Cohort retention methodology
Use /retention to see the latest published report for Hyperliquid HyperCore perpetual markets. The card shows how much of a fixed group of trading addresses remains active, how much it pays in trading fees, and how much maker volume it executes in later months.
The month shown on the card is the end of the data window. Requesting a report does not calculate new data or subscribe the chat to notifications. Updates use complete UTC months; source delays can leave the previous report available for longer.
Reading the heatmaps
Each row is a separate baseline month, called M0. The cohort contains addresses with positive executed perpetual notional in that month. Its membership stays fixed: addresses arriving in later months are excluded from that row. An address can belong to several monthly cohorts.
Columns M1 through M6 show what that same cohort did one to six months later. Blank cells have not reached that age yet; they are not zero retention. The headline is the median M6 value across cohorts with six complete follow-up months, calculated separately for each metric.
These are rolling active-address cohorts, not cohorts of newly acquired users. One person can control several addresses, and an address is not necessarily one person or company.
The three main metrics
Let C be the fixed M0 address set and h the number of months after M0.
| Metric | Calculation | Interpretation |
|---|---|---|
| Address retention | 100 × addresses in C with positive executed notional at Mh / addresses in C | The share still trading; bounded by 100%. |
| Trading-fee NRR | 100 × positive trading fees from C at Mh / positive trading fees from C at M0 | Includes churn, contraction and expansion; can exceed 100%. |
| Maker-volume NRR | 100 × executed maker notional from C at Mh / executed maker notional from C at M0 | Activity of the same addresses when acting as makers; can exceed 100%. |
For the fee headline, each fill contributes max(fill fee - builder fee, 0). The subtraction removes separately reported builder fees. Negative fee amounts are not netted against positive payer fees; observed protocol rebates are tracked separately. Protocol and HIP-3 deployer fees both remain in the trading-fee headline.
For example, if the cohort paid $100 in M0 and pays $60 at M6, its trading-fee NRR is 60%, even if a smaller number of addresses now contributes most of those fees. This is a gross-fee retention measure, not profit or net revenue after incentives.
Maker volume measures executed trades, not standing order-book depth, spreads or market-maker identity. Changes in maker-volume NRR alone do not prove that market-making firms left or explain why activity changed.
Payers, concentration and fee-weighted survival
The lower table adds diagnostics using the same positive trading fees excluding builder fees as the primary headline:
- Payer count retention: the percentage of M0 fee-paying addresses still paying positive fees at Mh.
- Fee-base survival: the percentage of M0 fees contributed by addresses that still pay at Mh. It weights survivors by their original fees and is bounded by 100%.
- Payer-only fee NRR: current fees from the original M0 payers divided by their baseline fees; excludes conversion of initially non-paying addresses.
- Fees from M0 non-payers: the share of current cohort fees coming from addresses that traded in M0 but paid no positive fees then.
- Top-10 share of M0 fees: how concentrated baseline fees were among the ten largest payers.
- Fee NRR ex top payer: fee NRR after excluding the single largest M0 payer from both periods.
Fee-base survival and fee NRR answer different questions: retaining a large share of the original fee-paying base does not mean those addresses still pay the same amounts. Diagnostic survival and expansion ratios relate within a single cohort; multiplying their independently calculated medians is not a valid cohort decomposition.
Market coverage and assumptions
The analysis combines the native perpetual market and HIP-3 namespaces in the selected period. Spot markets are excluded. Monthly aggregation keeps the same address together across these perpetual namespaces. USD-pegged quote collateral is valued at $1; this does not measure depegging risk.
The primary cohort includes liquidation activity and addresses with any positive baseline notional. Additional analyses check minimum-notional thresholds and the effect of excluding liquidation activity. Volume is measured from account-side fills; maker and taker activity are separate lenses and should not be added as if they were unique exchange trade volume.
Estimated protocol/deployer split
The protocol-fee figure is supplementary and explicitly estimated. Before 21 March 2026, the calculation infers a 50% deployer share of positive HIP-3 trading fees when the field is unavailable. From that date, an entire missing deployer_fee column fails validation. A row-level null in a file containing the field is treated as zero.
That last convention is an assumption about the archive encoding, not an official guarantee. Hyperliquid documents the economic fee split and zero-deployer cases, but those formulas do not establish the meaning of a null in a third-party Parquet file. The primary trading-fee NRR does not depend on how trading fees are divided between protocol and deployer.
Source checks and limitations
The source is Hydromancer's historical fills archive. Coverage is checked for each namespace and UTC date against the configured activity periods. Schema checks also verify required fee fields. Duplicate-key and essential-field checks on sampled files can detect problems, but do not prove every row in the archive is error-free.
For the August 2025–August 2026 reference report, the supplied validation records cover 1,979 selected source files, with zero post-cutoff HIP-3 files missing the deployer-fee column. A directed probe of 27 files contains 12,064,792 rows: 623,098 null deployer fees, no explicit zero values and 11,441,694 nonzero values. This probe documents the encoding observed; it does not prove that every null economically means zero.
Historical source data may arrive late or be corrected. File inventory checks detect added or removed objects; changes inside an existing object require a deliberate rebuild. Report integrity checks verify that its image and metadata belong together, not that the provider's data is complete or economically exact.
Gross fees exclude neither every referral payment nor every incentive or private arrangement. Fee tiers, market conditions and venue mix can change the ratio without a change in user loyalty. Overlapping cohorts are not independent observations, and the results do not establish ownership, wash trading or causality.
Sources
Return to the command guide or open the bot.