The Polymarket Up/Down Observatory

What this is

UpDown Charts records the Polymarket order book and the reference price feeds listed in the methodology for all 28 up/down markets (7 assets × 4 window lengths), sampled every 500 ms. This page measures one thing: how well the book's implied odds predicted the result, using the last two-sided quote at 60, 30 and 10 seconds before each window closed. Every current figure here comes from scripts/observatory/build-aggregates.mjs and is reproducible from the published aggregates (the earlier-build figures in the revision note above excepted).

Methodology and data notes · What is a Polymarket up or down market?

1. Base rates

Over 13,783 windows the UP rate is 50.63% [49.80%, 51.47%]. Pooled across the seven assets, 5m at 48.23% [46.58%, 49.89%] (n = 3,500), 1h at 52.03% [50.37%, 53.68%] (n = 3,500) and 4h at 51.81% [50.10%, 53.52%] (n = 3,283) have intervals that exclude 50%, but that pooling assumes independent windows and the seven assets share the same clock, so the effective sample is well below each n and the intervals are too narrow. Taken one market at a time, 20 of those 21 per-market intervals contain 50%. The same caveat applies to the pooled headline above. This is a record of what happened, not an edge: an outcome frequency says nothing about profit once the bid/ask spread and fees are paid.

All seven assets sit between 49.97% and 51.65% with overlapping intervals, so the data does not distinguish one asset from another on direction.

2. Calibration

A calibrated book is one where things priced at 20% happen about 20% of the time. Grouping windows into deciles of the implied UP probability (the mid of the UP token top-of-book), the observed rate never falls from one bucket to the next across the ten buckets at 60 seconds before close, over 13,081 windows. The two buckets straddling 0.5 are the thinnest cells (n = 181 and n = 180) and their intervals span about 14 points each, so the middle of the curve is the least precise part of it. Across the three checkpoints 7 of the 30 decile intervals exclude their own mean implied price, more than the one or two that 95% intervals would miss by chance alone. Away from the tails: [0.4, 0.5) at 60s before close, priced at a mean 0.450 and resolved UP 34.25% [27.73%, 41.43%] (n = 181) and [0.1, 0.2) at 30s before close, priced at a mean 0.138 and resolved UP 6.01% [3.39%, 10.44%] (n = 183).

The tails are the best-powered cells in the calibration table. At 60 seconds the bottom bucket was priced at a mean 0.014 and resolved UP 1.03% of the time ([0.79%, 1.33%], n = 5,447); the top bucket was priced at a mean 0.987 and resolved UP 99.47% ([99.24%, 99.63%], n = 5,626). Of the six tail intervals across the three checkpoints, 5 exclude their own mean implied price, all the same way: heavy favourites won slightly more often than priced and longshots slightly less. It is still a record and not an edge: these are frequencies at a quoted mid, before the spread and fees a position would pay, and calibration is a statement about the pricing rather than about whether a strategy would have made money.

3. Favourite watch

When the book priced a side at 90% or higher with 60 seconds left, that side lost 0.78% of the time: 87 of 11,103 windows, Wilson 95% [0.64%, 0.97%].

These are outcome frequencies at a quoted mid, before the bid/ask spread, fees and slippage any position would pay. 90% is only the entry threshold: sides in this set were quoted at a mean of 98.6c, not 90c.

The rate falls as the deadline approaches, while the eligible set grows by 15% as prices harden: 0.78% at 60s before close (n = 11,103), 0.36% at 30s before close (n = 12,229), 0.27% at 10s before close (n = 12,737).

At 60 seconds the rate runs from 0.27% on 4h (8/2,980) to 1.24% on 5m (31/2,509). Those two intervals do not overlap, though both pool seven assets on one clock and so run narrow, and they cover different calendar periods (see the base-rate caption). The comparison is between two outcome frequencies, not between two returns: both sit before spread, fees and slippage, and neither says a position taken at either price would have paid.

4. The hourly gap

Two numbers look like the same price and are not. Our own first Chainlink reading after a window opens sits 3.2 bps away from the open Polymarket actually settles against on 1h windows (median over 1,204 windows, p95 10.2 bps, 31.48% of windows above 5 bps). On 5m, 15m and 4h windows the same comparison is 0.24, 0.22 and 0.1 bps. It is mostly one-sided rather than jitter: on 1h windows the Chainlink reading is above the settlement open in 6.7% of windows, signed median -3.16 bps, and every asset with a Chainlink feed across the whole sample shows the offset, at medians of 3.03 to 3.39 bps.

It matters because a reference level taken at the top of the hour is not the level the 1h market resolves against, so a window can look decided on one number and settle on the other. Site tables print the settlement open for that reason; the methodology page documents both numbers. Its 9.24 to 9.78 bps range is not from this build: it was measured on an earlier, smaller sample using a 2-decimal stored copy of the same value, and is marked there as superseded. The offset also moves over time (this page's 2026-07-26 build measured 7.67 bps the same way), so cite the figure above with its sample: the most recent 200 1h windows per market as of 2026-09-24.

5. Window replays

3 windows where a side above 93% lost, one per asset and window length, 240 recorded points each. The green line is our own exchange recording (Binance, falling back to Pyth or Chainlink where Binance has no feed). The two dashed lines are the settlement open and the close stored for the window, both from Polymarket's price-to-beat feed sampled at the window boundaries, so the green line does not have to end on the close. The outcome is Polymarket's settled result, not a recomputation from those two lines: since 2026-08-07, 5m, 15m and 4h windows settle on a Chainlink TWAP they do not reproduce. Replays are limited to windows whose stored close, read against the settlement open, gives the settled outcome. XRP and DOGE are left out of replay selection, a rule kept from when their window closes were stored at 2 decimals (closes before 2026-07-26 still are).

Browsing every window that matches a filter, going back through the full retained history, and exporting the rows as CSV are dashboard features on Pro. See pricing or open the dashboard.

Premium adds programmatic access to the same per-window data through the REST API and the hosted MCP server: API and MCP reference.

6. Data and citation

Every aggregate on this page is published as a single JSON file: /observatory-aggregates.json. It is free to use with attribution and a link back to this page, under CC BY 4.0. It contains the aggregates on this page plus the 3 window replays shown above, each downsampled to 240 points. The rest of the per-window dataset and the recorded price archive are not part of it.

Limitations and coverage

• Outcome frequencies are a record of what happened. They exclude the bid/ask spread, fees and slippage, so they are not a statement about profitability.
• A checkpoint is excluded for a window when the last recorded two-sided quote is more than 20s old, when the book is one-sided, or when the spread exceeds 0.2. Exclusion rates are reported per checkpoint and per market.
• Window close prices recorded before 2026-07-26 were stored at 2 decimals, which cannot represent a normal short-window move for XRP and especially DOGE; closes since then are full precision. Outcomes are the result Polymarket settled (history reconciled against it on 2026-09-24), not recomputed from these prices.
• Pyth stopped serving on 2026-08-26 16:20 UTC and the recorder stored its last price until 2026-09-24 15:00 UTC. Those Pyth values are ignored: replay price paths in that range use Binance, else Chainlink, and HYPE windows from that range (Pyth was its only feed) are not replay candidates.
• HYPE had no Chainlink or Binance feed before 2026-09-24; its windows before then have no captured open and are absent from the captured-vs-settlement comparison.
• The sample is the most recent resolved windows per market at build time, at most 500 (fewer where fewer have resolved), not the full archive. Recorded price series are retained for 90 days.
• Windows for different assets cover the same clock intervals, and concurrent crypto moves are correlated, so any cell that pools assets has an effective sample well below its N. Wilson intervals on pooled cells assume independent windows and are therefore optimistic; per-market cells are the ones to read.
• Because the sample is up to the most recent 500 windows per market, the shorter timeframes cover far less calendar time than the longer ones. Per-timeframe date ranges are in coverage.byMarket and are published next to the base-rate table.
• Quotes are sampled by our own collector every 500 ms. A checkpoint mid is the last observation at or before that instant, not the exact book state at that instant.
• Quote coverage at 60 seconds before close is at least 96% on 25 of 28 markets. The exceptions are hype-1h at 15.2%, hype-4h at 79.1% and doge-1h at 90.2%. A checkpoint counts only when a two-sided quote no wider than 0.2 was recorded in the 20 seconds before it, so any claim about those markets should be scoped to the windows that qualified.
• The sample is capped at 500 windows per market: 500 per 5m market, 500 per 15m market, 500 per 1h market and 469 per 4h market.