Methodology v1.0
The Scopes Desk
1Purpose and scope
The Scopes Calibration Index (SCI) measures how well prediction-market prices anticipated the outcomes of the contracts they priced. It is published as a family of series, by venue, by desk and by horizon, on a fixed schedule, computed from the timestamped, append-only record maintained by Scopes Reference Data, LLC (the Administrator).
The Scopes Fair Value is not scored by the SCI and is never an input to it. The Administrator does not benchmark its own estimates within a benchmark it administers.
The SCI is a reference statistic. It is not investment advice and is not designed to be tracked by a financial product; product use requires a separate license. The SCI is not a regulated benchmark. This document follows the structure of the IOSCO Principles for Financial Benchmarks (2013), and the Administrator publishes an annual self-assessment against them, without any claim of regulatory compliance.
2The family
One series per live desk, scored on the venues the record covers for that desk, plus the all-desks composite (SCI-ALL).
| Series | Desk | Venues scored |
|---|---|---|
| SCI-D-RATES | Rates | Kalshi, Polymarket |
| SCI-D-POLLS | Polls | Kalshi, Polymarket |
| SCI-D-SPORTS | Sports | Kalshi |
| SCI-D-ENERGY | Energy | Kalshi |
| SCI-D-STOCKS | Stocks | Kalshi |
| SCI-ALL | Composite | All venues |
Each venue is also reported separately within a desk. Excluded venues: Gemini, whose market data agreement bars use in any benchmark or index; and Polymarket US, which enters a desk only after it clears the record's evidence gate for that series. A venue added later enters as a methodology change (section 14), never retroactively.
Because each venue's value covers the events that venue listed, two venues on one desk are also compared on the events both listed. For each pair, each venue is scored on the shared events only, under the same gate; the difference (first venue minus second) carries a 90% interval from resampling the shared events, keeping each drawn event's contracts on both venues. This venue comparison is a diagnostic published with the values, not a series in the family.
3Definitions
- Contract: a binary event contract on one venue with a defined resolution rule.
- Outcome: 1 if the outcome resolved yes, 0 if no. On the Rates and Polls desks it is the result recorded in the Scopes record, and one recorded result per outcome applies to every venue that priced it. On the Sports, Stocks and Energy desks, whose record stores prices by Kalshi market rather than an outcome per contract, it is that market's settlement as published by Kalshi, retrieved at computation and kept with the observations. (Corrected October 31, 2026.)
- Anchor: the scheduled event start where the record has one (a game's start time), otherwise the event's scheduled resolution time.
- Horizon: a fixed interval before the anchor at which the price is sampled. The horizon timestamp is the anchor minus the horizon.
- Observation: one (contract, venue, horizon): the sampled price p and the outcome o.
- Record date: when the Administrator captured a price (the known-when time). Only prices captured at or before a horizon timestamp can be sampled for it.
- Cell: one (venue, desk, horizon, period). Every statistic is computed per cell.
4Horizons
| Desk family | Desks | Horizons |
|---|---|---|
| Long-dated | Rates, Polls | H90, H30, H7, H1 (days) |
| Short-dated | Sports, Stocks | H3, H1 (days), H6h (6 hours) |
| Short-dated, energy | Energy | H3, H1 (days) |
Energy is scored on short-dated horizons because its contracts include daily settlements. It has no H6h horizon because that would fall inside its 24-hour near-expiry window (section 5). A horizon that falls before the contract was listed, or inside the event, is not sampled for that contract.
5Sampling and eligibility
For each resolved outcome, each venue that priced it, and each horizon of its desk:
- Take the last captured price at or before the horizon timestamp. A later price is never used.
- The price is the midpoint of the best bid and ask where both exist, otherwise the last trade, flagged last_trade. A bid exists only if it is above 0 and an ask only if it is below 100%, because the venue records an empty side as 0 or 100% (stated November 1, 2026).
- If that price was captured more than 1 hour (short-dated) or 6 hours (long-dated) before the horizon timestamp, it is kept and flagged proxied. If no price was captured within 3 days (short-dated) or 14 days (long-dated), the observation is excluded as no_price. No price is ever estimated or filled by judgment.
- Each outcome of a multi-outcome event is its own binary observation.
An observation is excluded, with the reason counted and published, if it fails any rule:
| Reason | Rule |
|---|---|
| unresolved | No recorded outcome yet. |
| voided | The event was cancelled or voided. |
| no_price | As in step 3. |
| listing_age | The contract was first captured less than 3 days (short-dated: 24 hours) before the horizon timestamp. |
| illiquid | The sampled price had no two-sided quote (a bid above 0 and an ask below 100%) and reported zero or no 24-hour volume. |
| not_comparison_valid | Captured in play, or inside the near-expiry window (stocks 2 hours, energy 24 hours before settlement). A sports contract whose recorded start time falls on a different US Eastern date than the game date in its own market ticker cannot be shown to be pregame and is excluded under this reason (added October 31, 2026). |
The liquidity rule uses only what the record captures. Trade counts and multi-day notional are not recorded, so a trade-count floor is not used.
The record's routine capture runs several times a day, with minute-level capture only near resolution, so many short-dated observations will be proxied. The proxied share is published with every value.
6Statistics
For a cell with N eligible observations (p_i, o_i):
ō = (1/N) Σ o_i
UNC = ō (1 − ō)
SCI value (Brier skill score) = 1 − BS / UNC
The headline SCI value is the Brier skill score against the cell's own base rate: how much better the prices did than a constant price equal to the share of outcomes that resolved yes. Higher is better; 0 means no better than knowing the base rate; it can be negative. If every outcome in the cell is the same, UNC is 0 and the value is published as undefined, never as 0 or 1. Accuracy (BS, lower is better) is published beside it.
Calibration. Observations are sorted by price (ties by observation id) into K = 10 equal-count bins: the observation at rank r goes to bin floor((r − 1) × K / N) + 1. For bin k with n_k observations, mean price p̄_k and observed frequency ō_k:
RES = (1/N) Σ_k n_k (ō_k − ō)²
ECE = (1/N) Σ_k n_k |p̄_k − ō_k|
WBV = (1/N) Σ_k Σ_(i in k) (p_i − p̄_k)²
WBC = (2/N) Σ_k Σ_(i in k) (p_i − p̄_k)(o_i − ō_k)
BS = REL − RES + UNC + WBV − WBC (exact)
REL is the calibration line (lower is better), with ECE as its plain-language companion. The familiar three-term decomposition, BS = REL − RES + UNC, is exact only when every price in a bin is identical; with binned prices the two within-bin terms complete it. Every published cell reconciles to its BS through this identity.
Gates. A cell is published only when N is at least 30 and its observations come from at least 10 distinct events (a release day, a game, an election); below either, it shows "insufficient data" with its N and event count, never an estimate. Many strikes on a few release days are not independent evidence, and a bootstrap over fewer than 10 events cannot give a reliable interval. (Event minimum added October 31, 2026.) The reliability diagram is drawn only when N is at least 100; below that, REL and ECE are published as numbers.
7Uncertainty
Every published value carries a 90% interval from a cluster bootstrap: 1,000 resamples drawing events with replacement (all of an event's observations travel together, so the outcomes of one multi-outcome event are never treated as independent), with the SCI value, BS and REL recomputed on each. The interval is the 5th to 95th percentile. The random seed is fixed and published with each value, so the interval reproduces exactly.
8Composite
SCI-ALL is computed at H1, the one horizon every desk shares, so no contract is counted more than once. Each desk's H1 value pools that desk's eligible observations across its venues. The headline composite is the equal-weight mean of the desk values that clear the gate, so no single high-volume desk dominates; the desks included are listed with it. A contract-weighted mean (each desk weighted by its N) is published as a secondary line. The composite is published only when at least two desks clear the gate.
Desks enter and leave the composite as they clear or miss the gate, so each publication also reports why the trailing composite moved since the previous one. The change is split exactly, with nothing left over, between desks present both times and desks that entered or left. With C the desks present both times and m the mean of their values: a continuing desk contributes its own change divided by the size of C; a desk that entered contributes its value minus m (at the later publication) divided by the number of desks then; a desk that left contributes the negative of its value minus m (at the earlier publication) divided by the number of desks then. For the contract-weighted line, a continuing desk's contribution splits into a score part and a weight part (its share of the continuing desks' N changed), and an entering or leaving desk's is its N times its distance from the N-weighted mean of C, over the total N. The breakdown restates published values; it changes none.
9Periods and publication
An observation belongs to the calendar month (UTC) in which its event was scheduled to resolve, not the month it was priced, so a published value never contains an unresolved contract. Monthly values are published as points; the headline series is the trailing 12 months. Until July 2027 the trailing window starts at the beginning of the record, July 15, 2026, and is labeled that way.
Values publish monthly on the fifth business day for the prior month, from midnight Eastern time. A business day is Monday to Friday, excluding US federal holidays (a holiday falling on a Saturday is observed the Friday before, on a Sunday the Monday after). The first publication, on November 2, 2026, covers July 15 through October 31, 2026. Results of the November 3, 2026 elections enter the November values in the normal monthly publication.
10LIVE and BACKFILL
An observation is LIVE only if both hold: the day its price was captured was anchored the same day in the daily integrity digest (anchor mode same_day, see /integrity), and its outcome was recorded on or after this version's effective date. A value is LIVE only if every observation in it is LIVE; otherwise it is BACKFILL, shown in a separate table. A BACKFILL value is never relabeled LIVE.
The second condition exists because these rules were fixed after the outcomes in the first publication were known. Those values are computed mechanically under rules published before they were computed, but not before the outcomes, so they are all BACKFILL.
11Special cases
- Voided or cancelled events are excluded and counted.
- Disputed resolutions are held out until the venue's process concludes; the final result governs, and a change to a recorded outcome after publication is a correction (section 14).
- A contract whose terms change mid-life is treated as two contracts at the change date; observations before the change are excluded unless an Oversight Log ruling finds the change immaterial.
- The same event listed on several venues is scored independently on each venue.
12What is published with every value
The SCI value and its interval; BS; REL and ECE with their intervals; RES, UNC, WBV and WBC; N and the number of events; the gate result; exclusions by reason; the proxied and last-trade shares; the reliability diagram where N is at least 100; the LIVE or BACKFILL label; the methodology version; the code version; a hash of the input observations; the bootstrap seed; and the integrity digest date with its timestamp proof. Each publication also carries the composite change breakdown (section 8) and the venue comparison (section 2). Values are derived statistics: the observation-level venue prices they are computed from are not published or redistributed.
13Governance
The Administrator is solely responsible for the record, the application of this methodology and publication. No benchmark determination function is outsourced. Cloud hosting and database services are infrastructure under the Administrator's control; contractors act on its instructions and have no discretion over inputs or methodology.
Oversight is exercised through the Oversight Log, a dated, append-only record of every methodology decision, ruling, correction and complaint, and through one named external reviewer who reviews the log and a sample of rulings annually and signs a short statement. The reviewer will be named in the Oversight Log on appointment; until then, no external review has taken place.
Conflicts: the Administrator and its principal hold no positions in contracts covered by the index, and attest to this annually. The Administrator accepts no payment from venues for inclusion or treatment. Revenue is disclosed by category in the annual statement, each January 31.
Controls: prices are captured automatically into an append-only record; no captured value is altered in place. Each day's record is hashed and the hash committed to a public timestamp service. Access to the record is limited to the Administrator and designated systems, and all writes are logged. Calculation code is versioned. Records, code, publications, proofs and the Oversight Log are retained for at least 7 years.
If publication ceases, the Administrator gives 90 days' notice, publishes the full history and record hashes in an open format, and offers the methodology and record to a successor at no cost.
14Changes, corrections and complaints
Changes. A change ships as a new version (v1.1 for a non-material change, v2.0 for a material one) with its rationale in the Oversight Log. A material change is published for consultation at least 30 days before it takes effect, and the prior version continues to be published in parallel for 12 months. A published value is never re-cut under a later version.
Corrections. An error found after publication is corrected within 5 business days. Within 12 months of the original publication the corrected value is published with the original struck through; beyond 12 months the original stands with a correction note. Every correction is logged.
Complaints about a published value or a ruling go to hello@scopes.com and are answered in writing within 15 business days; each is summarized, anonymized, in the Oversight Log.
15Worked example
A test cell any implementation must reproduce to 6 decimals: N = 12 across four events, with K = 4 bins for the example (production uses K = 10).
| Obs | Event | p | o |
|---|---|---|---|
| 1 | E1 | 0.05 | 0 |
| 2 | E1 | 0.20 | 0 |
| 3 | E1 | 0.70 | 1 |
| 4 | E2 | 0.10 | 0 |
| 5 | E2 | 0.35 | 1 |
| 6 | E2 | 0.60 | 0 |
| 7 | E3 | 0.45 | 0 |
| 8 | E3 | 0.80 | 1 |
| 9 | E3 | 0.90 | 1 |
| 10 | E4 | 0.15 | 0 |
| 11 | E4 | 0.55 | 1 |
| 12 | E4 | 0.95 | 1 |
| Bin | Observations | p̄_k | ō_k |
|---|---|---|---|
| 1 | 1, 4, 10 | 0.100000 | 0.000000 |
| 2 | 2, 5, 7 | 0.333333 | 0.333333 |
| 3 | 11, 6, 3 | 0.616667 | 0.666667 |
| 4 | 8, 9, 12 | 0.883333 | 1.000000 |
REL = 0.006528 · RES = 0.138889 · ECE = 0.066667
WBV = 0.005000 · WBC = 0.005556 · REL − RES + UNC + WBV − WBC = 0.117083
Composite check: desk values 0.531667 (N 12) and 0.200000 (N 40) give an equal-weight composite of 0.365833 and a contract-weighted composite of 0.276538.
16Version history
| Version | Date | Change |
|---|---|---|
| v0.9 | October 1, 2026 | Consultation draft: governing principles only. |
| v1.0 | October 31, 2026 | Full specification. Effective November 2, 2026, with first values. |
| v1.0 correction | October 31, 2026 | Before taking effect: outcome source for Sports, Stocks and Energy (Kalshi settlement); minimum 10 events per cell; sports start-time check. |
| v1.0 correction | November 1, 2026 | Before taking effect: a two-sided quote requires a bid above 0 and an ask below 100%, as the venue records an empty side as 0 or 100%. |
The index · Oversight Log · Licensing and use
Published by Scopes Reference Data, LLC. Research and information only, not investment advice.