Steve Thawornwong

1
Engineering Management Graduate Seminar
Using Technical Analysis Indicators as Inputs to
Neural Networks for Predicting Stock Movements

• By
• Steve Thawornwong
• Dr. David Enke, Dissertation Supervisor
• September 24, 2001

2
Introduction

• Stock trend prediction
• Monetary rewards
• Difficult to perform
• Many factors are involved
• Complex relationships
• Technical Analysis Past Perspective
• Has been criticized
• Built on weak foundations
• Does history repeat itself?

3
Introduction

• Technical Analysis Current Perspective
• Viable analytical option
• Markets may not be fully efficient
• A mixture of human, political, and economical
  behaviors
• Use data of past prices and volumes
• Rely on strong empirical regularities
• Not always evident, masked by noise, and vary
  from security to security
• Difficult for producing consistent results

4
Introduction

• Artificial neural networks
• Non-linearity
• No priori assumptions required
• Useful in security investment

5
Research Investigation

• Used neural networks to enhance the effectiveness
  of technical analysis
• Objective Can neural networks be used to better
  uncover the regularities?
• Five popular technical indicators were selected
• Three neural networks were examined
• Three major stocks were selected
• Trading strategies were developed
• Predictability and profitability were evaluated

6
Literature Review

• Two forecasting approaches
• Technical analysis
• Fundamental analysis
• Technical analysis
• Price and volume movements
• Charts are a primary tool
• Some patterns are expected to repeat
• Motivation Identify changes in trends at an
  early stage

7
Literature Review

• Fundamental analysis
• Economic forces of supply and demand
• Relevant information
• Determine intrinsic value
• Samples of information
• Economic variables
• Industry specific
• Company specific

8
Literature Review

• Stock prediction with neural networks (NNs)
• Various studies used fundamental factors
• Historical time-series of prices were used
• Some technical indicators were included in a few
  studies
• The abilities of technical indicators when using
  NNs were rarely addressed
• At best, they relied on a few indicators or a
  single NN
• Technical indicators were not effectively tested

9
Why neural networks?

• Strong regularities of technical analysis were
  captured using linear relationships
• No linear evidence for human, political, and
  economical behaviors
• Conflicting signals often occur
• Investors may get confused
• Diverse rationales from various technical
  analysts
• Depended on experience, criteria, observation,
  etc.
• NNs can help manage unclear regularities and
  contradicted signals

10
Why neural networks?

• Preprocessing steps were required to improve NN
  performance
• Applicable to certain aspects of technical
  indicators
• NN abilities could be further improved if
  original stock prices and volumes are
  preprocessed gt technical indicators
• Using technical indicators would allow NNs to
  concentrate on details for accurate predictions

11
Stock Selection

• Three major stocks
• Lockheed Martin (LMT), Caterpillar (CAT), and
  Delta Air Lines (DAL)
• Short-term prediction
• 08/22/1996 to 12/29/1999 (846 trading days)
• Adjusted for splits and dividend distributions
• Data were divided into two periods
• First period 08/22/1996 06/30/1999 (720 days)
• Second period 07/01/1999 12/29/1999 (126 days)
• First period for training and validating the NNs
• Second period for out-of-sample evaluation

12
Indicator Selection

• Five technical indicators
• MA, RSI, MFI, SO, and MACD
• Commonly used criteria were considered
• Moving Average (MA)
• Average value of a stock price

13
Indicator Selection

• Relative Strength Index (RSI)
• Compares price gains to price losses

14
Indicator Selection

• Money Flow Index (MFI)
• Accounts for volumes

15
Indicator Selection

• Stochastic Oscillator (SO)
• Indicate overbought or oversold
• Displayed by K and D (3-day MV of K)

16
Indicator Selection

• Moving Average Convergence/Divergence (MACD)
• A difference between an ? 0.075 and an ? 0.15
  exponential moving average (EMA)

17
Data Preprocessing

• Technical trading criteria are based on
  particular patterns and recent movements
• Eight variables exist from five technical
  indicators
• Closing Price (CP), RSI, MFI, MA, K, D, MACD,
  and Signal Line (SL)
• These variables were preprocessed to replicate
  the movement patterns for neural networks
• (1) CPt CPt-1
• (2) MAt CPt, (3) Kt Dt, (4) MACDt SLt
• No preprocessing for (5) RSI and (6) MFI

18
Data Preprocessing

• Three-day lags to account for recent movement
• CPt CPt-1, CPt-1 CPt-2, CPt-2 CPt-3, RSIt,
  RSIt-1, RSIt-2, MFIt, MFIt-1, MFIt-2, MAt CPt,
  MAt-1 CPt-1, MAt-2 CPt-2, Kt Dt, Kt-1
  Dt-1, Kt-2 Dt-2, MACDt SLt, MACDt-1
  SLt-1, and MACDt-2 SLt-2
• Eighteen inputs were used to predict the
  direction of the next day stock price (Ct1)
• Time lags were used throughout the experiment
• Similar to the real world

19
Trading Practices

• Profited from an increase (long position) and a
  decline (short position) in stock prices
• Charged 1 round-trip transaction cost
• Close a long position and create a short position
• Cover a short position and open a long position
• Transaction cost v.s. excessive trading
• Developed a trading strategy in connection with a
  predicted signals
• Assisted in making trading decisions and allowed
  for systematic trading

20
Trading Strategies

• Make a daily decision whether maintaining a
  current position or making a trade
• Money becomes illiquid in the position for that
  day
• Trading Criteria
• If Ct1 1, then maintain the current long
  position, or cover a short position and open a
  long position (pay the 1 transaction cost and
  receive the profit/loss of the covered short
  position)
• Else (if Ct1 1), then maintain the current
  short position, or close a long position and
  create a short position (pay the 1 transaction
  cost and receive the profit/loss of the closed
  long position)

21
Trading Strategies

• Similar trading strategies and practices were
  used for technical indicators
• A short position when a sell signal is generated,
  and vice versa
• Current position maintained when no signal
• No trading when a signal suggested a similar
  position

22
Profitability Measure

• Accumulation of gains/losses for the whole
  trading period
• Day 0 a buy signal gt Day 1 a long position at
  x1
• Day 1 a short signal gt Day 2 a short position
  at x2
• Total transaction cost 0.005(x1) 0.01(x2)
• Total investment x1 0.005(x1) 0.01(x2)
• Gain/Loss x2 x1
• Rate or return x2 x1 0.005(x1)
  0.01(x2) /x1
  0.005(x1) 0.01(x2)
• No reinvestment of profits

23
Neural Network Modeling

• Three neural networks were implemented
• Scaled input values into range 1 and 1
• Feed-forward neural network (FNN)
• Single layer architecture, sigmoid tangent
  function, and resilient learning algorithm
• Early stopping for generalization
• First period (720 days) 576 days (80) for
  training and 144 days (20) for validating
• The hidden neurons and learning rate were
  determined during network training

24
Neural Network Modeling
25
Neural Network Modeling

• Probabilistic Neural Network (PNN)
• A two-layer network based on the estimation of
  probability density function
• First layer computes distances of input vectors
• Second layer classifies probabilities into a
  class
• Neither validation nor early stopping is required
• 720 days (first period) for network modeling
• A smoothing parameter equal to 1.00

26
Neural Network Modeling
27
Neural Network Modeling

• Learning Vector Quantization Network (LVQ)
• A two-layer (competitive and linear) network
• Competitive layer indicates subclasses of input
  vectors
• Linear layer combines subclasses into a class
• LVQ2 algorithm was used
• 0.5 upward and downward class percentages
• Similar methodologies as those used in the FNN

28
Neural Network Modeling
29
Empirical Results

• Evaluated using the untouched out-of-sample data
  (Second period)
• Predictability Results
• SIGN Proportion of time that the signals of
  stock price changes are correctly predicted
• Buy-and-Hold Always investing in each stock
• PortNN Outputs of three NN models were combined
• e.g. 1 of FNN, 1 of PNN, and 1 of LVQ gt an
  upward direction

30
Empirical Results
31
Empirical Results

• Profitability Results
• of Trades Total number of transactions
• Total R Total rate of return on investment
• Port R Equally weighted portfolio of the three
  securities
• Port R wRLMT wRCAT wRDAL

32
Empirical Results
33
Conclusions

• Used neural networks to improve the predictive
  ability of technical indicators
• Three stocks were tested to support the
  robustness
• Different technical indicators did not work well
  for all securities
• Neural networks help manage contradicting signals
  from technical indicators
• Neural networks help improve the predictability
  and profitability of the technical indicators

34
Engineering Management Graduate Seminar
Using Technical Analysis Indicators as Inputs to
Neural Networks for Predicting Stock Movements

• By
• Steve Thawornwong
• Dr. David Enke, Dissertation Supervisor
• September 24, 2001