🌍 English ∙ German

Reinforcement Learning for Tactical Asset Allocation

This project contains the training and testing of multiple reinforcement learning agents given a portfolio management environment.

The interaction between environment and agent is given by:
Environment <> Runner <> Agent <> Model

Important files and folders

• Data aquisation: data.py
• Environment: environment.py
• Hyperparametertuning of the prediction model: predictor.py
• Training of the prediction model: predictor_network.py
• Runner: runner.py
• Basic configuration parameters: config.py
• Training: train.py
• Testing: test.py
• Agent configurations: agent_config
• Neural net configurations: net_config

Dependencies

The proposed implementation was done mainly in python, therefore python version >=3.6 is required. Furthermore following additional python packages are required:

• h5py==2.7.1
• Keras==2.1.3
• matplotlib==2.1.0
• numpy==1.14.1
• pandas==0.20.3
• pandas-datareader==0.5.0
• scikit-learn==0.19.1
• scipy==1.0.0
• seaborn==0.8.1
• tensorflow==1.4.0
• tensorflow-tensorboard==0.4.0rc3
• tensorforce==0.3.5.1

Running the train file

For agent training you should run the file train.py using the console.
For example:

python ~/path/to/file/run/train.py -at "clipping" -v 1

Changes to the environment and/or run parameters can be selected through the train file, config file and through pre specified flags shown below.

Modifications of the agents and models must be specified through the belonging config file (json format).

Flags:

Flag 1	Flag 2	Meaning
-d	--data	path of the environment.csv file
-sp	--split	train/test split
-th	--threaded	(bool) threaded runner oder single runner
-ac	--agent-config	path to the agent config file
-nw	--num-worker	number of threads if threaded is being selected
-ep	--epochs	number of epochs
-e	--episodes	number of episodes
-hz	--horizon	investment horizon
-at	--action-type	action typ: 'signal', 'signal_softmax', 'direct', 'direct_softmax', 'clipping'
-as	--action-space	action space: 'unbounded', 'bounded', 'discrete'
-na	--num-actions	number of dicrete actions given a discrete action space
-mp	--model-path	saving path of the agent model
--eph	--eval-path	saving path of the agent model of the evaluation files
-v	--verbose	console verbose level
-l	--load-agent	if given agent will be loaded based on a prior save point (path)
-ds	--discrete_states	discretization of the state space if true
-ss	-standardize-state	(bool) standardization or normalization of the state
-rs	--random-starts	(bool) random starts for each new episode

Running the test file

The execution of the test file is very similar to the one of the train file. There has to be a checkpoint in saves for the selected agent.

python ~/path/to/project/run/test.py -l /project/model/saves/AgentName

The folder saved_results contains multiple parameter constellations of pretrained agents.

Flags

Flag	Flag 2	Meaning
-d	--data	path of the environment.csv file
-ba	--basic-agent	selection of a BasicAgenten: 'BuyAndHoldAgent', 'RandomActionAgent'
-sp	--split	train/test split
-ac	--agent-config	path to the agent config file
-e	--episodes	number of episodes
-hz	--horizon	investment horizon
-at	--action-type	action typ: 'signal', 'signal_softmax', 'direct', 'direct_softmax', 'clipping'
-as	--action-space	action space: 'unbounded', 'bounded', 'discrete'
-na	--num-actions	number of dicrete actions given a discrete action space
--eph	--eval-path	saving path of the agent model of the evaluation files
-v	--verbose	console verbose level
-l	--load-agent	if given agent will be loaded based on a prior save point (path)
-ds	--discrete_states	(bool) discretization of the state space if true
-ss	-standardize-state	(bool) standardization or normalization of the state

TensorBoard

The files predictor.py as well as train.py integrate TensorBoard. TensorBoard can be loaded by typing

tensorboard --logdir path/to/project/env/board
tensorboard --logdir path/to/project/run/board

Got to localhost:6006 for the results.

Credits

• TensorForce