UniPat AI releases EchoZ prediction model with a 63% success rate in live trading on Polymarket, "surpassing human traders"

PolyMarket’s annual trading volume has already reached several billions of dollars, but more than 90% of traders lose money long-term (Dune Analytics, March 2026). In a game centered on “predicting the future,” most people are simply paying for the decisions of a small number of better decision-makers.

If who’s better at judging probabilities determines the outcome, then the question becomes: can this capability be replicated?

UniPat AI’s EchoZ-1.0 provides a quantifiable answer to exactly that. In a comparison against Polymarket’s human traders, its win rate on political questions reaches 63.2%, and 59.3% in long-term forecasting. The team built 5 EchoZ Agents for live trading: 4 of them were profitable, and the best performer achieved a 15% return within one week.

This isn’t the result of “trading技巧.” It’s more like the spillover of model capability. UniPat AI’s core members come from major model teams such as Qianwen, Kimi, Xiaomi, and Seed, and have long worked on building reasoning models and complex decision systems. In a prediction market—an environment fundamentally driven by “probability games”—they attempt to systematically replace intuition with models and repeatedly validate this capability in real markets.

More importantly, this is not just a model that looks impressive in a report. It’s a predictive capability that can be called directly. UniPat AI is productizing EchoZ and plans to open it up to the public in the form of an API. For developers and institutions, this means that in the future they can directly input a question and receive a complete output including conclusions, probability distributions, an evidence chain, and counterfactual analysis.

Before it’s truly open, a more valuable question to break down is: where exactly does EchoZ’s advantage come from?

What 63% win rate means

People who’ve done probability games know that in a zero-sum market where most people lose money, a statistical 60%+ win rate is a kind of advantage at a meaningful magnitude. Above 50% is positive expected value, and 60% is already enough to build stable profitability strategies.

EchoZ’s win rates against Polymarket human traders by scenario:

Politics and governance: 63.2%

Long-term forecasting (more than 7 days): 59.3%

High-uncertainty range (human confidence 55%-70%): 57.9%

The pattern is clear: the more hesitant humans are, the harder the scenario—long time horizons, multi-factor contests, and fragmented information—EchoZ’s advantage grows.

These are exactly the most valuable decision scenarios. Regulatory policy directions, macroeconomic variables, on-chain governance proposals, and token listing timing—all mostly fall into high-uncertainty, long-horizon, multi-factor intertwined problems. Whoever can consistently make more accurate probability judgments in these scenarios has alpha.

EchoZ ranks #1 on the General AI Prediction Leaderboard with an Elo of 1034.2, ahead of Gemini-3.1-Pro (1032.2), Claude-Opus-4.6 (1017.2), and GPT-5.2. The leaderboard covers 12 models, 7 domains, and 1,000+ active questions.

Is this ranking credible?

If you build your own leaderboard, the first reaction is always, “you’re just giving yourself an award.” UniPat AI did something very Crypto Native: all data is public.

All prediction questions, the probability distributions output by the models, and the final settlement results are all publicly available at echo.unipat.ai, so anyone can trace and verify them.

In addition, four sets of stress tests are also disclosed:

Adjust the core parameters of the scoring framework (σ from 0.01 to 0.50, 9 total groups). Under all settings, EchoZ ranks first; it’s the only model with zero ranking fluctuation. GPT-5.2 fluctuates significantly between ranks 2 and 9.

Randomly drop 10%-70% of the data—the ranking remains stable.

Remove models 1-6 from the leaderboard—the remaining order is almost unchanged.

After a new model is added, it converges to a stable ranking in 5.4 days.

Transparent, verifiable, resistant to interference.

How does it make money?

EchoZ autonomously searches for information, reads news, queries data, and then outputs a structured prediction report: probability distribution, evidence chain, and judgment rationale—every step of reasoning is traceable.

Look at three real cases:

NVIDIA market-cap prediction. On March 18, 2026, EchoZ answered the question “Which company will have the highest global market cap on March 31?” and gave NVIDIA an 98% probability. The basis for the judgment isn’t a single piece of information, but multiple independent evidence chains cross-validating each other: NVIDIA market cap ~$4.43T-$4.45T, leading Alphabet and Apple by about $4.43T; within nine trading days it’s almost impossible to be caught up; on March 13, the U.S. Department of Commerce withdrew AI chip export-control rule changes, eliminating the biggest regulatory risk before the target date; implied volatility in the options market is only ±1.98%, and the derivatives market does not experience a crash in a single pricing event that could wipe out a 15% lead advantage; Qatar’s helium facility shutdown creates supply-chain risk, but TSMC has not shut down. Four strands of evidence lock in the conclusion from four dimensions: market-cap math, regulation, derivatives pricing, and supply chain.

ETH new-high prediction. On March 18, 2026, EchoZ answered “Will ETH/USDT set a new all-time high before March 31?” and gave a 99% probability—predicting “No.” The reasoning chain is clear: the current price is around $2,220-$2,340, with the all-time high at $4,956.78; within 13 days it would need a 112%-123% increase; the U.S. Federal Reserve keeps interest rates unchanged at 3.50%-3.75% combined with the U.S.-Iran conflict, suppressing any explosive rally in risk assets; USDT is anchored as a stablecoin, and Binance’s ETH/USDT depth is ample (with $35M liquidity within a 2% price range), ruling out nominal price anomalies caused by stablecoin de-anchoring. Three independent evidence chains cross-validate each other, and Polymarket consensus also gives <1% probability.

NBA Western conference #1 seed prediction. Again on March 18, EchoZ predicted the NBA Western conference #1 seed for the 2025-26 season and gave the Thunder a 89.9% probability. Core logic: the Thunder are 54-15, leading the Spurs by 3 games, with each team having 13 games remaining; although the Spurs hold the head-to-head advantage (4-1), all they need to do is match that—but the Spurs face the hardest remaining schedule in the entire league (opponents’ win rate .560); the Thunder’s magic number is only 11, so they just need to perform normally to lock it in. The Lakers can at most reach 57 wins; mathematically they’re already out, confirming this is the battle between the two teams.

The key is that these predictions are not selected after the fact. For each question, the prediction timestamp, probability output, and settlement result are all publicly verifiable.

Why can’t GPT, Claude do this?

Simply put, the training approach is different.

Mainstream large models are trained for prediction using historical data, but historical data has two problems: when the model searches the web, it can easily “bump into” the answer (data leakage), and real-world randomness can cause the model to learn noise—when a good analysis meets a black swan it gets punished, while a blind guess meeting good luck gets rewarded.

EchoZ’s training paradigm is Train-on-Future: it directly asks the model to predict events that haven’t happened yet, evaluating the quality of the reasoning process rather than waiting for the answers to be revealed. A good analyst can also be wrong occasionally, but with a high long-term win rate—EchoZ’s training logic is the same.

But who defines what “good reasoning” means? The differences are huge across domains. UniPat’s approach is data-driven search for scoring standards (Rubric Search): prepare a set of candidate scoring dimensions, score and rank the model’s reasoning process with these dimensions, then compare against Elo rankings based on real outcomes—the higher the alignment, the closer the standard is to the real characteristics of “good reasoning.” Search separately by domain, and optimize each iteration.

The results that come back are very interesting. In the politics domain, the best scoring standard has 20 dimensions, including “absence signal recognition”—whether the model treats “nothing happened” as an important signal (for example, no new court filings, no new military communiqués, which in itself is information). There’s also “distinguishing statements and actions”—separating politicians’ verbal claims on social media from the actual execution steps entering legal processes. These dimensions are all discovered through data; human intuition alone wouldn’t think of this level of granularity.

What can you do after the API is opened?

Prediction API will soon be available for enterprises and developers. It supports asking a prediction question in natural language and returns a complete structured report:

Probability distribution: quantitative judgments of the event’s various outcomes

Evidence chain: multiple independent evidence items supporting the judgment, ordered by weight

Counterfactual analysis: how probabilities move when key variables change

Monitoring recommendations: signals to continuously track and trigger conditions

For exchanges and prediction market platforms, this means they can directly provide an AI prediction layer to users. When users browse a prediction contract, they can see EchoZ’s probability judgments, key rationales, and critical variables right beside it. For quant teams, these structured probability outputs can be directly used as strategy factors. For DeFi protocols, event probability becomes a brand-new on-chain data dimension—condition-triggered options, prediction-based insurance pricing, dynamic risk-control parameters. Currently, there are almost no reliable on-chain event probability data sources, and that’s exactly the gap EchoZ is trying to fill.

This is a new category: prediction capability as a callable infrastructure.

Why are these people doing this?

UniPat AI’s core team comes from top large-model teams such as Qianwen, Kimi, Xiaomi, and Seed, with more than ten researchers. Their focus is on reinforcement learning, Agent systems, data synthesis, and model evaluation. They have already received support from multiple leading USD funds.

This team composition explains Echo’s product form. Building prediction intelligence needs to solve three problems at once: how to train it (RL + process rewards), how to evaluate it (a dynamic evaluation system), and how to make the model find information itself to make judgments (Agents). These three tasks align exactly with the three directions the team is strongest at.

They choose to build prediction infrastructure because prediction capability is naturally quantifiable, verifiable, and profitable—one of the few large-model categories that can directly connect to commercial value.

UniPat AI states: “Prediction capability is one of the few AI abilities that can directly be tied to commercial value. When probability judgments can be structured, validated, and called, it will become a foundational input into trading and financial systems.”

Next steps

In the past few years, capabilities that have been API-ized include text, images, and code.

The next capability to be API-ized may be judgments about uncertainty itself. When probability judgments about the future become a callable, integratable, verifiable parameter, the decision pathways it can embed—trading strategies, risk-control models, product pricing, compliance early warnings—are far broader than prediction markets themselves.

Echo’s goal can be summarized in one sentence: turn “what will happen next in the world” into developer-callable input.

ECHO official website:

Technical blog：