btgym.research package

btgym.research.gps subpackage

• btgym.research.gps.aac module
• btgym.research.gps.policy module
• btgym.research.gps.loss module
• btgym.research.gps.strategy module
• btgym.research.gps.oracle module

btgym.research.casual_conv subpackage

• btgym.research.casual_conv.layers module
• btgym.research.casual_conv.networks module
• btgym.research.casual_conv.policy module
• btgym.research.casual_conv.strategy module

btgym.research.strategy_gen_4 module

class btgym.research.strategy_gen_4.DevStrat_4_6(**kwargs)

Objectives:

external state data feature search:

time_embedded three-channeled vector:

• Open channel is one time-step difference of Open price;
• High and Low channels are differences between current Open price and current High or Low prices respectively

internal state data feature search:

time_embedded concatenated vector of broker and portfolio statistics time_embedded vector of last actions recieved (one-hot) time_embedded vector of rewards

reward shaping search:

potential-based shaping functions

Data:

synthetic/real

Parameters:**kwargs – see BTgymBaseStrategy args.

class btgym.research.strategy_gen_4.DevStrat_4_7(**kwargs)

4_6 + Sliding statistics avg_period disentangled from time embedding dim; Only one last step sliding stats are used for internal state; Reward weights: 1, 2, 10 , reward scale factor aded;

class btgym.research.strategy_gen_4.DevStrat_4_8(**kwargs)

4_7 + Uses full average_period of inner stats for use with inner_conv_encoder.

class btgym.research.strategy_gen_4.DevStrat_4_9(**kwargs)

4_7 + Uses simple SMA market state features.

class btgym.research.strategy_gen_4.DevStrat_4_10(**kwargs)

4_7 + Reward search: log-normalised potential functions. Nope.

class btgym.research.strategy_gen_4.DevStrat_4_11(**kwargs)

4_10 + Another set of sma-features, grads for broker state

class btgym.research.strategy_gen_4.DevStrat_4_12(**kwargs)

4_11 + sma-features 8, 512;