experiment1.md

Experiment 1

• Create categories for buy, hold and sell based on change in stock price.

• Predict future change in stock price n days ahead. Based on:

• Predicting the Direction of Stock Market Index Movement Using an Optimized Artificial Neural Network Model Technical Indicators from Table 2.

• Predicting the Direction of Stock Market Index Movement Using an Optimized Artificial Neural Network Model.pdf

Main

Experimental settings that are constant throughout the experiment:

• Objective: Predict future change in stock price n days ahead.
• Task type: Classification.
• Architecture: Feedforward.

Variation 1

• Input data: 3 stocks. E.g. [GOOG, IBM, AAPL]
• Output: One hot categorical prediction with threshold t+n time steps in the future.
  • n = [1, 2, 3, 5, 8, 13, 21]
  • Threshold [up_thr, down_thr]

Trading logic

If Output > up_thr: BUY Elif Output < down_thr: SELL Else HOLD

Variation 1

• Input data: 3 stocks. E.g. [GOOG, IBM, AAPL]
• Output: 3 output nodes. [[BUY(t+1), HOLD(t+1), SELL(t+1)], [BUY(t+2), HOLD(t+2), SELL(t+2)], ...,[BUY(t+n), HOLD(t+n), SELL(t+n)]]
  • n = [1, 2, 3, 5, 8, 13, 21]
  • Threshold [up_thr, down_thr]

Trading logic

If Output > up_thr: BUY Elif Output < down_thr: SELL Else HOLD

• categories = [[large pos, pos, no change, neg, large neg], [pos, no change, neg]]
• nr of input nodes
  • weekdays: Yes/No
  • On Balance Volume (OBV)
  • Simple Moving Average (SMA)
  • BIAS
  • Psychological Line (PSY)
  • ASY

Variations

• Input data: 3 stocks. [GOOG, IBM, AAPL]
• Output: One hot categorical prediction t+n time steps in the future.
  • n = [1, 2, 3, 5, 8, 13, 21]
  • categories = [[large pos, pos, no change, neg, large neg], [pos, no change, neg]]
• nr of input nodes
  • weekdays: Yes/No
  • On Balance Volume (OBV)
  • Simple Moving Average (SMA)
  • BIAS
  • Psychological Line (PSY)
  • ASY

• type of technical indicators
• nr of layers
• nr nodes in each layer
• dropout setting
• threshold values for output classes
• weight initialisation
• batch size

Notes

• Make sure to balance the number of outputs / category. Ideally they are equal.

Results

Prediction

Mean Average Percent Error (MAPE):

Backtest

Backtest on Out-Of-Sample data only.

Performance metrics are:

• Total Returns
• Average Annual Return
• Maximum drawdown
• Sharpe ratio sharpe = (returns/std)*sqrt(365.25) std = std_dev(returns)
• Sortino ratio

Comparision Strategies:

• Buy&Hold
• Naive
• Trend following

https://github.com/Crypto-toolbox/pandas-technical-indicators/blob/master/technical_indicators.py