Part 22: AI-Assisted Trading

Large Language Models (LLMs) like Claude and GPT-4 bring powerful new capabilities to algorithmic trading: analyzing sentiment in news articles, generating portfolio recommendations, and creating comprehensive market reports. In this chapter, you will learn how to integrate LLMs into your trading system using Puffin’s puffin.ai module.

Architecture

The AI-assisted trading pipeline follows a layered design. Raw financial text flows through LLM providers, gets scored by the sentiment engine, converts into tradable signals, and finally reaches the portfolio agent for allocation decisions.

flowchart TD
    classDef provider fill:#2d5016,stroke:#1a3a1a,color:#e8e0d4
    classDef sentiment fill:#1a3a5c,stroke:#0d2137,color:#e8e0d4
    classDef signal fill:#6b2d5b,stroke:#4a1a4a,color:#e8e0d4
    classDef agent fill:#8b4513,stroke:#5c2d0e,color:#e8e0d4

    NEWS["News Articles & Filings"]:::provider
    RSS["RSS Feed Fetcher"]:::provider

    CLAUDE["ClaudeProvider"]:::provider
    OPENAI["OpenAIProvider"]:::provider
    CACHE["Response Cache"]:::provider

    SENT["analyze_sentiment()"]:::sentiment
    BATCH["batch_sentiment()"]:::sentiment
    DECAY["Time-Decay Weighting"]:::sentiment

    GEN["NewsSignalGenerator"]:::signal
    THRESH["Threshold Filter"]:::signal
    SIGNALS["Trading Signals (DataFrame)"]:::signal

    AGENT["PortfolioAgent"]:::agent
    RECS["Allocation Recommendations"]:::agent
    REPORT["generate_market_report()"]:::agent

    NEWS --> RSS
    RSS --> CLAUDE
    RSS --> OPENAI
    CLAUDE --> CACHE
    OPENAI --> CACHE
    CACHE --> SENT
    SENT --> BATCH
    BATCH --> DECAY
    DECAY --> GEN
    GEN --> THRESH
    THRESH --> SIGNALS
    SIGNALS --> AGENT
    AGENT --> RECS
    AGENT --> REPORT

    linkStyle default stroke:#4a5568,stroke-width:2px

What You Will Build

By the end of this chapter you will have a working pipeline that:

  • Connects to multiple LLM providers through a single abstract interface
  • Extracts bullish/bearish sentiment scores from financial news with confidence levels
  • Converts raw sentiment into filtered, time-weighted trading signals
  • Delegates portfolio allocation to an AI agent that reasons over positions, market data, and signals
  • Generates formatted market analysis reports on demand

Module Overview

The puffin.ai package exposes the following components:

Component Module Purpose
LLMProvider puffin.ai.llm_provider Abstract base class with caching
ClaudeProvider puffin.ai.providers Anthropic Claude integration
OpenAIProvider puffin.ai.providers OpenAI GPT integration
analyze_sentiment puffin.ai.sentiment Single-text sentiment extraction
batch_sentiment puffin.ai.sentiment Batch analysis with time decay
NewsSignalGenerator puffin.ai.signals Sentiment-to-signal conversion
PortfolioAgent puffin.ai.agent AI portfolio management
generate_market_report puffin.ai.reports Markdown report generation
fetch_rss_news puffin.ai.news RSS news article fetcher

All LLM providers cache responses by default (keyed on provider, model, text, and prompt) to minimize API costs. You can tune the TTL with cache_ttl.

Chapters

  1. LLM Providers – Abstract provider interface, Claude and OpenAI implementations, caching, fallback strategies, and async patterns
  2. Sentiment & Signals – Sentiment extraction, batch analysis with time decay, news-driven signal generation, and strategy integration
  3. AI Agent Portfolio Manager – Portfolio agent recommendations, audit trail, executing trades, market report generation, and the complete news-driven bot

Prerequisites

Before starting this chapter, make sure you have:

LLM API calls incur real costs. Start with short texts and use caching aggressively during development. Monitor your token usage before running batch analyses.

Notebook: Run the examples interactively in operational.ipynb

Source Code

Browse the full implementation: puffin/ai/

Next Steps

Start with LLM Providers to set up your first provider connection, then work through sentiment analysis and the portfolio agent.

Table of contents