prediction-market-portfolio

prediction-market-portfolio

A Claude skill that turns a prediction-market screenshot + your budget into a 3-tier portfolio recommendation. Multi-sport (football + NBA), multi-agent (News Scout + Quant Modeler + Portfolio Optimizer). Works on Claude mobile, claude.ai web, and Claude Code.

What it does

1. You share a screenshot of a Binance Prediction Market event (or Polymarket / Predict.fun / Kalshi). Optionally state a budget.
2. The skill parses the screenshot, identifies sport and outcomes, then runs two subagents in parallel:
   • News Scout — searches for injuries, lineups, suspensions, rule changes; converts to a Bayesian adjustment factor per team.
   • Quant Modeler — derives a base-rate probability per team using sport-specific math (Elo + Poisson for football, Elo + best-of-7 binomial for NBA). Deliberately ignores market prices.
3. The main thread ensembles the three signals (market, quant, news), checks for arbitrage, and builds three risk-tiered portfolios:
   • Tier 1 Conservative — pure arbitrage (Dutching) if positive
   • Tier 2 Balanced — value-tilted, max-loss capped at 15%
   • Tier 3 Aggressive — concentrated edge plays + lottery longshots
4. Adjusts for fees, slippage, and opportunity cost. Outputs a single recommendation with honest caveats.

Why screenshot input

Binance Prediction Market does not expose a public REST/WebSocket API, and its web pages are blocked by some browser-automation safety layers. Manual screenshot is the only universal path — and works the same on mobile, web, and desktop. The skill is built around this reality.

If you want auto-discovery without screenshots, you can adapt the skill to call Polymarket's public REST API (different liquidity pool, different prices — useful as a cross-reference).

Why multi-agent

Single-shot LLM analysis tends to anchor on the market price. Splitting into:

• News Scout (private info, market-blind for reasoning)
• Quant Modeler (base rate, market-blind)
• Optimizer (synthesizer with explicit weights)

reduces this anchor and lets you tune the ensemble weights when one signal is weaker (e.g., Quant Modeler that didn't fold injury info gets weight reduced automatically).

Sport coverage

Other sports (NFL, MLB, tennis, F1) can be added by writing a <sport>_specifics.md reference file.

Prerequisites

The skill auto-detects the environment and falls back gracefully.

Installation

Option 1: Claude Code (desktop)

git clone https://github.com/YunseobShin/prediction-market-portfolio.git ~/.claude/skills/prediction-market-portfolio

(Optional) Add PERPLEXITY_API_KEY to your user environment variables for enhanced News Scout. The companion perplexity-research skill handles the API call.

Option 2: Mobile / claude.ai web

1. Clone or download this repo
2. Zip the repo contents (top level should contain SKILL.md, references/)
3. Upload via Claude Settings → Capabilities → Skills

Repository structure

prediction-market-portfolio/         # repo root = skill folder
├── README.md                        # this file
├── LICENSE                          # MIT
├── SKILL.md                         # multi-agent orchestrator
└── references/
    ├── probability_models.md        # universal models + sport-specific math
    ├── portfolio_templates.md       # 3-tier formulas + worked examples
    ├── football_specifics.md        # football-only guidance
    ├── nba_specifics.md             # NBA-only guidance
    └── data_sources.md              # injury, Elo, market sources

Trigger phrases

Korean:

• "예측시장 분석해줘" (with screenshot attached)
• "이 배당으로 어떻게 베팅할까?" (with screenshot)
• "$500 있는데 이 마켓에서 포트폴리오 짜줘"

English:

• "Analyze this Binance Predict screenshot, $500 budget"
• "Build a 3-tier portfolio for this NBA Champion market"

The skill activates whenever a screenshot of a prediction market is attached, or when "예측시장 / prediction market / 배당 / portfolio" keywords appear with relevant context.

Output

For each parsed market:

• Market structure summary (arb %, N outcomes)
• News Scout findings table (form, injuries, adjustment factors)
• Probability comparison table (P_market vs P_quant vs P_news vs P_final)
• 3-tier portfolio with per-outcome P&L and SD
• Net EV after fees + slippage + opportunity cost
• One-line recommendation + 3 reasons + caveats

If all edges are < 3pp the recommendation is "skip this market" — the skill won't force a trade on an efficient market.

Disclaimer

Analytical framework, not investment advice. Prediction markets involve real financial risk. Positive expected value does not guarantee profit on any single trade. The skill does not execute trades — only produces analyses. For amounts under a few thousand dollars, traditional ETFs or stablecoin staking may statistically outperform after fees + variance.

License

MIT — see LICENSE.

Acknowledgements

Originally derived from a football-only skill used to analyze the 2025-26 UCL winner market on Binance. Generalized to multi-sport + multi-agent and validated against the 2026 NBA Western Conference Champion market on Polymarket (which correctly identified the market as essentially efficient — no value play recommended).