How to read a prediction market AI score — AI probability vs market price explained

An AI scoring service tells you a Polymarket question has 38% probability of resolving YES. The market is currently trading YES at 12¢. Should you bet? On which side? And how much? Most retail traders look at the AI score and either trust it blindly or dismiss it entirely. The honest interpretation lives in the gap between AI probability and market price — and the rules for reading that gap correctly are simpler than they look.

The two numbers you are comparing

On any binary prediction market, there are two probabilities at any moment:

• AI probability — what your model thinks the chance is that YES resolves true. Numbers between 0 and 1 (or 0% and 100%).
• Market-implied probability — what the current YES price says the consensus thinks. The market YES price IS the implied probability (a YES price of 0.30 means the market thinks YES has 30% probability).

The difference between these two is "edge." If AI probability is 0.38 and market YES is at 0.12, the edge is 0.38 − 0.12 = 0.26, or 26 percentage points. AI thinks YES is much more likely than the market does.

The decision tree

Three possible relationships, three different decisions:

Case 1: AI probability > market price + threshold → BUY YES

AI thinks YES is more likely than the market does. The mispricing is in YES's favor. You buy YES at the current market price.

Example: AI says 0.38, market YES at 0.12. Edge = +26pp. Above your action threshold (say 10pp). Recommendation: BUY YES.

If you're right, you bought a 38%-probability event at 12¢ — payout is $1 per share if YES, so net profit per share is $0.88. If you're wrong (NO resolves), you lose your $0.12 per share. Expected value: 0.38 × 0.88 − 0.62 × 0.12 = +$0.26 per share. Positive expected return; bet sized as you choose.

Case 2: AI probability < market price − threshold → BUY NO

AI thinks YES is less likely than the market does. The mispricing is in NO's favor. You buy NO at (1 − market YES price).

Example: AI says 0.60, market YES at 0.85 (NO at 0.15). Edge = −0.25, or 25pp toward NO. Above threshold. Recommendation: BUY NO.

If you're right (NO resolves), you bought a 40%-probability event (NO) at 15¢ — payout $1, net +$0.85. If wrong (YES resolves), you lose $0.15. Expected value: 0.40 × 0.85 − 0.60 × 0.15 = +$0.25 per share.

Case 3: |AI probability − market price| < threshold → SKIP

AI and market are close enough that the EV is small or zero. Not worth the bet.

Example: AI says 0.55, market YES at 0.52. Edge = +3pp. Below threshold. Skip.

Why threshold? Because every bet has implicit costs: spread, AMM slippage, fees on Polymarket, AI scoring error. A small edge gets eaten by those costs. The threshold is the size at which the math is robust to your assumptions about cost.

Why threshold matters more than people realize

Setting the action threshold too low is the most common error. If you bet on every 3-percentage-point edge, you're effectively betting on AI scoring noise. The AI probability is not perfectly calibrated — it has a margin of error around its point estimate.

A typical AI scoring model produces probabilities with ±5pp standard error. If you bet on every 5pp edge, half your bets are noise (the true edge could be zero or negative). If you bet on every 15pp edge, almost all your bets are real signal (the true edge is unlikely to be near zero).

Most production prediction market services target a 10pp threshold for actionable bets. Below that, the bet is too noisy to justify. Above that, the bet has high expected EV with reasonable robustness to scoring error.

The thing about underdog bets

The math in Case 1 above is striking: AI probability of 0.38 vs market 0.12. AI gives YES three times more likely than market does. If AI is right, win rate is 38% per bet. That sounds bad — most retail traders would never bet on something they expect to lose 62% of the time.

But each $1 bet wins +$7.33 (1/0.12 − 1) when YES resolves. So:

• 38 wins × $7.33 = +$278.54
• 62 losses × −$1 = −$62
• Net: +$216.54 per $100 staked. Per-bet ROI: +$2.17 or +217%

Same AI, same probability estimates, same bets — but with 38% directional accuracy you generate +217% ROI per bet. The headline accuracy looks awful; the economics are exceptional.

This is the core reason most retail traders should not run prediction market AI strategies on their own gut judgment. The math says bet underdogs; their psychology refuses to do so. We covered this in detail in why accuracy and profitability are two different things.

How to verify an AI is calibrated

A calibrated AI probability is one that matches actual frequencies over time. If the AI says "40% probability" on 100 markets, you'd expect roughly 40 of them to resolve YES.

The standard test: group AI outputs into deciles (0-10%, 10-20%, ..., 90-100%) and check that actual resolution frequencies match the bucket mean. A good calibration plot has dots lying on the diagonal. A bad one has dots way off.

If your AI service won't publish their calibration, that's a warning sign. Calibration is the most honest summary of probabilistic accuracy. We publish ours alongside resolution outcomes on our PREDICT history page so subscribers can verify directly.

The five-step interpretation routine

When an AI scoring service gives you a signal, run this checklist in 60 seconds:

1. Edge magnitude. Calculate |AI probability − market price|. If less than 10pp, skip. The bet is too marginal.
2. Edge direction. If AI > market → BUY YES. If AI < market → BUY NO. The signal is one direction or the other.
3. Market price interpretation. If the bet side price is below 20%, you're betting an underdog with asymmetric payoff. If above 80%, you're betting a heavy favorite with low payoff. Different risk profile, different sizing.
4. Liquidity check. A 20pp edge on a market with $200 liquidity is unrunnable. A 10pp edge on a market with $200k liquidity is excellent. Check the depth before sizing.
5. Resolution criteria. Read the market description once. AI scoring sometimes misreads the actual resolution rule. If the market criteria is ambiguous or AI seems to be misreading it, skip.

That's the entire decision routine. 60 seconds per signal. If you cannot do those five steps before placing a bet, do not place the bet.

What our own service does

PREDICT runs the equivalent of this five-step routine automatically on every market it scores. Markets that fail the threshold or liquidity tests get SKIP rather than BUY. Markets that pass get an actionable signal delivered to subscribers with all the relevant metadata: AI probability, market price, edge in basis points, recommendation, and the underlying reasoning.

This is the workflow we built so subscribers don't have to manually verify five things on every bet. The signal arrives with enough context that a glance is enough to act on. Trial is free.