FDRNN

Implementing Deng, Y., Bao, F., Kong, Y., Ren, Z., & Dai, Q. (2016). Deep direct reinforcement learning for financial signal representation and trading. IEEE transactions on neural networks and learning systems, 28(3), 653-664.

Preparation

1. Prepare the index data as CSV file. The file must include column CloseDiff, which represents the index difference CloseDiff[t] = Index[t] - Index[t-1]. The CSV files must arrange as following directory structure:

   +-- Data
   | +-- futures
   | | +-- future_2018-01-01.csv
   | | +-- future_2018-01-02.csv
   | | +-- ...
   

2. To reduce the training time, it is strongly recommended that computing the parameters of fuzzy representation before training. The vanilla index file can be transformed into fuzzy version via applying FuzzyStreamer in handler.py.

   fromhandlerimportFuzzyStreamer#streamer = FuzzyStreamer(<window size>, <fuzzy degree>)streamer=FuzzyStreamer(lag, fuzzy_degree)
   # streamer.transform(<folder of original index files>, <folder of fuzzy index files>)streamer=streamer.transform('./Data/futures/train', './Data/fuzzy_futures/train')

3. Adjust the required parameters in config.ini

   [default]# Number of training epochsepochs = 1000
   # Save the model each n epochssave_per_epoch = 20
   # Transaction costc = 0.05
   # Window sizelag = 50
   # Data pathdata_src = ./Data
   # Log pathlog_src = ./Pickle
   [fddrl]fuzzy_degree = 3

Run

Running FDRNN - The proposed method in the paper

python main.py

Running baseline DDRL - The proposed model without fuzzy representation

python baseline_ddrl.py

Running baseline DRL - The proposed method without fuzzy representation and autoencoder

python baseline_drl.py