NeuralNetIndicator

The goal of NeuralNetIndicator is to provide functionality to calculate a financial trading indicator based on artificial neural network price predictions. The model bases price forecasts on reference stock prices of the same sector. The focus is on the utilization of lag and time delays in price movements of similar stocks in order to profitably forecast price changes. With a lot of customizable parameters for the network training algorithm, training windows sizes and a range of Tukey’s smoothing functionalities, the indicator function offers a useful tool to test various parameter settings for their profitability. The information contained in the indicator values and usage suggestions of the indicator functions can be found under the subheadings Indicator Values and Optimization and Backtesting at the end of this document.

Along the walk-forward trained model with various training-algorithm adjustments and tweaking possibilities, this also contains a function to download historical stock prices into a suitable xts object as well as functionality to plot the oscillating indicator along the underlying stock prices as a chart.

Installation

In order to install NeuralNetIndicator from GitHub, you will first need to install devtools.

# if not installed
install.packages("devtools")

devtools::install_github("n1tecki/NeuralNetIndicator")

Example

This is an example which shows you how to access historical data and train the forecast-model:

library(NeuralNetIndicator)

getReferenceSymbols

First, the getReferenceSymbols downloads historical data into the xts object symbols. Along the stock prices of interest, reference assets (preferably from the same (GISC) sector) should be listed with their stock symbols. In this case a time span of approximately 20 years will be used.

## basic example code
symbols <- getReferenceSymbols(x = c('XOM','BP','COP','OXY','PXD','TOT','VLO'), 
                               start_date = '2000-10-10', end_date = '2020-10-10')
                               
head(symbols, 4)
#>            XOM.Close BP.Close COP.Close OXY.Close PXD.Close TOT.Close VLO.Close
#> 2000-10-10  12.95871 21.11406  5.070281  2.667497  13.19047  7.102553  1.131466
#> 2000-10-11  13.01947 21.45461  5.090010  2.659743  13.24884  7.102553  1.131466
#> 2000-10-12  13.07154 22.06273  5.242908  2.690761  13.54066  7.314998  1.154994
#> 2000-10-13  12.56812 21.23568  5.050553  2.566691  13.13211  6.999201  1.107939

walkForwardNeuralnet

The xts object symbols of closing price of all stocks is used as the input for the walkForwardNeuralnet. The traded asset at focus is XOM, hence the first column is selected for the parameter trad_col. With a in-sample window of 700 days and out-of-sample of 60, the parameter slid - number of days of how far into the future the forecast should be calculated - is selected to be 4. The indicator values in this example will not be smoothed, hence the smoothing parameter is left as default NULL.
The closing prices of XOM, price predictions and indicator value for each day are exported into the xts object indicator.

indicator <- walkForwardNeuralnet(symbols, trad_col=1, neurons_int = c(3,5), 
                                  in_sample = 700, out_sample = 60, slid = 4)
                                  
head(indicator, 4)
#>            XOM.Close forecast indicator
#> 2003-07-29  21.22882 21.43451 0.9689374
#> 2003-07-30  21.16291 21.43184 0.9689374
#> 2003-07-31  21.31870 21.43704 0.9689374
#> 2003-08-01  21.16890 21.42723 0.9689374

indicatorChart

The indicatorChart function allows to plot a chart of the indicator, along the stock prices of interest (symbol) and zoom in on points of interest. The indicator values indicator$indicator from the previously received indicator object are passed on, along the stock code of the corresponding stock prices.

indicatorChart(indicator$indicator, symbol = 'XOM', zoom = '2018-07-30::2019-07-01')

It is also possible just to plot the indicator itself to visualize the values.

indicatorChart(indicator$indicator, zoom = '2017-07-27::2018-10-01')

Indicator Values

The oscillating values of the indicator are the percentage difference between the closing price of that day and the price forecast for slid days ahead calculated by the neural network.
For example, an indicator value of 1.4 means that the neural network estimates a 1.4% increase in price in slid days. The same goes for negative values - when the indicator shows a value of -2.3, the forecast is that the price will fall 2.3%. Based on those values buy or sell positions can be generated. For example if the indicator reaches a value of 2, a long position is entered; when the value is above 3, the long position is increased.

Optimization and Backtesting

As the walkForwardNeuralnet function has many parameters that can be changed, it is encouraging the user to test different combinations of parameter values (in-sample/out-of-sample length, number of nodes, different combination of stocks, back-propagation algorithm of the neural network, error threshold, …) for their profitability. One recommendation is to use a genetic algorithm and back-testing the different combinations of parameters for their profitability.