Overview
Every Ensemble index is a versioned rulebook with five parts: eligibility, selection, weighting, valuation and rebalancing. The rulebook is code, so the backtest, the live basket and this page can never disagree. An index changes only through a new version, announced at least seven days before it takes effect.
Indexes come in two types. Theme indexes track a family of related outcomes, such as the path of Fed policy. Hedge indexes are sized against a target notional, so a basket pays a set amount if the hedged event happens. Strategy sleeves, which weight by expected edge, are a third type published only after they clear their costs in testing.
Eligibility
Every market on each connected venue is screened hourly. A market qualifies only if:
- The venue is available in the account holder's state.
- Its category and wording pass the exclusion list: no war, terrorism or assassination markets, no mention markets, and no sports at launch.
- It resolves inside the index's window, for example between 1 and 120 days.
- Its price sits inside the index's band, by default 3 to 97 cents.
- Its order book shows enough depth near the mid and a spread no wider than the rulebook allows.
- Its resolution source is named and public.
A held leg that later fails eligibility is not sold immediately. It is flagged and sold at the next rebalance, unless the failure is a venue halt or a state restriction, which freeze the leg at once.
Selection and grouping
Eligible markets are ranked by the index's own score: theme relevance, hedge fit or expected edge after costs. Markets that hinge on the same underlying outcome share a correlation group, and each group is capped, by default at 20% of the index. Mutually exclusive outcome sets, such as a "who wins" market with several candidates, are treated as one group, and the sum of their prices is checked before any of them is held.
Each index sets a leg count, typically 5 to 25. Ties in the score are broken by tighter spread, then deeper book, then earlier resolution.
Weighting
A binary contract bought at price p settles at 1 or 0. Per dollar invested, the variance of its return is:
var(YES at p) = (1 − p) / pvar(NO at p) = p / (1 − p)
Three weighting schemes are supported:
- Equal weight: the same dollars in every leg. Simple, but a 5-cent contract carries far more risk than a 90-cent one.
- Risk parity (default): each leg weighted by the inverse of its volatility, then group caps applied and weights renormalized.
- Edge weighted: for strategy sleeves only, weight by expected edge after fees, capped by risk.
w_i = (1 / σ_i) / Σ_j (1 / σ_j)edge_i = q_i − p_i − fee(p_i)
Here q is a calibrated probability from historical resolution rates of similar contracts, and fee(p) is the venue's published fee at price p. Exchange fees follow a p × (1 − p) curve, so as a share of money spent they are highest on cheap contracts. A strategy sleeve is published only if its edge stays positive after fees and spreads in testing.
Worked example
Three YES contracts priced at 10, 50 and 80 cents. Risk parity puts the most money in the steadiest contract:
| Contract price | Volatility σ | 1 / σ | Risk parity weight | Equal weight |
|---|---|---|---|---|
| $0.10 | 3.00 | 0.33 | 10% | 33% |
| $0.50 | 1.00 | 1.00 | 30% | 33% |
| $0.80 | 0.50 | 2.00 | 60% | 33% |
Index level
Each index starts at 100 on its launch date. Between rebalances the level moves with the mark value of the basket; at each rebalance the level carries over and a new base is set, so the level is chained across rebalances:
level(t) = level(r) × mark(t) / mark(r)
Here r is the most recent rebalance. Fees and spreads paid at a rebalance reduce the basket's mark value, so trading costs show up in the level rather than being hidden. An exit level, computed the same way from exit value, is published beside it.
Reference index calculation
The five reference indexes are paper portfolios calculated by the open engine every 15 minutes from public Kalshi and Polymarket US quotes. Each one started at 100 on November 2, 2026. The steps are the same in the published levels and in backtests:
- Launch: the index buys its first basket at the ask, pays each venue's taker fee, and is then rebased to 100, so the launch value is the basket at mid. Every later trade pays its full cost.
- Reconstitution: the rulebook picks the constituents again on its schedule, after any constituent settles, and when a constituent comes within the rulebook's minimum time to close. That last case is a roll: Fed Path leaves each meeting a day before the decision.
- Reweighting: between reconstitutions the constituents are fixed. If a weight drifts past its trigger, or idle cash passes 10% of the level, the index trades back to target weights in the same contracts. It never adds new contracts then, which keeps turnover and costs down.
- Trades: buys pay the ask and sells receive the bid, each plus the venue's taker fee. Trades smaller than 0.25% of the level are skipped.
- Marks: the level marks each position at mid. A market that is closed but not yet settled keeps its last mark until it settles, then moves to cash at its payout.
- Exit level: the value if every position were sold at the bid after fees, published beside the level.
- Capacity: the dollars the basket could add, in proportion, without paying more than 2 cents above the best ask on any leg, read from each order book at the time of the calculation.
- Data quality: if any venue source an index depends on fails to load, that index skips the run and keeps its last level rather than publish a partial calculation.
Every trade, level and hourly order book snapshot is recorded. The calculation, the rulebooks and the tests that check the published code against the Python engine live in Ensemble's engine repository, which design partners can review and use to rerun the track record from the recorded books.
Valuation
Two values are shown for every index, because the gap between them is the real cost of leaving early:
mark value = cash + Σ quantity_i × mid_iexit value = cash + Σ quantity_i × bid_i − fees
When a book is one-sided or empty, the leg is marked at its last trade and flagged as stale. Legs whose outcome is disputed or delayed are carried at their last trade and flagged. Settled legs move to cash at their actual payout, which may be a fraction of a dollar if a venue settles that way.
Fees
Every quote, backtest and index level uses each exchange's published fee schedule, applied per contract at the order's price p:
| Venue | Taker fee | Maker fee | Taker fee at 40 cents |
|---|---|---|---|
| Kalshi | 0.07 × p × (1 − p), times the series multiplier, rounded up to the cent per order | 0.0175 × p × (1 − p) | 1.68 cents |
| Polymarket US | 0.0695 × p × (1 − p) | None | 1.67 cents |
Schedules change. Ensemble reads the current schedule before every quote and records the one used for each order and backtest. Sources: Kalshi fee schedule, Polymarket US fees.
Rebalancing and rolling
- Scheduled: each index has a cadence, daily, weekly or monthly, with a randomized start inside a window so the timing can't be front-run.
- Triggered: any leg drifting more than five points from its target weight, any leg failing eligibility, or settled cash above 10% of the index.
- Rolling: settlement cash buys the next eligible contracts under the same rules.
- Trade minimization: changes below a threshold, $25 per leg by default, are skipped, so fees don't eat small corrections.
Every rebalance is a proposal. It lists each order, its expected fee and the net cash change, and nothing trades until it is approved.
Execution
Orders are limit orders only. By default Ensemble rests orders at the touch as a maker, waits a set time, then crosses the spread only within the slippage cap. Large orders are sliced against visible depth. Every order carries a unique client ID, so a retry after a timeout can never double a position.
- Tick grid: prices snap to each market's tick size before an order is sent, since off-grid prices are rejected.
- Legging: if one leg cannot fill within its cap, the rest of the basket still fills and the gap is reported, rather than chasing the price.
- Kill switches: global, per index and per venue, plus Kalshi order groups that cancel resting orders when fills exceed a cap.
Backtesting
Backtests replay a rulebook version against recorded order books, not closing prices:
- Point in time: only markets listed on each day are eligible, using the wording and rules in force that day.
- Depth-limited fills: a taker order fills only against the size shown at each price level; a maker order fills only if the market trades through its price.
- Real costs: each venue's fee schedule at the time, plus the spread actually paid.
- No look-ahead: resolutions, disputes and voids enter the backtest on the date they were announced.
Reports show gross and net returns, fees, turnover, drawdown and capacity, the size at which the index's own trades would move prices. Every backtest names its rulebook version and data window and is labeled hypothetical.
Data and calculation
Market data comes directly from each venue: REST snapshots for markets and order books, and WebSocket streams for book changes, trades and market status. Any gap in a stream triggers a fresh snapshot before trading resumes. Self-hosted installations pull live data straight from the venues; Ensemble's cloud publishes rulebooks, reference index levels and the historical order books used in backtests.
Index levels are recalculated whenever a leg's quote changes. Reference levels shown on this site are simulated until live publication begins.
Settlement handling
| Event | How the index handles it |
|---|---|
| Leg settles before the next rebalance | Cash is held and rolled at the next rebalance |
| Market closes early because the event happened | Treated as a normal settlement; no fixed end date is assumed |
| Outcome disputed or delayed | Leg marked pending at last trade and excluded from rebalancing until final |
| Market voided or settled at a partial value | Payouts are stored as decimals, never as yes or no flags |
| Venue changes a market's rules | Leg re-screened daily and sold at the next rebalance if it no longer qualifies |
| Event postponed | Leg held with the new expected date; long delays trigger a sell rule |
| Trading halted in one market | That leg is frozen and flagged; the rest of the basket trades normally |
| Venue outage or data gap | No orders on that venue until a fresh snapshot matches the ledger |
| Same question on two venues resolves differently | Legs are tied to one venue and one market and never netted across venues |
Versioning and governance
Each rulebook carries a version number. Changes are announced at least seven days before they take effect, and the previous version stays published. Emergency changes, such as excluding a market after a venue halts it, can take effect at once and are published with the reason the same day. Backtests always state which version and which data window they used, and are labeled as hypothetical.
Glossary
| Term | Meaning |
|---|---|
| Leg | One contract position inside a basket |
| Mark value | What the basket is worth at mid prices |
| Exit value | What selling every leg at the bid would return after fees |
| Drift | How far a leg's current weight has moved from its target, in percentage points |
| Correlation group | Legs that depend on the same underlying outcome, capped together |
| Maker and taker | A maker order rests on the book; a taker order trades against one that is already there |
| Slippage cap | The most an order may pay above the mid before it stops crossing the spread |
| Hedge ratio | Contracts held per dollar of exposure being offset |
| Capacity | The size an index can trade before its own orders move prices materially |