Skip to content

autogen.md

Latest commit

 

History

History
219 lines (168 loc) · 10.5 KB

autogen.md

File metadata and controls

219 lines (168 loc) · 10.5 KB
Raw

💸 AutoGen Reference Implementation

This is the AutoGen reference implementation.

The Semantic Kernel workflow provides the maintained deterministic pipeline. This implementation remains available for framework comparison and migration study.

It has an independent Poetry project requiring Python 3.11–3.12. It is incompatible with this repository's Python 3.13 root environment; use a separate environment if you need to run it.

🔑 Key Differences from Microsoft Agent Framework

Aspect AutoGen reference Microsoft Agent Framework
Orchestration GroupChat + select_speaker WorkflowBuilder + edges
Message Flow Broadcast to all agents Data flows through edges
Agents AssistantAgent, ConversableAgent ChatAgent (stateless)
Tools FunctionTool class class or @ai_function decorator
State Built into agents AgentThread for context
Pattern Control-flow (event-driven) Data-flow (workflow-based)
Routing Custom logic in manager Conditional edges

📊 Architecture Comparison

AutoGen Reference Architecture

flowchart TD
    classDef user         fill:#e6e6fa,stroke:#333,stroke-width:2px
    classDef orchestrator fill:#f0fff0,stroke:#999,stroke-width:1px
    classDef agent        fill:#f0f8ff,stroke:#333,stroke-width:2px

    A[User Input]:::user --> B[UserProxyAgent]:::user
    B --> C[GroupChatManager<br/>Custom speaker selection]:::orchestrator
    C --> D[Stock Analysis Agent]:::agent
    C --> E[Signal Analysis Agent<br/>Python code executor]:::agent
    C --> F[Code Executor Agent]:::agent
    D --> G[Tools]:::user
    E --> G
    F --> G
    G --> H[Manual Result Collection]:::orchestrator
Loading

Best time to buy/sell: Sell in May, buy in October

Overview

💸Agent-based stock analysis and investment strategy application using 🎰AutoGen framework.

  • Review and backtest the historical stock performance for a specific ticker and date range.
  • Autogen Agents for finding the best timing to invest.
  • To run the main workflow: python agent_workflow_e2e.py
  • Generated AutoGen artifacts are written to output/autogen. Both executable entry points print this directory when they start.

Important: The code in this repository was developed during a hackathon and implemented within a limited timeframe. It is intended for proof-of-concept purposes only.

Sequence diagram

sequenceDiagram
    %% User
    participant User
    %% Generates Investing Strategies
    participant StrategyIdeaAgent
    %% Generates CAGR/MDD graph
    participant StockReportAgent
    %% Manages group chat discussions
    participant GroupChatManager
    %% Executes code to create signal data
    participant UserProxyAgent
    %% Fetches price data from Yahoo Finance and performs backtesting
    participant StockAnalysisAgent
    %% Generates code for creating buy/sell signals for backtesting
    participant SignalAnalysisAgent
    %% Public API for fetching stock price data for specified periods
    participant YahooFinanceAPI
    %% Public web search for finding technical indicators and investing strategies
    participant BingSearch

    autonumber

    rect rgb(204, 255, 204)
        User ->> StrategyIdeaAgent: Request to create investment strategies
        StrategyIdeaAgent ->> StrategyIdeaAgent: Stores investment strategies as 'strategy_ideas.json'
    end

    loop Iterate through investment strategies
        User ->> GroupChatManager: Sends an investment strategy to group chat
        GroupChatManager ->> GroupChatManager: Initializes group chat
        GroupChatManager ->> StockAnalysisAgent: Calls StockAnalysisAgent
        StockAnalysisAgent ->> YahooFinanceAPI: Fetches stock data from Yahoo Finance
        YahooFinanceAPI -->> StockAnalysisAgent: Provides stock data
        StockAnalysisAgent ->> StockAnalysisAgent: Stores stock price dataset as 'stock_data.csv'
        StockAnalysisAgent -->> GroupChatManager: Notifies GroupChatManager of task completion
        GroupChatManager ->> SignalAnalysisAgent: Calls SignalAnalysisAgent
        SignalAnalysisAgent ->> SignalAnalysisAgent: Generates code for creating buy/sell signals
        opt
            SignalAnalysisAgent ->> BingSearch: Searches for examples using the `ta` library
            BingSearch -->> SignalAnalysisAgent: Provides examples
        end
        SignalAnalysisAgent -->> GroupChatManager: Notifies GroupChatManager of task completion
        GroupChatManager ->> UserProxyAgent: Calls UserProxyAgent
        UserProxyAgent ->> UserProxyAgent: Executes code created by SignalAnalysisAgent
        UserProxyAgent ->> UserProxyAgent: Stores buy/sell signal dataset as 'stock_signals.csv'
        UserProxyAgent -->> GroupChatManager: Notifies GroupChatManager of task completion
        GroupChatManager ->> StockAnalysisAgent: Calls StockAnalysisAgent
        StockAnalysisAgent ->> StockAnalysisAgent: Performs backtesting using 'stock_data.csv' and 'stock_signals.csv'
        opt
            StockAnalysisAgent ->> BingSearch: Searches for descriptions of indicators
            BingSearch -->> StockAnalysisAgent: Provides descriptions
        end
        StockAnalysisAgent ->> User: Sends a summary of the investment strategy performance
    end

    rect rgb(230, 242, 255)
        User ->> StockReportAgent: Requests creation of a graph
        StockReportAgent ->> StockReportAgent: Generates CAGR/MDD graph
    end
Loading

Input and Output

  • User Input
Run complete stock analysis for MSFT from 1995-01-01 to today. 
Generate buy/sell signals using 20-day Moving Average, TRIX, UO.
Based on the generated signals, backtest the strategy and provide performance metrics.
  • Output
### Strategy Performance Summary

**Indicators Used:**
- 20-day Moving Average (MA)
- TRIX (Triple Exponential Average)
- Ultimate Oscillator (UO)

**Strategy Summary:**
The strategy employs a combination of technical indicators to generate buy and sell signals. The 20-day Moving Average smooths price data to identify trends, while the TRIX indicator assesses momentum and potential trend reversals. The Ultimate Oscillator helps gauge market momentum by considering multiple time frames.

Buy signals are generated when the stock price is above the 20-day MA, TRIX is positive, and UO is below 50, indicating a potential upward trend. Conversely, sell signals are triggered when the price is below the 20-day MA, TRIX is negative, and UO is above 50, suggesting a downward trend. This strategy aims to capture momentum while managing risk effectively.

### Performance Metrics

| Metric               | Value    |
|----------------------|----------|
| **Cumulative Return** | 188.17%  |
| **CAGR**             | 2.15%    |
| **Maximum Drawdown** | -4.47%   |
| **Sharpe Ratio**     | 0.04     |

CAGR calculation

  • Most backtrading systems already integrate buy/sell signals and CAGR (Compound Annual Growth Rate) calculation into a single interface. However, this approach limits the flexibility to use custom signals for CAGR calculation.

  • To resolve this issue, the app uses an approach that separates signal generation and CAGR calculation.

    • Assumptions

      • To streamline the process, we make the following assumptions:

        1. All trading decisions are based on the signal generated from the previous day's data.
        2. The calculation of daily returns depends on the type of signal and the price movement:

        • Buy Signal: On the trading day, the daily return is calculated as the change of the Adjusted Close price.

        • Sell Signal: On the trading day, the daily return is calculated as the gap between the previous day's Close price and the current day's Open price: Daily Return = ({Open} / {Prev Day Close}) - 1

    • Signal Validation

      • Each trade checks the validity of the signal:

        1. A sell signal cannot be executed without a preceding buy signal.
        2. Consecutive identical signals (e.g., multiple buy signals in a row) are treated as a Hold action, where no new trade is initiated.

Configurations and Setup

poetry install --no-root
  • This application library supports ta, which provides commonly used indicators (pure Python).
  • Another backtesting framework will be supported in the future.
  • Rename .env.template to .env and OAI_CONFIG_LIST.template.json to OAI_CONFIG_LIST.json. Then, set your Bing Search API and OpenAI keys.

Python Libraries for Quant Trading

  • Zipline: Maintained and updated by the community after Quantopian shut down. git
  • backtrader: Python Backtesting library for trading strategies git
  • QuantStats: git
  • fastquant: git
  • TA-Lib Official Site: 200 indicators such as ADX, MACD, RSI, Stochastic, Bollinger Bands etc. Candlestick patterns. Faster (C-based). recognition. git 
python -m pip install TA_Lib-0.4.32-cp311-cp311-win_amd64.whl
  • finta: Common financial technical indicators implemented in Pandas. archived repository. git
  • FinRobot uses GroupChat for multiple agents
    • GroupManager controls mutiple agents. It can be used to manage multiple agents in a group chat.

Official Documentation

Official AutoGen documentation  |  AutoGen GitHub repository