What is the 7-action space in this skill?

It expands basic Buy/Sell/Hold into a Hold action plus three size tiers (25%, 50%, 75%) for both Buy and Sell directions, allowing models to decide how much to allocate.

Why did the observation dimension increase from 5600 to 5900?

The skill adds three new features (position size %, available capital %, and normalized account value) across a 100-step observation window.

Is it compatible with older model versions?

No, because the observation dimensions change, v2.6 and older models cannot be used with v2.7 environments without retraining or mapping.

How does this handle 100% position allocation?

To prevent 'all-in' behavior, 100% allocation is achieved through scaling in, where the model issues consecutive signals in the same direction.

Trading Position Sizing & Action Space

Name: Trading Position Sizing & Action Space
Author: smith6jt-cop

bysmith6jt-cop

データサイエンスとML

Implements a 7-action space with integrated position sizing and small account simulation for reinforcement learning trading models.

概要

This skill upgrades trading reinforcement learning (RL) models by expanding the decision-making process from simple directionality to nuanced capital allocation. It introduces a 7-action space—allowing the model to choose between conservative (25%), standard (50%), or aggressive (75%) sizing for long and short entries. Designed specifically for small account simulations ($1,000 to $25,000), it includes safety buffers, reward scaling based on position size, and expanded observation features to help models learn risk-adjusted trading strategies that better match live market execution.

主な機能

Scale-In Logic: Enables reaching 100% allocation through consecutive same-direction signals rather than single high-risk trades.
0 GitHub stars
Expanded Observation Matrix: Adds capital availability, account value, and position size features for deeper model memory.
Small Account Simulation: Configurable $1,000+ initial account values with 30% safety buffers.
7-Action RL Space: Multi-tier sizing (25%, 50%, 75%) for both Long and Short positions.
Position-Scaled Rewards: Dynamically scales reward signals based on the actual percentage of capital at risk.

ユースケース

Standardizing the action space between training environments and live trading execution bots.
Simulating realistic trading environments for accounts under the $25,000 Pattern Day Trader (PDT) threshold.
Training RL models to manage capital and risk rather than just timing market entries.

概要

主な機能

Scale-In Logic: Enables reaching 100% allocation through consecutive same-direction signals rather than single high-risk trades.
0 GitHub stars
Expanded Observation Matrix: Adds capital availability, account value, and position size features for deeper model memory.
Small Account Simulation: Configurable $1,000+ initial account values with 30% safety buffers.
7-Action RL Space: Multi-tier sizing (25%, 50%, 75%) for both Long and Short positions.
Position-Scaled Rewards: Dynamically scales reward signals based on the actual percentage of capital at risk.

ユースケース

Standardizing the action space between training environments and live trading execution bots.
Simulating realistic trading environments for accounts under the $25,000 Pattern Day Trader (PDT) threshold.
Training RL models to manage capital and risk rather than just timing market entries.