FinWorld Configuration Guide

Comprehensive guide to configuring FinWorld using Python-based configuration files

Configuration System Overview

FinWorld uses a Python-based configuration system built on mmengine. Each configuration file is a Python script that defines all necessary parameters for a specific task or component.

                    Key Features
                    Python-based: Full Python syntax support for complex configurations
mmengine Integration: Built on mmengine for unified experiment management
Registry Mechanism: Flexible component management with type parameters
Seven-Layer Architecture: Organized by hierarchical layer structure
Task-oriented: Separate configs for different financial tasks

                

Configuration Structure

# Basic configuration structure
task_type = "trading"           # Task type
symbol = "AAPL"                 # Asset symbol
workdir = "workdir"             # Working directory
method = "ppo"                  # Method name
tag = f"{symbol}_{task_type}_{method}"  # Experiment tag

# Dataset parameters
history_timestamps = 64
future_timestamps = 32
start_timestamp = "2015-05-01"
end_timestamp = "2025-05-01"

# Model parameters
embedding_dim = 256
num_heads = 4
depth = 4

# Configuration dictionaries
dataset = dict(...)
environment = dict(...)
agent = dict(...)
trainer = dict(...)
task = dict(...)

Seven-Layer Architecture Configuration

FinWorld's configuration system is organized around its seven-layer architecture. Each layer has specific configuration parameters:

1. Data Layer

Download & Storage

Download Configuration

downloader = dict(
    type="PriceDownloader",  # Specify downloader type
    source="fmp",
    assets_path="configs/_asset_list_/dj30.json",
    start_date="1995-05-01",
    end_date="2025-05-01"
)

Dataset Configuration

dataset = dict(
    type="SingleAssetDataset",  # Specify dataset type
    symbol="AAPL",
    data_path="datasets/exp",
    history_timestamps=64,
    future_timestamps=32
)

2. Process Layer

Data Processing

Processor Configuration

processor = dict(
    type="AggProcessor",  # Specify aggregator processor type
    procs_config=[
        dict(
            type="Processor",  # Specify processor type
            assets_name="dj30",
            data_type="price",
            feature_type="Alpha158"
        )
    ]
)

3. Model Layer

AI Models

Embedding Configuration

embed_config = dict(
    type="TradingPatchEmbed",  # Specify embedding layer type
    dense_input_dim=150,
    latent_dim=256,
    dropout=0.0
)

Encoder Configuration

encoder_config = dict(
    type="TransformerEncoder",  # Specify encoder type
    input_dim=256,
    depth=4,
    num_heads=4
)

4. Agent Layer

AI Agents

RL Agent

agent = dict(
    type="PPO",  # Specify PPO agent type
    task_type="trading",
    action_dim=3,
    actor_output_dim=3,
    critic_output_dim=1
)

LLM Agent

agent = dict(
    type="FinanceAgent",  # Specify finance agent type
    model="Qwen3-32B",
    max_steps=20
)

5. Environment Layer

Trading Environment

environment = dict(
    type="EnvironmentPatchTrading",  # Specify trading environment type
    mode="train",
    initial_amount=1e5,
    transaction_cost_pct=1e-4
)

Portfolio Environment

environment = dict(
    type="EnvironmentPatchPortfolio",  # Specify portfolio environment type
    mode="train",
    initial_amount=1e5,
    num_assets=29
)

6. Trainer Layer

Training Logic

RL Trainer

trainer = dict(
    type="PPOTradingTrainer",  # Specify PPO trainer type
    policy_learning_rate=1e-5,
    total_steps=int(1e8),
    device="cuda"
)

ML Trainer

trainer = dict(
    type="MLTradingTrainer",  # Specify ML trainer type
    n_estimators=1000,
    learning_rate=0.05
)

7. Task Layer

Task Orchestration

Task Configuration

task = dict(
    type="Task",  # Specify task type
    trainer=None,
    task_type="trading"
)

Async Task

task = dict(
    type="AsyncTask",  # Specify async task type
    trainer=None,
    task_type="trading"
)

Task-Specific Configurations

FinWorld provides comprehensive configuration examples for different financial tasks. All configuration files are organized in the configs/ directory: