AI Predictions on Polymarket: How Machine Learning Beats the Market

The Problem AI Solves

Polymarket has thousands of active markets and hundreds of smart money wallets making trades daily. No human can process all of this data in real-time. The challenge is: given that 150 smart money wallets hold positions across 50+ markets at any moment, which specific markets have the highest expected value for a retail trader to enter?

This requires synthesizing multiple data streams: wallet reputation (PnL, win rate), position sizes, market liquidity, current pricing, time to expiry, and the degree of agreement among independent wallets. AI is uniquely suited to this kind of multi-factor analysis.

Edge Radar’s AI doesn’t try to predict outcomes from scratch. Instead, it uses smart money positions as the primary signal and then applies its own probability estimation to calculate the edge. This hybrid approach — on-chain data as input, AI as evaluator — is what produces the 60.8% win rate.

The Pipeline: Step by Step

Our daily picks are generated twice daily at 14:00 and 22:00 UTC. Here’s the complete pipeline:

Step 1: Wallet Selection

The system selects approximately 150 unique smart money wallets from three tiers: top 100 by lifetime PnL (excluding market makers), top 100 by win rate (65%+ across 500+ trades), and top 50 by 7-day PnL (with positive lifetime profit and 500+ trades). After deduplication, this produces a curated list of the most reliable signal sources on the platform.

Step 2: Position Scanning

For each wallet, the system fetches their complete live positions from Polymarket’s Data API. With up to 8 concurrent connections and rate limiting to avoid API throttling, this scan completes in 2–3 minutes. Each position includes the market, outcome, size, current price, and expiry date.

Step 3: Market Filtering

Positions are filtered through multiple criteria:

• Market must expire within 12 hours (short-term resolution)
• Position value must exceed $1,000
• Current price must be between 10¢ and 90¢ (avoid extreme pricing)
• Market volume must exceed $10,000
• Sport whitelist: NBA and soccer only (highest SM alpha)
• No crypto markets (excluded since March 2026)

Step 4: Aggregation

Remaining positions are grouped by market and side. For each market, the system calculates: total SM volume on each side, number of SM wallets, dominant side (the side with more SM money), and the average entry price. Markets where multiple independent wallets agree on the same side receive higher priority.

Step 5: AI Probability Estimation

This is where Claude AI enters the pipeline. The model receives the top candidate markets sorted by SM volume, along with each market’s question, current price, SM wallet count, and volume. The AI’s job is to estimate the true probability of each outcome, independent of the current market price.

The AI doesn’t just output a probability — it also provides a confidence score and reasoning. For an NBA spread market, it might say: “Based on team records, recent form, and the injury report, I estimate a 70% probability that the Lakers cover -5.5, while the market prices it at 55%. The 15-cent edge reflects the market underweighting their home court advantage.”

Step 6: Edge Calculation and Confidence

Edge is calculated as: edge = AI_probability - market_price. Picks are then assigned confidence levels:

• HIGH — Edge ≥ 20¢ (the AI thinks the outcome is 20+ percentage points more likely than the market price)
• MEDIUM — Edge ≥ 10¢
• LOW — Edge ≥ 5¢
• SKIP — Edge < 5¢ (not enough edge to recommend)

Step 7: Dedup and Output

The final step ensures no more than one pick per game/event (the highest-edge pick wins) and outputs 3–5 picks. These are saved to the database, sent via Telegram to all active subscribers, and published on the Daily Picks page.

Why AI + Smart Money > AI Alone

Pure AI prediction models for sports betting typically achieve 52–55% accuracy — barely above the break-even threshold. The problem is that sports are inherently noisy, and the market is efficient enough to price in most publicly available information.

Smart money data provides what pure AI cannot: real-time information about what the most profitable traders believe. When a wallet with $3.7M in lifetime profit and 91% win rate takes a $25K position, they likely have information or analysis that isn’t yet fully priced into the market. This signal is what pushes our accuracy from the ~53% baseline to 60.8%.

The AI’s role is to evaluate and calibrate the smart money signal. Not every whale trade is equally informative. The AI considers the market context, pricing, time to expiry, and whether the smart money positioning aligns with fundamentals. This filtering is what prevents false signals from degrading overall accuracy.

Results and Transparency

Every pick is logged with its entry price, recommended side, confidence level, and reasoning. After resolution, the outcome (win/loss), final price, and PnL percentage are recorded. This complete history is available on the Daily Picks page.

We don’t cherry-pick results. Every pick generated by the system is tracked, including losses. The 60.8% figure includes all 102 resolved picks from the system’s inception through March 2026.

Algorithmic Fallback

If the Claude API is unavailable (rate limits, outages), the system falls back to a purely algorithmic pick generation method. This backup uses the same data pipeline but assigns confidence based on SM wallet count, volume, and historical hit rate of similar signals. The algorithmic fallback has a slightly lower but still positive expected value.

AI Wallet Analysis

Beyond daily picks, Edge Radar uses AI to analyze individual smart money wallets. Every day at 06:00 UTC, the system generates strategy analysis for top wallets, evaluating:

• Market focus (which categories they trade most)
• Trading style (aggressive vs. selective, large bets vs. many small ones)
• Risk management patterns (position sizing, diversification)
• Recent performance trends (hot streak, cold streak, or steady)

These analyses are available on each wallet’s profile page via the Top Wallets leaderboard.