How OddsShift Works
One verdict per market.
Graded when it resolves.
OddsShift reads the on-chain positions of 166 tracked Polymarket wallets, tiers them daily by their record on resolved markets, and gives every market where one of them holds a position a single AI verdict — BUY YES, BUY NO or HOLD — frozen before resolution and graded after. Today that is 162 Radar markets out of 791 tracked, and 206 of 295 graded BUY calls matched (April – September 2026). See the track record →
Not all money is equal.
Prediction markets aggregate thousands of bets into a single number — the odds. But that number mixes everything together: a tourist dropping pocket money, a bot arbitraging pennies, and a wallet whose record on resolved markets we have actually measured.
The odds you see on Polymarket are an average. An average of noise and informed money, mixed together. You can’t tell which is which from the price alone. That’s the problem we solve.
We read wallet balances on-chain. The record decides the tier.
We keep a list of 166 Polymarket wallets and read their positions straight from the blockchain: the balance of each outcome token in each wallet on Polygon, refreshed every 5 minutes. No screenshots, no self-reported trades — which side a wallet holds, and how much, is what the chain says it holds.
Every day we re-grade the whole list against the markets that have resolved. A wallet is tier-1 when it has been right on 80% or more of 15 or more resolved markets, and tier-2 at 70–80% on 10 or more. The bar is fixed; the membership is not. A wallet that stops being right stops being tier-1 the next morning.
The tracked wallets are an input to the verdict, never the verdict itself. Their leaderboard and per-wallet records live in Copy Trading →
How a verdict is built.
A tracked wallet holding a side is the starting point, not the answer. Every night, each Radar market — every tracked market where at least one tracked wallet holds a position — goes through two layers of models on top of the wallets’ side.
Three math models
We don’t say “this looks good.” We calculate a fair value.
Bayesian Update starts from the market price and moves it by how accurate the wallets on each side have been on resolved markets. PIN Model estimates how much of the market’s volume looks informed rather than retail. Compound Signal combines which side the tracked wallets are on, how large their positions are and their measured record into one score.
Together they produce a fair value for the market. The distance between that fair value and the current price is what a BUY verdict is built on.
- Input
- market price · tracked wallets’ side and size
- Bayesian Update
- price, moved by the wallets’ record
- PIN Model
- informed vs. retail share of volume
- Compound Signal
- side × size × record
Up to five AI models, one question
We don’t trust a single model. Each Radar market is put to up to five independent AI models, each with a different brief: a fundamentals analyst that updates probabilities as evidence arrives, a quant that looks at volume patterns and base rates, a contrarian that asks what the market is not pricing in, a cross-validator that weighs the evidence evenly, and a macro analyst that tracks the wider context. They see the market, the tracked wallets’ side and recent news for the event.
Only the models that answered on a given night count, and the market page shows which ones did. When they agree with the wallets and the math, that is a call. When they split, the market is genuinely uncertain — and we say so.
See tonight’s model runs on the Radar →| Model | Brief | Counts |
|---|---|---|
| Claude Analysis | fundamentals | when it answers |
| DeepSeek Quant | volume & base rates | when it answers |
| Grok Contrarian | what is not priced | when it answers |
| Gemini Validation | cross-check | when it answers |
| Kimi Macro | wider context | when it answers |
| Only responders are shown. A model that did not answer that night is left out of the verdict and off the page — never filled in. | ||
One verdict per market: BUY YES, BUY NO or HOLD.
The wallets’ side, the three math models and the AI models that answered collapse into one line on the market page. It is either a call — BUY YES or BUY NO — or it is HOLD, which means no entry.
A BUY comes with an entry (the price when the verdict was set), a target (where the fair value says the price should go) and the fair value itself. HOLD is shown so you can see that we looked and passed; it is never graded and never counted in the record.
- BUY YES
- tracked wallets on YES, fair value above the price · entry and target shown
- BUY NO
- tracked wallets on NO, fair value below the price · entry and target shown
- HOLD
- no entry — inputs disagree or fair value is too close to the price to pay · shown, never graded
Under the hood, every Radar market: 166 tracked wallets read on-chain, 3 math models (Bayesian Update, PIN Model, Compound Signal), up to 5 AI models with only the responders counted, and recent news for the event. Markets refreshed twice a day, verdicts nightly.
How the record is graded.
Every BUY YES and BUY NO is frozen before the market resolves — side, entry price, date. Nothing is edited afterwards. When the market resolves, the frozen call is graded against the outcome: matched or missed.
To turn that into money terms we hold the same fixed stake, $1,000 per call, to resolution and report the ROI across all graded calls. Misses are listed one by one, not netted away. HOLD verdicts are shown in the record but never graded.
| Series | Graded | Matched | Rate | ROI at $1,000/call |
|---|---|---|---|---|
| BUY verdicts vs. resolved outcome | 295 | 206 | 70% | -2.3% |
| HOLD (shown, not graded) | 167 | — | — | — |
206 of 295 graded BUY calls matched, with 120 BUY calls still open and 1,228 tracked markets resolved so far. The hit rate and the ROI are two different things: a call can match and still lose money if the entry was expensive. Both are published, as of September 10, 2026. Every graded call, including the misses, is on the track record → and month by month on the scoreboard →
What is free, what is PRO.
Every tracked market page with price, volume and news. Whether a verdict exists. The full track record and the monthly scoreboards, misses included. The wallet leaderboard.
The verdict itself with entry and target. Model fair value. Which tracked wallets hold each market, with entries and P&L. Wallet follow, and alerts within about a minute of a tracked wallet’s trade.
Of 791 tracked markets, 162 carry a verdict today. The rest show price, volume and news and say “no verdict” plainly — no tracked wallet holds them, so there is nothing to call.
Copy Trading is a separate product: the tracked wallets themselves — leaderboard, per-wallet pages with their record, and follow alerts within a minute of a trade on PRO. X-Tweet Market is another: tweet-count forecasts per persona, graded on its own pages.
Most tools show you the odds. We show you one verdict, and grade it.
Most tools show what Polymarket already shows for free — in a different layout. Some track whale trades. Some show charts. Some send alerts.
Few answer the question that matters — is this price wrong, and by how much? — and fewer still write the answer down before the market resolves and grade it after.
| Feature | Other tools | OddsShift |
|---|---|---|
| Raw odds | ✓ | ✓ |
| Whale alerts | some | ✓ |
| Wallet positions read on-chain | some | ✓ |
| Wallet tiers from the record on resolved markets, daily | ✗ | ✓ |
| Fair value from math models and AI models | ✗ | ✓ |
| One verdict per market: BUY YES / BUY NO / HOLD | ✗ | ✓ |
| Every BUY frozen before resolution, graded after | ✗ | ✓ |
| Public track record with the misses listed | ✗ | ✓ |
| Monthly scoreboards | ✗ | ✓ |
What we don’t know.
We could show only the wins. Every other tool does that. Here’s the truth:
Tracked wallets aren’t always right. A wallet’s record varies by category and by period, and even tier-1 wallets miss. That is why the tiers are recomputed every day and why the wallets are an input, not the call.
The models disagree sometimes. When the AI models split, or split from the wallets and the math, that is not a strong call — it is genuine uncertainty. It usually ends in HOLD. No false confidence.
We don’t know the future. A verdict is what the tracked wallets, the math and the models say now. The record shows the calls have matched more often than not; it also shows the ROI, which is a harder test, and the misses.
We’re transparent about bugs and limits. The track record is frozen before resolution and published with every miss. If we find a data issue, we disclose it.
Why say this? Because trust built on honesty lasts. If you want guaranteed profits, we’re not it. If you want to see where measured smart-money wallets are positioned, what the models make of it, and how that has graded so far — that’s us.
Frequently asked questions
Are tracked wallets always right?
What does HOLD mean?
How is the verdict graded?
How often is the data refreshed?
Do the AI models ever disagree?
Does OddsShift execute trades?
What is PRO?
See today’s verdicts.
PRO from $29/mo: full access to every verdict, its entry and target, and the tracked wallets behind it. The track record is free either way — check it first.