194 lines
10 KiB
Markdown
194 lines
10 KiB
Markdown
# Intelligence Report: Polymarket Arbitrage & Edge Detection Bot (spark-003)
|
||
|
||
**Analyst:** ARI | **Date:** 2026-02-14 | **Classification:** INVESTMENT THESIS EVALUATION
|
||
|
||
---
|
||
|
||
## VERDICT: HOLD — Technically Feasible but Structurally Disadvantaged at This Capital Level
|
||
|
||
**Conviction: 4/10**
|
||
|
||
---
|
||
|
||
## 1. CONTEXT
|
||
|
||
Case proposes building an AI-powered bot that identifies mispriced Polymarket prediction markets by comparing market odds against Claude-estimated probabilities derived from news scraping. Paper trade first, then deploy $200-500 real capital.
|
||
|
||
---
|
||
|
||
## 2. FINDINGS
|
||
|
||
### 2.1 Market Size, Volume & Liquidity
|
||
|
||
**[HIGH CONFIDENCE]**
|
||
|
||
- Polymarket is valued at **$9 billion** (Feb 2026), backed by ICE's $2B investment (Oct 2025)
|
||
- Top markets routinely see **$1M-$10M+ daily volume** (confirmed via live API pull: govt shutdown market had $6.4M 24hr volume)
|
||
- Liquidity per market varies wildly: top political/macro markets have **$100K-$500K+ liquidity pools**; long-tail markets have **$5K-$50K**
|
||
- Market operates on Polygon blockchain using USDC — transactions are cheap (~$0.01) but require crypto wallet setup
|
||
- Official Python CLOB client (`py-clob-client`) available on PyPI — well-documented, supports read-only and trading
|
||
- Minimum order size: **$5**
|
||
|
||
### 2.2 Legal/Regulatory Status — US Users
|
||
|
||
**[HIGH CONFIDENCE — THIS IS THE CRITICAL RISK]**
|
||
|
||
- Polymarket was **fined $1.4M by CFTC** in Jan 2022, received cease & desist
|
||
- **Blocked US users from 2022 to December 2, 2025**
|
||
- Trump administration eased regulatory environment; CFTC/DOJ ended probe in July 2025
|
||
- Donald Trump Jr. is now an **advisor** to Polymarket; 1789 Capital invested
|
||
- US access is **re-opened** but regulatory status remains a **legal gray area**
|
||
- FBI raided founder Shayne Coplan's home in Nov 2024 over US user access violations
|
||
- Polymarket is **banned in France, Singapore, Switzerland, Poland** — regulatory risk is real and ongoing
|
||
- Insider trading concerns flagged by Rep. Ritchie Torres — a Jan 2026 account made $400K+ on Venezuela strikes positions, raising insider trading scrutiny
|
||
- **Bottom line:** US users CAN trade now, but automated bot trading in a gray-area crypto prediction market adds legal risk layers. Not illegal, but not clearly legal either. No explicit regulatory blessing for algorithmic trading on prediction markets.
|
||
|
||
### 2.3 Existing Competitors & Market Sophistication
|
||
|
||
**[MEDIUM CONFIDENCE]**
|
||
|
||
- The 2024 election showed **sophisticated players already exist**: one French trader controlled 4 accounts, placed $30M in Trump bets, won $85M
|
||
- Nate Silver (FiveThirtyEight founder) is a **Polymarket advisor** — the smartest probability modelers in the world are already watching these markets
|
||
- Known competitor tools/approaches:
|
||
- **Superforecaster communities** (Good Judgment Project) actively trade prediction markets
|
||
- **Quantitative hedge funds** are entering prediction markets as the space scales
|
||
- **Open-source bots** exist on GitHub for Polymarket automated trading
|
||
- **Market makers** provide liquidity algorithmically — they're the counterparty, and they're sophisticated
|
||
- **Key insight:** The "participants trade on vibes" thesis is **increasingly outdated**. As Polymarket hit $9B valuation and $3.3B was wagered on the 2024 election alone, professional capital has entered. The easy edges are being arbed away.
|
||
|
||
### 2.4 AI vs Market Accuracy
|
||
|
||
**[MEDIUM CONFIDENCE — CONFLICTING SIGNALS]**
|
||
|
||
- Prediction markets have historically been **more accurate than polls and pundits** for binary political outcomes
|
||
- However, markets have shown **clear mispricings**:
|
||
- VP pick 2024: Market had Shapiro at 68%, Walz at 23% — Walz was picked (market was wrong)
|
||
- Trump whale trades shifted odds by 10+ points beyond fundamentals — Silver himself called it "larger swing than justified"
|
||
- AI probability estimation faces challenges:
|
||
- LLMs can synthesize news quickly but **lack calibration** — they don't have trained probability distributions
|
||
- Academic research on LLM calibration is mixed; GPT-4/Claude are decent at relative rankings but poor at precise probability assignment
|
||
- **The edge would come from speed** (reacting to news faster than the market adjusts) more than accuracy
|
||
- Markets with **less attention** (long-tail, non-political) are more likely to be mispriced — but also have **less liquidity**, limiting profit potential
|
||
|
||
### 2.5 Realistic Profit Potential — $200-500 Capital
|
||
|
||
**[HIGH CONFIDENCE — THIS IS THE DEALBREAKER]**
|
||
|
||
Let's do the math honestly:
|
||
|
||
| Scenario | Edge | Capital Deployed | Positions/Mo | Monthly Return | Annualized |
|
||
|----------|------|-----------------|-------------|----------------|------------|
|
||
| Optimistic | 10% edge | $500 across 20 positions | 20 | $50-75 | $600-900 |
|
||
| Realistic | 5% edge | $400 across 15 positions | 15 | $20-30 | $240-360 |
|
||
| After losses | 3% net edge | $300 effective | 10 | $9-15 | $108-180 |
|
||
|
||
**Problems at this capital level:**
|
||
- $5 minimum order × 20 positions = $100 deployed, $400 idle — **capital utilization is terrible**
|
||
- Many positions lock up for weeks/months until resolution — **illiquid capital**
|
||
- A single correlated loss wipes weeks of gains — **risk of ruin is high at small scale**
|
||
- Claude API costs for continuous news scraping + probability estimation: **$20-50/month** — potentially eating all profits
|
||
- Time investment to build, maintain, and monitor: **20-40 hours initial, 5-10 hrs/week ongoing**
|
||
- **Effective hourly rate at realistic returns: $1-3/hour**
|
||
|
||
**Comparison:** $500 in a high-yield savings account = ~$25/year risk-free. The bot needs to deliver >5% annual return *after costs* just to beat a savings account. At $200-500 capital, the math doesn't work.
|
||
|
||
### 2.6 Technical Feasibility
|
||
|
||
**[HIGH CONFIDENCE]**
|
||
|
||
The good news — this is very buildable:
|
||
|
||
- **Polymarket API** is well-documented with official Python client
|
||
- **Gamma API** provides market metadata (free, no auth for reads)
|
||
- **CLOB API** supports order placement with API keys
|
||
- Feed Hunter infrastructure provides news scraping capability
|
||
- Claude API handles probability estimation
|
||
- Pipeline: `Scrape markets → Scrape related news → Claude estimates probability → Compare to market price → Flag divergences > threshold → Paper trade / execute`
|
||
- **Build time estimate:** 2-3 weeks for MVP including paper trading
|
||
- **Infrastructure cost:** Near-zero incremental (runs on existing homelab)
|
||
- Polygon gas fees negligible (~$0.01/tx)
|
||
|
||
The technical feasibility is **not the bottleneck**. The economics are.
|
||
|
||
---
|
||
|
||
## 3. ANALYSIS
|
||
|
||
### Strengths
|
||
- Technically straightforward to build with existing infrastructure
|
||
- Polymarket is now legally accessible to US users (for now)
|
||
- AI news analysis provides genuine informational edge on speed
|
||
- Paper trading phase limits downside risk
|
||
- Excellent learning project for prediction market mechanics
|
||
|
||
### Weaknesses
|
||
- **Capital is 100-1000x too small** for meaningful returns
|
||
- Professional capital is already in the market
|
||
- Legal status remains gray — automated bot trading adds risk
|
||
- AI probability calibration is unproven vs market consensus
|
||
- Capital lockup in illiquid positions
|
||
|
||
### Opportunities
|
||
- Long-tail markets with low attention may have persistent mispricings
|
||
- Cross-platform arbitrage (Polymarket vs Kalshi vs Manifold) could offer structural edges
|
||
- If paper trading validates edge, could scale capital later
|
||
|
||
### Threats
|
||
- Regulatory reversal (next administration could re-ban)
|
||
- Platform risk (Polymarket could restrict API access or bot trading)
|
||
- Market sophistication increasing rapidly as institutional money enters
|
||
- Correlated event risk (e.g., all political bets go wrong together)
|
||
|
||
---
|
||
|
||
## 4. CONFIDENCE ASSESSMENT
|
||
|
||
| Claim | Confidence |
|
||
|-------|-----------|
|
||
| Polymarket is accessible and liquid | HIGH |
|
||
| US legal status is gray but currently allowed | HIGH |
|
||
| AI can identify some mispricings | MEDIUM |
|
||
| $200-500 is insufficient for meaningful returns | HIGH |
|
||
| Sophisticated competitors already exist | MEDIUM |
|
||
| Technical build is feasible | HIGH |
|
||
|
||
---
|
||
|
||
## 5. SO WHAT
|
||
|
||
This is a **technically cool but economically unviable idea at the proposed capital level**. The math is brutal: even with a genuine 5-10% edge (which is optimistic), $200-500 deployed across prediction markets with weeks-long lockups yields **$10-75/month before API costs**. After accounting for Claude API usage and time invested, the effective hourly rate is below minimum wage.
|
||
|
||
The idea becomes interesting at **$5,000-10,000+ capital**, where a 5% monthly edge yields $250-500/month — enough to justify the time investment and cover API costs. But at that point, you're taking real financial risk in a quasi-regulated market.
|
||
|
||
**Recommendation: HOLD** — Don't build this now. If Case wants to explore prediction markets:
|
||
1. Paper trade manually for 30 days to validate intuition
|
||
2. Track whether Claude probability estimates actually diverge meaningfully from market prices
|
||
3. If manual paper trading shows consistent edge, revisit with $2,000+ capital allocation
|
||
4. Prioritize spark-002 (consulting) and spark-006 (QA service) which have 10-50x better return on time invested
|
||
|
||
---
|
||
|
||
## 6. MONEY
|
||
|
||
| Metric | Value |
|
||
|--------|-------|
|
||
| Build Cost | $0 (existing infra) + 20-40 hrs labor |
|
||
| Ongoing Cost | $20-50/mo (Claude API) |
|
||
| Monthly Revenue (realistic) | $10-30 at $500 capital |
|
||
| Monthly Revenue (optimistic) | $50-75 at $500 capital |
|
||
| Break-even Timeline | Never at this capital level |
|
||
| ROI vs Alternatives | spark-002 consulting: $2K+/mo. This: $20/mo. No contest. |
|
||
|
||
---
|
||
|
||
## FOLLOW-UP VECTORS
|
||
|
||
1. **Cross-platform arbitrage analysis** — Compare identical markets on Polymarket vs Kalshi vs Manifold for structural price differences (potentially more profitable than single-platform edge detection)
|
||
2. **Backtest Claude probability estimation** — Feed Claude 100 historical resolved markets, compare its estimates to final outcomes vs market odds at time of estimation. This would validate/invalidate the core thesis with zero capital risk.
|
||
3. **Monitor regulatory developments** — CFTC stance under current administration could shift; any enforcement action would be a kill signal for this idea
|
||
|
||
---
|
||
|
||
*Report generated by ARI | Research & Intelligence Division*
|
||
*Sources: Polymarket API (live), Wikipedia, Polymarket documentation, Gamma API*
|