Where the numbers on this site come from

Every figure on this site originates in one of four source layers: trading venues, blockchains themselves, application contracts, and published disclosures. Each is produced by different parties, updates on a different rhythm, and fails in a characteristic way. The methodology page states the current rules field by field; this lesson explains the shape of the pipeline so that those rules are readable.

The four layers

• Market data measures prices, volumes, depth and derivatives positioning. It is produced by exchanges and by aggregating vendors, and its characteristic errors are fabricated volume, stale quotes and venues that quietly stop reporting.
• On-chain data measures supply, fees, transactions, addresses and staking. It comes from full nodes and from indexers that transform raw blocks into queryable form, and its characteristic errors are definitional drift in labels such as active addresses, and provisional values near the chain tip that a reorganization can change.
• Application data measures value locked, protocol income and lending balances. It is assembled by reading contracts across many chains, and its characteristic errors are double counting through wrapped and restaked positions, and sensitivity to the price feeds used to value deposits.
• Disclosure data measures fund holdings, corporate treasuries and reserve reports. It comes from regulatory filings and issuer publications, and its characteristic errors are reporting lag and the limited scope of what any given statement actually covers.

The distinction that matters most is between measured and reported. On-chain figures are measured: anyone with a node can recompute them from the same history and reach the same answer. Market and disclosure figures are reported: they depend on a party choosing to publish, and on that party's own definitions. A number's reliability tracks that distinction far more closely than it tracks the reputation of whoever displays it.

What is observed and what is derived

Few fields on a data page are raw observations. Price is an aggregate, as the previous lesson described. Fees are measured in the native asset on the chain and then converted to dollars, so a dollar fee series mixes activity with the asset's own price move, and the conversion convention must be stated for the figure to be reproducible. Protocol income over thirty days depends on where the boundary is drawn between payments retained by liquidity providers or block producers and payments directed at the asset itself, a choice with no universal standard across the industry.

Annualization is another derivation, and a lossy one. A yield stated from a short window and scaled to a year assumes that window repeats, which it frequently does not. Real staking yield differs from the nominal figure precisely because it subtracts the dilution caused by new issuance, and a nominal number alone overstates what a staker's proportional position gains. Ratios inherit the error in both inputs, so a ratio between two derived series carries two layers of convention before any interpretation begins.

As-of, backfill and revision

Every series has a timestamp convention, and that convention is a substantive part of the number rather than a footnote. Snapshots taken at a fixed hour differ from rolling windows, and values near the chain tip stay provisional until enough blocks have accumulated that a reversal is implausible.

Two processes change historical values legitimately. Backfill fills gaps once a source that was unavailable returns, or once a new chain's history is indexed for the first time. Revision corrects a definition or a bug and restates the past on the new basis. Both are normal in serious data work, and both mean a chart today is not necessarily identical to a screenshot taken last month. What matters is that such changes are disclosed rather than silent.

Two biases flatter almost every dataset in this field and deserve naming. Look-ahead bias appears when a value known only later is placed at an earlier timestamp, making historical analysis look sharper than the information available at the time allowed. Survivorship bias appears when assets and protocols that failed drop out of the universe, so an aggregate computed over today's list quietly describes a group selected for having survived. Both are structural rather than careless, and the correction is to keep failed assets in the record, which is why the incidents pages sit alongside the asset pages.

Where trust is unavoidable

Some figures cannot be verified independently from a chain, and saying so plainly is more useful than implying precision. Reserve figures behind a fiat-backed stablecoin rest on a reserve attestation, which is a point-in-time report on assets under an agreed scope rather than a continuous audit of both assets and liabilities. Institutional holdings disclosed in quarterly filings describe a past quarter-end and cover only certain instrument types. A fund's net asset value is struck once per day by the issuer. Any on-chain system that prices external assets depends on an oracle, and every figure derived from such a protocol inherits that dependency, including value-locked figures that look purely mechanical.

Coverage, gaps and why a blank is better

The set of assets shown is itself a methodological choice. Inclusion rules based on liquidity, venue availability and data quality determine which assets appear, and any aggregate computed across that set describes the set as much as the market. When a chain is added, its history enters through backfill and aggregate series can shift, which is why coverage changes belong in a change log rather than in a silent update.

Where a figure is unavailable or not meaningful for an asset's category, the field is left empty rather than filled with a substitute. A plausible-looking placeholder is worse than a gap, because it propagates into every ratio built on top of it and cannot be distinguished from a real measurement afterward.

Each field has an entry in the metric catalog giving its definition, unit, source layer and update frequency, and every ratio entry names both inputs, since most disagreements between sites resolve into a different denominator rather than a different measurement. The complete list of vendors, venues and node infrastructure is on the data sources page. The next lesson applies all of this to the single most quoted figure in the field.