Finding the Edges of Prediction Markets
How Far Can Prediction Markets Expand? A Quantitative Liquidity Perspective
Prediction Markets work. 2025 showed that. 1 Polymarket called the US election when polls didn’t.2
They work by aligning financial incentives with truth-seeking. Put money where your conviction is. Contrarian bets that prove right get paid. Following the crowd into wrong answers costs you.
I came across a16z’s 2026 predictions. One made me stop:
Prediction markets go bigger, broader, and smarter.3
The a16z thesis: AI agents will provide attention and liquidity for long-tail prediction markets. Markets for everything. Bet on a TikTok star’s rise. Your colleague’s promotion. London housing prices. Your company’s quarterly performance. A political initiative’s success.
Anatoly Yakovenko, co-founder of Solana, pushes this further in a recent podcast:
As intelligence gets cheaper, you have more markets that become viable—everything being decided by market forces. [...] I can't fit all these markets in my head. [...] This is probably the most optimal direction for society to make decisions and move forward, the more market-based it is. [...] The forcing function of losing money is a good way to course-correct bad intelligence. 4
Allow me to summarize it: “More intelligence enables more markets.”
We launch meme coins at near-zero cost. 5 Why not prediction markets for everything?
Of course AI and agents need crypto. No bank accounts. Permissionless payments. That’s why this comes from the crypto corner of the internet.
But this isn’t a crypto post. Let’s go back to square one: When do prediction markets actually work? Technical requirements. Social hurdles. Will markets go broader? What are the limiting factors? Is this the a16z thesis true?
The history runs deep. 1503: betting on papal successors.6 Humans have always loved to wager on outcomes.
What is a prediction market? Binary outcome, fixed end date, bet on the result. Basically: binary dated options.
I’m talking about real-money platforms. With Polymarket and Kalshi as the by far largest ones by transactions and volume.7
Not Metaculus—that’s reputation-based, not real money. Real money changes everything.
Where do prediction markets work?
Two layers: technical and social.
Technical requirements:
Verifiable resolution criteria. Market liquidity.
Social requirements:
High expert bias environments. Moderate timelines. Regulatory acceptance.
Let me illustrate:
Technical requirements
Verifiable resolution.
Someone must decide if the event happened. This is the hardest problem in prediction markets: determining real-world truth in a provably fair way.
Polymarket uses UMA’s Optimistic Oracle—a hybrid decentralized system. Anyone can propose an outcome by posting a bond. Two-hour challenge period. If undisputed, it’s final. Challenged? Escalates to UMA token holders who vote. Takes 48-96 hours. This resolves 98.5% of cases without disputes.8
Kalshi uses centralized resolution. Platform staff verify against pre-specified sources (Bureau of Labor Statistics, Associated Press). Resolution in 1-12 hours. Simple. Fast. But you must trust the platform on ambiguous cases. 9
Take “Anthropic IPOs before 2027?” What counts as an IPO? Direct listing? SPAC? What if they get acquired instead? The oracle—whether decentralized voters or platform staff—must answer. Get it wrong, traders flee.
Clear, objective, measurable outcomes. No ambiguity. Without that, markets collapse.
Market liquidity.
You need traders. Tight spreads. Easy execution. Without liquidity, prices become fiction—wide spreads, massive slippage, nobody wants to trade.
Polymarket learned this the hard way. Started with automated market makers (AMMs) algorithmic pools that automatically quote prices. In the original design (Fixed Product Market Maker), liquidity providers deposited equal value of both outcome tokens, and the pool used a constant-product curve to price trades. Liquidity providers consistently lost money. Why? In prediction markets, one side resolves to $1, the other to $0. Unlike token swaps where both assets maintain tradeable value, this created structural losses for anyone providing liquidity—their losing tokens became worthless while winners walked away with the valuable ones. AMMs failed.
Late 2022: Polymarket pivoted to a central limit order book (CLOB)—the traditional exchange model where traders post buy and sell orders at specific prices. Off-chain matching for speed, on-chain settlement for security. Now professional market makers post resting orders, earning the bid-ask spread instead of suffering losses.10
This isn’t amateur territory. Susquehanna International Group, short SIG for anyone in quant finance—handling $2 trillion yearly in options—became Kalshi’s first institutional market maker in April 2024. Polymarket pays ~$433,000 monthly in liquidity rewards. Also check LinkedIn: here in Switzerland market making firms are hiring for prediction markets quants.
Kalshi operates a traditional centralized order book as a CFTC-regulated exchange, while Polymarket uses a hybrid-decentralized CLOB—off-chain matching with on-chain settlement via Polygon smart contracts. Both architectures succeed because they support professional market makers.
Other attempts and revival; PredictIt’s retail-only continuous double auction and Augur’s fully decentralized AMM both failed to attract sustained institutional liquidity, limiting their growth despite functional order mechanisms. 1112
Prediction markets need CLOBs, the ones without it don’t scale.
Social requirements:
High expert bias environments.
Politics. Sports. Corporate strategy. Anywhere emotions run hot and experts are reliably wrong. Traditional forecasting fails when status and hierarchy matter more than truth. Prediction markets cut through that—they don’t need any individual trader to be precise. As long as the group is unbiased on average, individual errors cancel out across enough participants. Financial stakes keep the crowd centered. Anonymity prevents coordinated bias. The result: collectively accurate prices from individually noisy bets.
Moderate timelines.
Too long, nobody cares. Lock up capital for five years on a sure bet? You’re losing to risk-free rates. Ever wonder why long-tail Polymarket markets don’t resolve to 0% even when the outcome is obvious? Because rational traders won’t tie up $10,000 for a year to make $100. Capital has opportunity cost.
Case in point: Will Jesus return in 2025? It took a year for odds to fall from 3% to 0%. Not because traders believed in the second coming. Because truth competes with risk-free returns elsewhere. This is proof markets work—they price in time.
Social acceptance.
Needed for large public markets. The US classifies them as gambling or unregistered securities. Kalshi fought the CFTC for years. Polymarket went hybrid—centralized operator, blockchain settlement. Most jurisdictions and organizations see prediction markets as threats. This kills most ideas before launch, under smoke and mirrors of gambling laws.
Some corollaries follow:
Corporate markets work technically, fail politically. Google’s Prophit (2005-2011) outperformed management forecasts by 20-25%. Got shut down. Revived as Gleangen (2020-present). HP, Ford, Intel—same story. Not accuracy problems. Political problems. Markets bypass hierarchy, surface uncomfortable truths. Management didn’t like that.
Dan Schwarz tells the full story: The Death and Life of Prediction Markets at Google.
The insider trading paradox. All betting faces scrutiny. See the Trump election narrative. But that’s where real value comes from—insiders, whistleblowers, people speaking uncomfortable truths. One person’s insider trader is another’s truth-teller.
Long-tail liquidity is hard. Just as crypto’s long tail struggles with liquidity, so do prediction markets. You can’t generate markets for everything. Anyone who’s traded illiquid markets knows—wide spreads, massive slippage, prices become fiction. This is why big platforms curate. Discord proposals, community suggestions, but internal teams decide: Is there enough interest? Will someone provide liquidity? Most ideas die here.
This last corollary brings us back to the thesis. It will enable us to place a lower bound on the size of prediction markets. But first: Does abundance of intelligence change any of that?
As enthusiastic as I am about AI, no. AI struggles without fast feedback loops. Software and finance work—tight loops, immediate results. Biology? Cells take time to grow. Human social dynamics? Slow, messy, unpredictable. No quick test to run. That's why AI makes comparatively slow progress here.
LLMs are the mean of the internet’s collective intellect. Not contrarian. Not sober-minded. Not what prediction markets need. Call me an AI blasphemer.
But. If API credits are paid, AI cares 24/7 about Zelensky wearing a suit or not. Scrapes all social media. Interprets new narratives constantly. That’s not more intelligence. That’s more attention.
Let me reshape the a16z thesis:
Not “More Intelligence → More Markets.”
Rather: “More Attention → More Markets.”
Now this thesis I do see. Prediction markets will grow into social and political narratives. Endless social media chatter, ever-changing, hard to quantify with traditional sentiment analysis. LLMs’ flexibility lets them interpret these narratives. Automated betting on narrative shifts.
This is what liquidity providers will do for long-tail markets. AI enables smaller markets to exist. Whether AIs trade or make markets doesn’t matter—they provide the attention these markets need to survive.
But limits remain. AI provides plumbing. Humans provide edge. The exciting movements still come from insiders and contrarian thinkers.
So where’s the edge? Not in the technology—CLOBs have existed for decades. The question is: which markets can sustain professional liquidity?
Crypto markets offer a cautionary tale. The long tail of token markets has struggled with liquidity for years. Let’s use them to map what’s viable.
What are the smallest viable markets on major crypto exchanges? I pulled data from Binance, Coinbase, Kraken, and KuCoin—not the top by volume, but a representative sample. Binance: the gold standard, where well-funded projects list if they want centralized liquidity. Coinbase: user-friendly, US-regulated. Kraken: trusted, established. Ages 8-14 years. All run successful CLOBs. A good analogy for prediction markets.
Who defines the edge of viable markets? Market maker economics. Each new market means infrastructure overhead, coordination costs. Is it worth operating there?
The simplest proxy for market maker profit:
Profit ≈ Volume × Spread − Fees − Operating Costs
Every trade crosses the spread. Oversimplified, yes—but it gives the order of magnitude. The long tail of crypto markets already shows where this breaks down.
Here’s the data as of January 10-11, 2026.13 Bid-ask spread versus 24-hour volume for prediction markets and major crypto exchanges.
(full resolution plot here.)
The differences are striking.
Prediction market median spread is 95x crypto median spread.
Breaking it down by venue:
Prediction markets see 2-3 orders of magnitude less volume than crypto
Polymarket median spread: ~200 bps. Kalshi median spread: 903 bps. That’s 9%—prohibitive.
Polymarket respects natural CLOB limits. Kalshi pushes further with more markets, but traders pay for it.
Using our naive formula: Binance’s lower quartile shows 13 bps spread × $111,765 daily volume ≈ $145/day profit potential. With good automation, that’s viable.
The chart reveals the required diagonal. Crypto CLOBs cluster lower-right: high volume, tight spreads. Polymarket’s top markets barely reach Binance’s lowest quartile. Kalshi chose a different path—high spreads, broad coverage.
The tradeoff: more markets are possible if traders pay wider spreads. Kalshi allows broader coverage at higher cost. But will you bet if placing an order costs 9% immediately? I wouldn’t.
So which prediction markets are viable? Push the threshold: $50,000 daily volume with reasonable spreads. Someone will make that market. Below that? Probably not.
A note: Polymarket remains unprofitable despite $15 billion in cumulative volume—subsidized by $2.3 billion in VC.14 The market for markets is still being made.
To answer a16z: we’re already testing the limits of liquidity. We won’t have markets for everything. AI doesn’t change that.
Personally? I’d like to see corporate and scientific reproducibility markets return. A radical declaration of transparency and meritocracy. Most academic computational research is vapor. Or never materializes. Let people bet on it.
Why do prediction markets excite me? They capture a true feedback loop. Unapologetic. No permission needed. Technical dynamics meet social limits. When markets work, they surface truth.
Data and methodology for this analysis:
github.com/gordonkoehn/prediction-market-microstructure
https://www.theblock.co/post/333050/polymarkets-huge-year-9-billion-in-volume-and-314000-active-traders-redefine-prediction-markets
https://arxiv.org/html/2507.08921v1
https://www.coindesk.com/coindesk-news/2025/12/10/most-influential-pump-fun
Rhode, Paul; Strumpf, Koleman (2008). [”Historical Election Betting Markets: An International Perspective”](http://users.wfu.edu/strumpks/papers/Int_Election_Betting_Formatted_FINAL_NoComments.pdf) (PDF). _Perspectives on Politics_.
https://dune.com/datadashboards/prediction-markets
https://legacy-docs.polymarket.com/polymarket-+-uma
https://help.kalshi.com/markets/markets-101/market-outcomes
https://gamblingharm.org/wp-content/uploads/2025/11/Polymarket-Wash-Trading-Study.pdf
https://wifpr.wharton.upenn.edu/blog/a-primer-on-prediction-markets/
11: https://www.panewslab.com/en/articles/0e22d6dd-1044-4f29-8074-0eefb0d54195
Data sources: Cryptocurrency: CoinGecko, Prediction Markets: Polymarket (Gamma API), Kalshi (Public API) - Data Collection: January 10-11, 2026
https://www.cbsnews.com/news/polymarket-ceo-shayne-coplan-online-betting-platform-60-minutes-transcript/