On-Chain Flows: Following the Money

On-chain flows track the movement of tokens between addresses that can be identified as belonging to a particular kind of holder. The most common version measures net transfers into and out of exchange-associated wallets, on the theory that deposits precede selling and withdrawals precede holding.

That theory is reasonable and incomplete. On-chain flows describe custody changes, not intent, and the gap between the two is where most misinterpretations begin. A deposit can be a sale, a collateral top-up, a custody migration or an internal reshuffle between two wallets under the same control.

• Flows measure custody, not sentiment. Moving coins to an exchange changes who holds them, not necessarily who wants to sell them.
• Labeling is a heuristic. Exchange attribution comes from clustering and public tags, and it is wrong often enough to matter.
• Internal transfers dominate. Much of the apparent movement is housekeeping between wallets owned by the same entity.
• Scale matters more than sign. A net inflow that is small relative to daily volume carries no information at all.

What on-chain flows actually measure

A flow metric measures the change in balance of a defined set of addresses over a defined window. Everything else — the interpretation, the signal, the narrative — is an inference layered on top of that accounting.

Inflows, outflows and netflow

An exchange inflow is a transfer from a non-exchange address to an address attributed to an exchange. An outflow is the reverse. Netflow is inflows minus outflows over the window, and it is usually expressed in tokens or in notional value.

The sign of netflow is less informative than its persistence. A single day of large inflows can be one custodian reorganising wallets; two weeks of steady inflows usually reflects something real about how holders are positioning.

Labeled and unlabeled entities

Attribution depends on labels, and labels come from three sources: addresses published by the entity itself, clustering heuristics based on shared spending behaviour, and manual tagging by analytics providers. Each source has a different error profile.

Self-published addresses are reliable but incomplete. Clustering is broad but produces false positives, especially with custodial services that batch withdrawals from pooled wallets. Manual tagging is accurate where it exists and silent everywhere else.

A deposit to an exchange is a statement about where coins are stored. It is not a statement about what the owner intends to do with them.

Why on-chain flows are noisy

Blockchain data is precise about transfers and vague about meaning. That combination produces metrics that look rigorous and behave unpredictably.

Change and consolidation. In account-based chains, an entity moving funds between its own wallets generates flows that look identical to a customer deposit. Custodians consolidate constantly, and consolidation days produce spikes that carry no signal.

Derivatives and lending. Coins sent to a venue may be posted as collateral rather than sold. In that case the flow is a leverage signal, and reading it as selling pressure inverts the conclusion.

Custody migration. When a large holder changes custodians, or when a fund rotates between qualified custodians, the transfer appears as a withdrawal from one labeled address and a deposit elsewhere. The net across all labeled addresses can be zero while each individual series looks dramatic.

Off-chain settlement. Increasingly, large positions change hands on venues without touching the chain. Flows then measure only the subset of activity that happens to settle on-chain, which is a shrinking and unrepresentative sample.

Reading on-chain flows without overreaching

The useful question is not “are coins moving to exchanges” but “whose coins are moving, and what did that cohort do last time this happened”.

The entity problem

Aggregate netflow treats a market maker, a long-term holder and a retail participant as one population. Splitting by cohort — wallet age, historical holding period, size band — usually produces a cleaner signal than the aggregate, because different cohorts behave differently.

The cost is sample size. Once you slice by cohort and by venue and by week, the number of observations per cell falls quickly, and apparent patterns become indistinguishable from noise.

Time horizon

Flow signals are horizon-dependent. Daily netflow is dominated by operations; weekly and monthly netflow begin to reflect allocation decisions. A strategy that reads daily flows and holds for a month is measuring one thing and trading another.

It also pays to compare flows with a denominator. Net inflow of a few thousand tokens means nothing without knowing daily traded volume, free float and the size of the venue’s reserves.

Building a defensible flow metric

A flow metric earns trust by being explicit about its assumptions, not by being complex.

1. Fix the address universe and record exactly which labels are included, so the series can be reconstructed later.
2. Decide how to treat internal transfers and exclude movements between addresses attributed to the same entity.
3. Normalize by a denominator — daily volume, free float, or total venue reserves — so the number is comparable across time.
4. Aggregate to a horizon that matches your holding period rather than the horizon that updates fastest.
5. Split the series by cohort and check whether the aggregate is hiding offsetting behaviour.
6. Test the metric against forward returns on data that predates any published version of it.
7. Re-run the test after every labeling update, because retroactive relabeling changes history.

A worked illustration of the scale problem: suppose an asset trades 40 million units a day and a venue holds 300 million units. A net inflow of 2 million units is 5 percent of daily volume and 0.67 percent of reserves. That is a rounding error in operational terms; calling it accumulation requires a much longer window or a much larger number.

Where flow analysis fails

Flow analysis fails most often when it is treated as a leading indicator of price rather than as a description of holder behaviour.

During trending markets, flows follow price rather than lead it. Coins move to exchanges after a rally, not before one, because holders are taking profit. A model trained on that relationship will systematically expect continuation where the data is actually showing distribution.

During structural changes, the mapping breaks entirely. A new custody product, a change in venue practice, or a migration between chains can shift the baseline level of flows without any change in behaviour, and any threshold calibrated on the old baseline will fire constantly.

The honest framing is conditional: flows tell you that a cohort moved coins, and the historical base rate tells you what that cohort did next. Both parts are estimates, and neither is a forecast.