?
3. Explain what kernel estimation is.
4. Why are different kernel estimators available?
5. Explain what logistic regression is.
6. What does Beta Measure?
7. Why are Sharpe ratios a useful investment metric?
8. How does Beta differ from Sharpe ratios.
9. How do we measure mean absolute error?
10. Why is out of sample forecasting important?

3
Introduction

• Forecasting future stock price movement using
  financial indicators.
• Evidence from past for predictability power of
  financial factors e.g. Beta, E/P, B/M, past
  returns etc.
• Support Vector Machines (SVM), capable of
  handling large amount of unstructured, noisy or
  nonlinear data.
• SVM classification useful in prediction of future
  price direction (1,-1).

4
SVM in Classification

• SVM are characterized by
• Mapping input vectors into higher dimensional
  feature space.
• Structural risk minimization
• Non linear modelling with Kernel Functions
• Kernel density estimators are non-parametric
  density estimators with no fixed structure. They
  depend on all the data points to obtain an
  estimate.
• Classification of classes using optimal
  separating hyperplane.

5
SVM

• Optimal Separating Hyperplane.

6
SVM

• SVM use following kernel functions
• Linear
• Polynomial
• Radial Basis Function (RBF)
• Sigmoid
• Here and d are kernel parameters.
• Study Uses RBF kernel for its robustness on non
  linear data.

7
Data

• Dow Jones Industrial Average sample Stocks daily
  data for a period of 5 years (1/03/2005-9/03/2010
  ).
• Factors Used for forecasting

Factors Underlying rationale
Previous 2 days daily log returns. Indicator of the historical performance, which is widely used in time series analysis.
Beta (six months rolling window) Return dependence on the market return in the long run.
Price to Earnings Ratio Indicator of the current company value which effects the price movement.
Book to Market Ratio Fama- French (1992, 1993)
Traded Volume Indicator of the performance of the stock in the market.
Dividend Yield Indicator of company performance. Blume (1980)
8
Methodology

• Standardization of Data
• Direction of price change classified into binary
  -1 and 1 using
• Testing sample is created using last 130 days
  data.
• Kernel parameters, cost and gamma are optimized
  using grid search. A systematic way of seeking
  optima.
• The model is built on training data and is used
  for forecasting which is tested on out sample
  data (130 days) SVM results are compared with
  Logistic Regression results (with same training
  and testing data).
• Simple investment strategy used to check the
  predicted directions

9
Forecasting Results
Stocks Stocks Results SVM Logistic Regression
Stock 1 Stock 1 Correctly Classified Instances 77 (59.2308 ) 67 (51.5385)
C Gamma Incorrectly Classified Instance 53 (40.7692) 63 (48.4615 )
724 0.1 Mean Absolute Error 0.4077 0.5015
Stock 2 Stock 2 Correctly Classified Instances 112 (86.1538) 109 (83.8462 )
C Gamma Incorrectly Classified Instance 18 (13.8462) 21 (16.1538 )
1024 0.12 Mean Absolute Error 0.1385 0.316
Stock 3 Stock 3 Correctly Classified Instances 76 (58.4615) 67 (51.5385 )
C Gamma Incorrectly Classified Instance 54 (41.5385 ) 63 (48.4615 )
1448 0.003162 Mean Absolute Error 0.4154 0.4962
Stock 4 Stock 4 Correctly Classified Instances 76 (58.4615) 69 (53.0769 )
C Gamma Incorrectly Classified Instance 54 (41.5385 ) 61 (46.9231 )
724 3 Mean Absolute Error 0.4154 0.4963
Stock 5 Stock 5 Correctly Classified Instances 80 (61.5385) 59 (45.3846 )
C Gamma Incorrectly Classified Instance 50 (38.4615 ) 71 (54.6154 )
1448 0.56 Mean Absolute Error 0.3846 0.5091
10
Investment Strategy Results
The final net returns of the stocks are compared
using the Sharpe Ratio.
Final Return Final Return Sharpe Ratio Sharpe Ratio
SVM LOGISTIC SVM LOGISTIC
Stock1 20.10167056 -12.0362 17.42748 -13.0499
Stock2 7.246199093 6.009645 4.356055 3.369538
Stock3 16.33556329 15.30477 14.78509 13.72405
Stock4 14.33568424 5.611437 14.83901 4.495077
Stock5 18.27861273 -5.49125 14.62362 -6.39905
DJIA 10.12379524 10.12379524 8.10426878 8.10426878
11
Conclusion

• SVM classification outperforms logistic
  regression in classifying price direction.
• Simple stock trading strategy also reveals the
  efficiency of SVM in stock trading.
• Further applications can include prediction of
  other financial time series.
• SVM regression can be further tested for similar
  work