Prediction markets have gone from niche experiment to mainstream trading venue in under two years. Polymarket processed over $1 billion in volume during the 2024 US election. Kalshi received CFTC approval for event contracts. DraftKings and FanDuel now offer sports prediction markets alongside traditional sportsbooks.
But how accurate are they? Where is the liquidity? And can you actually make money trading them?
We set out to answer these questions using 25.6 million in-game data points, 1,100+ live trades across 9 sports, and cross-venue price comparisons from 6 different platforms. This is what we found.
Market Accuracy: How Well Do Prices Predict Outcomes?
Prediction market prices are implied probabilities. A contract trading at 65 cents implies a 65% chance of the event occurring. If the market is perfectly calibrated, events priced at 65% should happen 65% of the time.
We tested this by collecting closing prices (final price before settlement) across 60,000+ sports events and comparing to outcomes.
Finding 1: Sports prediction markets are well-calibrated overall.
Across all sports and platforms, the average calibration error is 2-3%. Events priced at 70% happen about 68-72% of the time. This is consistent with prior academic research on prediction market accuracy, including work by Justin Wolfers and Eric Zitzewitz and the Iowa Electronic Markets.
Finding 2: Calibration breaks down during live events.
Pre-game prices are more accurate than in-game prices. During live play, market prices lag reality by 10-30 seconds after score changes. This lag creates systematic mispricings:
- After NBA scoring events, prices adjust within 15-25 seconds
- After NFL touchdowns, adjustment takes 20-40 seconds
- After NHL goals, adjustment takes 30-60 seconds (thinner markets)
These windows are where algorithmic traders extract value. Our models detect approximately 5-15 actionable mispricings per game night across all sports.
Finding 3: Different venues have different calibration profiles.
| Venue | Calibration Error | Update Speed | Liquidity |
|---|---|---|---|
| DraftKings | 1.5% | Fast (5-10s) | Deep |
| FanDuel | 1.8% | Fast (5-10s) | Deep |
| Polymarket | 3.2% | Slow (15-30s) | Thin |
| BetMGM | 2.1% | Medium (10-15s) | Medium |
Sportsbooks (DraftKings, FanDuel) are better calibrated because they have more data, more sophisticated models, and more liquidity. Polymarket has wider mispricings but thinner books, which creates a tradeoff between edge size and execution quality.
Liquidity: Where the Money Actually Is
Liquidity determines whether you can execute trades at the prices you see. A market with a 10-cent edge but $50 of depth is worth less than a market with a 3-cent edge and $50,000 of depth.
Finding 4: Traditional sports on sportsbooks have 100-1000x more liquidity than prediction markets.
For a typical NBA game, DraftKings might have $500K-$2M in total handle. Polymarket might have $5K-$50K. This matters for position sizing and slippage.
Finding 5: Esports markets are the least liquid but most mispriced.
CS2 and League of Legends markets on Polymarket often have single-digit thousands in total volume. But the mispricings are the widest — 15-40 cent edges are common because so few sophisticated market makers cover these events.
Finding 6: Only 5.2% of live game time is actually tradable.
From our microstructure research on 13,568 market events: when you apply strict tradability filters (spread under 2 cents, depth over 5,000 contracts, quotes under 45 seconds old), only 5.2% of live game time qualifies. The other 94.8% looks tradable in historical data but isn't in practice.
What Actually Works (And What Doesn't)
We've tested dozens of strategies across 9 sports. Here's the honest assessment based on 1,100+ live trades.
Strategies that work:
-
Hold-to-settlement moneyline trading using independent WP models: 67% win rate on MLB, 64% on NHL, 60% on NBA. The key insight is that you never sell — you hold until the game resolves at 0 or 100 cents. This eliminates exit timing risk and adverse selection on sells. Full strategy breakdown.
-
Cross-venue edge detection comparing Polymarket prices to devigged DraftKings and FanDuel lines: when multiple sportsbooks agree on a price and Polymarket diverges, the edge is on Polymarket's side almost every time. Multi-venue methodology.
-
Esports value betting on CS2 and LoL with Elo-based models: wide edges (15-40 cents) compensate for thin liquidity and high slippage. Best on tier-2 tournaments where fewer sharp bettors participate.
Strategies that don't work:
-
Mean-reversion taker strategies: Buying dips after score changes. 0% win rate in live testing. Sports markets reprice permanently — there's no reversion.
-
Spread/total active trading: Taking positions on spread and total markets with exit targets. Markets reprice on every score, making exit timing impossible.
-
Low-edge high-volume: Trading every signal above 3-4 cents. After fees (2%) and slippage (3-5 cents average), these are negative EV.
The Technology Stack
Building a competitive prediction market trading system requires surprisingly little infrastructure:
Data collection: ESPN's undocumented API provides free, real-time scores for every major sport. No API key required. Full ESPN API tutorial.
Model training: scikit-learn for logistic regression, XGBoost for gradient boosting, with isotonic calibration. Trained on 25.6 million in-game snapshots stored in Apache Parquet format.
Execution: Polymarket's CLOB API for order placement, with py-clob-client as the Python wrapper.
Odds comparison: The Odds API for real-time sportsbook prices from 6+ venues.
Infrastructure: A $7/month Hetzner VPS, Caddy for HTTPS, SQLite for storage, Cloudflare for CDN. Total cost: $13/month. Full infrastructure breakdown.
The Open Questions
Will prediction market accuracy improve as liquidity grows? More liquidity should mean tighter spreads and better calibration, which means fewer mispricings. But it also means more sophisticated market makers, which could create new types of edges (latency arb, cross-venue arb) that favor algorithmic traders.
Will regulatory clarity help or hurt? Kalshi's CFTC approval legitimized event contracts but also brought compliance costs. If Polymarket faces similar regulation, it could reduce the retail flow that creates mispricings.
Can ML models maintain edge as markets mature? Our models currently exploit the gap between ESPN-trained fair value and Polymarket prices. As more algorithmic traders enter, this gap should narrow. The sustainable edge may shift from model alpha to execution alpha — trading the same signals faster and with less slippage.
Methodology
- Data: 25.6 million in-game state snapshots across 7 sports (NBA, NFL, NCAAMB, CFB, NHL, MLB, NCAAWB), collected via ESPN API from 2020-2026. 60,000+ game outcomes for Elo computation.
- Live trades: 1,100+ trades on Polymarket from March 9 to April 21, 2026, across 5 automated bots covering 9 sports (including CS2, LoL, Soccer, Tennis).
- Cross-venue data: Sportsbook odds from The Odds API (DraftKings, FanDuel, BetMGM, Caesars) collected every 2 minutes during game hours.
- Microstructure data: 13,568 market events from Kalshi sports markets, ~30-second polling resolution, December 2025.
All live trading results are publicly verifiable on-chain via our Polymarket wallet and PolygonScan.
This research is published by ZenHodl. We build ML-powered sports prediction models and publish our results transparently. Full methodology at zenhodl.net/methodology. Live results at zenhodl.net/results. White paper at zenhodl.net/research.