This program is for testing and experimenting with indicators to predict prices of stocks or cryptocurrencies. The prediction is based on a recurrent neural network. To make useful predictions based on this approach it is necessary to predict the indicators first. The goal of this program is not to make final predictions to decide whether to invest or not. But it's purpose is to experiment with indicators on the RNN approach to find more or less useful indicators and configurations.

The dataset in this example is a csv file of 5min BTC_XRP open, high, low, close ... data. It provides more than 120000 datapoints. As features for the network we will use indicators. The file indicators.py provides many indicators copied from the pandas-technical-indicators repository from Crypto-toolbox. Some of them are modified because the index did't work for me. Not all indicators are tested so maybe some of them have to be fixed first. The net we're using after preparing data is a recurrent neural network. At the end we'll plot everything with mpl_finance_ext (outdated! Please follow link for current version).

The first step is to prepare the data. We load the dataset and calculate the indicators. The dataset was downloaded from the Poloniex api. The indicators are copied from the pandas-technical-indicators repository from Crypto-toolbox as mentioned earlier.

# Load dataset
data = pd.read_csv('BTC_XRP_5min.csv', index_col=0).tail(5000)
data = data.drop(['date', 'quoteVolume', 'volume', 'weightedAverage'], 1)

# Calculate indicators
data = relative_strength_index(df=data, n=14)
data = bollinger_bands(df=data, n=20, std=4, add_ave=False)
data = exponential_moving_average(df=data, n=10)
data = moving_average(df=data, n=12)
data = macd(df=data, n_fast=12, n_slow=26)

# Cut DataFrame
data = data.iloc[40::]
# Reset index
data = data.reset_index()
# Delete old index
data = data.drop('index', 1)

# Normalize data
data_n = (data - data.mean()) / (data.max() - data.min())

To select a feature just enter the name of the column to the feature list:

features = ['MA_12', 'MACD_12_26']

To find out which names are existing just add print(list(data_n)). In this example available are ['close', 'high', 'low', 'open', 'RSI_14', 'bband_upper_20', 'bband_lower_20', 'EMA_10', 'MA_12', 'MACD_12_26', 'MACDsign_12_26', 'MACDdiff_12_26']

• Batch size: batch_size = 3
• Size of the test dataset. 0.1 means 10% of the complete dataset. The remaining 90% are testing data: test_dataset_size = 0.1
• Neurons of the RNN: num_units = 12
• Learning rate: learning_rate = 0.001
• Epochs: epochs = 8

For plotting we are using mpl_finance_ext (outdated! Please follow link for current version). from another Github repository of mine. The project provides a documentation.

Below you can see an example result of the parameter configuration described on this page. The white area shows the training data and the other area on the right in the upper chart shows the predicted data.

Here we take a closer look: