The classical efficient market hypothesis says that all available information is already incorporated into market prices. If that were strictly true, no model could consistently beat the market and no trading bot could consistently profit. Sports prediction markets in 2026 are not strictly efficient — but they are closer than most retail traders assume.
This post walks through what we have learned about market efficiency from a year of on-chain trading on Polymarket. The data is not theoretical. Every trade is logged. Every settlement is verifiable. The conclusions are constrained to what the data actually supports.
The Headline: Mostly Efficient, Sometimes Not
The aggregate result of a year of trading is that the market is right most of the time and wrong sometimes. Our model agrees with the market on roughly 65% of contracts (within 5 cents). On the remaining 35%, the model is right slightly more often than the market — the basis of the edge that produces a profitable bot.
But the inefficiency is not uniform. It clusters in specific structural patterns, and identifying those patterns is more useful than computing an aggregate "efficiency score." Three patterns matter most.
Inefficiency Pattern 1: Information Latency
The most reliable source of edge is being faster than the market at incorporating new information. When a starting NBA point guard is ruled out 20 minutes before tip, the market reprices over the next several minutes as different traders process the news. A model that polls the official NBA injury feed and reprices in under 5 seconds can place trades against a market that has not yet caught up.
This is not skill in the deep sense. It is plumbing. Whoever has the lower-latency data pipeline captures the edge. We run our injury-news pipeline at sub-5-second latency from the NBA feed and capture about 0.8 cents of edge per trade from this single source. That sounds small. Across 1,500 NBA trades a season it adds up to real money.
The corollary: this edge will compress over time. As more bots subscribe to the same feeds, the latency advantage flattens. The edge from being five seconds ahead becomes the edge from being five hundred milliseconds ahead, and eventually disappears for any single source.
Inefficiency Pattern 2: Thin Liquidity at the Tails
Polymarket order books are deepest in the middle of the distribution. A contract trading at 50 cents typically has tens of thousands of dollars resting on each side within a 2-cent spread. A contract trading at 10 cents or 90 cents has dramatically thinner books — sometimes just a few hundred dollars at the best bid or ask.
Thin books produce systematic mispricing in two directions.
On heavy favorites (90+ cent contracts), retail traders looking for "safe" plays push the price slightly above fair. A model that has the discipline to short these contracts via the corresponding underdog purchase can capture 0.5-1.5 cents of edge. The catch is that the absolute return is small (you are paying 92 cents to receive 100 cents — a 8.7% return at best, before fees), so you need volume to make it worth the bot's attention.
On heavy underdogs (10-20 cent contracts), the opposite pattern emerges. Lottery-ticket buyers push prices slightly above fair. Same dynamic, opposite direction. A disciplined model can short these contracts via the favorite side. Same caveat about volume.
These are real but small inefficiencies. The bots that exploit them are slow and patient, not aggressive.
Inefficiency Pattern 3: Cross-Sport Attention Asymmetry
Polymarket has uneven attention across sports. Major leagues (NBA, NFL, English Premier League) have many active traders watching every game. The market converges to fair value within minutes of a price move.
Minor leagues and esports have far fewer active traders. NCAAMB has thousands of games per season, most with sparse market attention. Tennis tournaments outside the Grand Slams attract limited liquidity. CS2 minor events trade with single-digit-trader books.
The attention asymmetry produces edge for any model that can scale across many low-attention markets. A bot that can score every NCAAMB game in the season, identify the small subset with mispriced markets, and trade them efficiently captures more total edge than the same bot focused on NBA. NBA edges are smaller because the market is more efficient. NCAAMB edges are larger because nobody else is watching.
This is why we built coverage across 11 sports rather than focusing on the highest-volume leagues. The low-attention markets are where the surviving edge lives in 2026.
Where the Market Wins
The market is consistently right in three places worth knowing.
Closing prices. The Polymarket closing price (the volume-weighted middle price in the last 60-120 seconds before a contract becomes binary) is hard to beat. It incorporates all the information available to all market participants. A model that produces a fair probability at game start is competing with that closing price as its benchmark. Our Closing Line Value post explains why CLV against the close is the right edge metric.
Extreme model outputs. When our model produces a probability more than 20 cents from market, the market is right more often than we are. The model is over-confident in a region where it has insufficient training data, and the market knows better. Every model has this regime. The fix is to cap the maximum edge you will trade and walk away from the rest.
Late-game thin books. In the final 5 minutes of a game, market prices reflect orderbook information that our model does not have access to. Specific traders are loading up on specific sides for specific reasons. Trying to trade against the late-game market without ingesting the orderbook signal is reliably negative for us.
How Efficiency Has Changed in 14 Months
Two trends in our data:
Edges have compressed slightly. Our average CLV across all sports was 3.4 cents per trade in early 2025; today it is 2.8 cents. The decline is real. New traders are entering the market, and the obvious mispricings get arbitraged faster than they did a year ago.
Latency requirements have tightened. A bot that needed sub-30-second response in 2024 needs sub-15-second response in 2026 to capture the same edge. The competitive frontier on speed continues to advance. Anyone running a bot on a 1-minute polling cadence is leaving most of the edge on the table.
The directional read: the market is becoming more efficient over time. Edge that exists today may not exist in 18 months. Anything you build needs continuous monitoring (we use drift alerts) and continuous retraining.
The Bottom Line
Polymarket is mostly efficient, partially inefficient, and trending toward more efficient. The edges that exist in 2026 are smaller than they were in 2024, more concentrated in specific structural patterns, and more dependent on infrastructure (latency, coverage, multi-sport scaling) than on cleverness.
If you are building a trading system, build for the world where the median edge is 5-10 cents and the median trade is just barely profitable after costs. That is the world we operate in. Models that need 20-cent edges to be profitable will not survive the next year.
The model is the entry ticket. The infrastructure is the moat.
Full per-sport CLV decomposition at zenhodl.net/clv. Live trade ledger at zenhodl.net/results. The methodology behind these conclusions is published at zenhodl.net/methodology.