Investment Prediction Agent

AI-powered multi-agent trading system for comprehensive stock & prediction markets analysis and automated trading decisions.

🚀 Overview

Auto Investment Broker combines specialized LLM agents to analyze markets, debate strategies, and execute trades. It features real-time data processing, a "Bull vs. Bear" debate engine, and a modern dashboard for visualization.

✨ Key Features

• Multi-Agent Architecture: LLM agents including Fundamentals, News, Technical, and Risk Managers.
• Top Traders Leaderboard: Real-time tracking of top performers from ZuluTrade and Polymarket.
• Interactive Dashboard: Modern UI with specific agent reports, history tracking, and technical charts.
• "Bull vs. Bear" Debates: Automated debates to assess risk and reward before every trade.

📚 Documentation

Full documentation lives at autoinvestment.broker/docs

Section	What's in it
Getting Started	Quick Start · Architecture · Configuration
Deployment	Deployment guide · Database · Scheduled jobs · Troubleshooting
Trading	Algorithmic strategies · Technical indicators · Fundamentals · Correlations · Brokers
Agents	Debate agents & prompts · Research paper
Reference	Workspace packages · Investment dictionary · Risk disclosure

Engineering notes that are not part of the published docs live in devdocs/.

📦 Monorepo Layout

The repo is a Turborepo workspace. Every command below can be run from the repo root, and turbo fans it out to the workspaces that define it.

.
├── apps/
│   └── ai-broker-web/        # Next.js app (UI, API routes, docs, D1 schema & migrations)
├── packages/
│   ├── investing/            # Trading agents, market data, prediction markets
│   ├── ai-broker-api-client/ # Generated API client
│   ├── fin-data-api/         # Financial data API service
│   └── mcp-server/           # MCP server generated from the OpenAPI spec
├── devdocs/                  # Internal engineering notes
├── third-party-trading-bots/ # Vendored reference bots (see its README)
├── turbo.json                # Task graph & caching
├── tsconfig.base.json        # Shared TypeScript compiler options
├── README.md
├── CHANGELOG.md
└── LICENSE.md

npm install                   # installs every workspace
npm run dev                   # turbo run dev
npm run build                 # turbo run build (packages first, then the app)
npm run test                  # turbo run test
npm run test:coverage         # turbo run test:coverage, then merge into coverage/lcov.info

# Scope a command to a single workspace
npx turbo run build --filter=ai-broker-web
npx turbo run test --filter=investing

Coverage

npm run test:coverage runs each workspace's suite with coverage on (Vitest for the app, investing and the API client; Jest for fin-data-api) and merges the per-workspace LCOV reports into coverage/lcov.info with repository-root-relative paths.

CI runs the same command on every pull request in .github/workflows/test.yml and uploads that single report to Codecov. The upload needs a CODECOV_TOKEN repository or organization secret. Thresholds, the per-workspace component breakdown, and the paths excluded from coverage (vendored bots, generated SDKs, migrations) live in codecov.yml; the project and patch checks are informational, so a coverage dip never blocks a merge on its own.

Each workspace documents itself:

Workspace	README
The app	apps/ai-broker-web
Trading agents & market data	packages/investing
Financial data API	packages/fin-data-api
Generated API client	packages/ai-broker-api-client
MCP server	packages/mcp-server
Cron routes	apps/ai-broker-web/app/api/cron
Engineering notes	devdocs

🤖 AI Agents & Strategies

Agent/Team	Role
Analyst Team	Gathers data: Fundamentals, Sentiment (Social), News, & Technical Analysis.
Researcher Team	Conducts "Bull vs. Bear" debates; assesses risk.
Trader Agent	Synthesizes reports to propose trades.
Portfolio Manager	Final decision maker; manages risk and position sizing.

Analysis Team

• Market Analyst: Technical analysis and liquidity assessment
• Sentiment Analyst: Social media sentiment and "undiscovered" status
• News Analyst: Recent events, catalysts, and jurisdiction risks
• Fundamentals Analyst: Financial scoring and valuation

Research Team

• Bull Researcher: Advocates for BUY opportunities
• Bear Researcher: Identifies risks and thesis violations
• Research Manager: Synthesizes debate and enforces thesis compliance

Execution Team

• Trader: Proposes execution parameters
• Risk Team (Risky/Safe/Neutral): Debates position sizing
• Portfolio Manager: Final authority on all trading decisions

☁️ Deploy to Cloudflare Workers

The full guide — creating the D1 database, secrets, Git-connected builds, custom domains, rollbacks, and a deployment checklist — is at docs.autoinvestment.broker/docs/deployment.

The app runs on Cloudflare Workers via vinext — the Next.js API surface reimplemented as a Vite plugin — with D1 as the database, better-auth for authentication, and Email Workers for transactional email (verification, password reset, invitations).

# Local dev (vinext dev server; the server runs in workerd, so D1 & email
# bindings from wrangler.jsonc are live)
npm run dev

# Apply drizzle migrations to D1
npm run db:migrate:local    # local miniflare D1
npm run db:migrate:remote   # production D1 (ai-broker-db)

# Preview the production worker locally
npm run preview

# Deploy
npm run deploy

Setup notes:

• The build is configured in apps/ai-broker-web/vite.config.ts (there is no next.config): the vinext() plugin holds the Next.js options, and @cloudflare/vite-plugin runs the server environment in workerd. apps/ai-broker-web/worker/index.ts is the Worker entry — it delegates to vinext's fetch handler and adds the cron scheduled handler.
• Git-connected deploys (Workers Builds) run both the build command and the deploy command from the repository root, which has no Wrangler configuration of its own. npm run build therefore ends with node .github/scripts/write-root-deploy-config.mjs, which mirrors the .wrangler/deploy/config.json redirect that @cloudflare/vite-plugin writes inside apps/ai-broker-web up to the repository root, so npx wrangler deploy and npx wrangler versions upload resolve the built Worker instead of failing with "Missing entry-point to Worker script or to assets directory". Pointing the Workers Builds root directory at apps/ai-broker-web is the alternative, but then the workspace packages it depends on are no longer built.
• Bindings are declared in apps/ai-broker-web/wrangler.jsonc: DB (D1 database ai-broker-db), SEND_EMAIL (Email Workers), cron triggers for the /api/cron/* routes. Read them in server code with import { env } from "cloudflare:workers".
• Page-level ISR is served from the Cloudflare edge cache (cache: { cdn: cdnAdapter() } in vite.config.ts, "cache": { "enabled": true } in wrangler.jsonc). Add data: kvDataAdapter() and a VINEXT_KV_CACHE KV binding if the "use cache" data cache is needed.
• Environment variables for the app belong in apps/ai-broker-web/.env (vinext reads .env* files next to the app, following Next.js precedence). drizzle.config.ts additionally falls back to a root .env.
• Secrets: wrangler secret put BETTER_AUTH_SECRET (likewise GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET, STRIPE_SECRET_KEY, STRIPE_WEBHOOK_SECRET, CRON_SECRET).
• Admin area: /admin (user directory and database controls) is gated on , a comma-separated allowlist of addresses — . With it unset nobody is an admin, including the first registered user, so a deployment that forgets it simply has no admin rather than an open one.

📂 Third Party APIs

• Alpaca: Broker & trading as a service.
• TradingView: TradingView PineScript automated trading.
• Polymarket: Top event predictions leaderboard & correlation with stock market.
• Finnhub: Stock market data.
• Lightweight Charts: Candlestick UI widget from TradingView
• TradingAgents: Implementation of LLM research bots (paper)

🧰 Reference Bots

third-party-trading-bots/ vendors ten open-source trading bots, market-data services, and exchange clients as prior art. They are outside the turbo task graph and are never built or deployed — they are read for how other people size positions, batch LLM calls, model prediction markets, and normalize venue APIs. Each subdirectory keeps its own upstream license and README; the index describes what each one contributes.

Project	What it is
Kalshi-Vibe-Bot	Autonomous Kalshi binary-markets bot: local vetting, batched Gemini/Grok calls, Kelly sizing, stop-loss-only exits
Polymarket-BTC-15-Minute-Trading-Bot	Seven-phase NautilusTrader bot for 15-minute BTC markets with signal fusion and self-tuning weights
Prediction-Markets-Trading-Bot-Toolkits	Rust/Tokio execution core running ten strategies across seven venues through one adapter stack
ai-hedge-fund	Eighteen-agent research fund modeled on named investors; generates orders, never trades
debate-agents	LangGraph "bull vs. bear" debate graph — the ancestor of this repo's debate engine
fin-data-api-python	OpenBB Platform core plus ~35 data providers behind one fetcher contract
kalshi-bot-api	PyKalshi — typed Kalshi client with WebSocket streaming and pandas output

Contributors