A public blockchain publishes every transfer, every balance and every contract call, permanently and for free. No other market has anything comparable — the equivalent in equities would be a live tape of every account's holdings. This is why on-chain analysis exists and why its output should be read with a specific kind of care: the raw data is exact, and almost every metric built on it is an estimate.
The first layer of inference is entity clustering. Addresses are not identities, and one person or firm may control thousands. Analysts group them using heuristics — most commonly, that inputs jointly spent in one transaction share an owner, and that change outputs can be identified by their form. These heuristics are usually right and are systematically wrong in the presence of collaborative spending, exchange batching and privacy tooling. Every figure about how many holders exist, or what long-term holders are doing, is built on this layer.
The second is labelling. Calling an address an exchange wallet, a miner or an institutional custodian is an attribution, made from observed behaviour and from off-chain information of varying quality. Labels are how exchange inflows and outflows are calculated, which are among the most quoted metrics in the market. They are also where the largest errors occur: a custodian reorganising its internal wallets can produce an apparent outflow of enormous size that means nothing, and this has happened repeatedly.
The third is behavioural interpretation. Metrics such as realised capitalisation, coin-days destroyed, and the various profit-and-loss ratios assign an economic meaning to the movement of coins — treating the price at the time an output was created as a cost basis, and its movement as a realisation. The arithmetic is sound; the assumption that moving a coin means selling it is not. Transfers between an owner's own wallets, into custody, into a lending contract or across a bridge all appear as movement. In a market where a growing share of coins sits in exchange-traded products and institutional custody, the share of movement that is not selling has been rising.
Two structural blind spots limit everything above. Activity on centralised venues is invisible: the overwhelming majority of trading is internal ledger updates that never touch a chain, so on-chain data misses the price formation entirely. And derivatives, which set the marginal price in most conditions, are almost entirely off-chain. On-chain analysis sees settlement, not markets.
Used properly, it remains one of the few genuine informational advantages available to anyone. The discipline is to ask, of any metric, which heuristic produced it, what would make that heuristic wrong, and whether a single large actor could have produced the reading by doing something administrative. Metrics that survive those questions — supply held at a loss, coin age distributions over long windows, the direction of net flows over months rather than days — carry real information. Single-day readings quoted without their construction usually do not.