Does Global Liquidity lead Bitcoin, and by how long?

Why do published estimates of the Bitcoin liquidity lag disagree?

Because each one measures a different series. There is no standard definition of global liquidity, so a lead time is a property of the measure rather than of Bitcoin.

The figures in the table below were produced on at least four different things, all of them called liquidity. The Bank for International Settlements means credit to non-bank borrowers. Most crypto research means the sum of national broad money across a varying set of economies, converted into US dollars. A third group means US net liquidity, which is not global. A fourth means a proprietary composite, which is what CrossBorder Capital, Coinbase Institutional, Keyrock and BlockPhi each build in their own way.

Once that is said out loud the disagreement mostly dissolves. A study measuring broad money and a study measuring Treasury bill issuance are not contradicting each other about Bitcoin; they are describing two different transmission channels that happen to share a word. What survives as a real disagreement is narrower and more interesting: among the studies that do measure broad money, the estimates still range across a season.

The practical consequence is a rule BlockPhi applies to its own work. A lead time published without naming the series it was measured on cannot be checked, cannot be reproduced, and cannot be defended when someone measures a different series and gets a different answer.

What lead time does each publisher report?

Lead estimates cluster near ten to thirteen weeks on broad money measures, with shorter broad-money results near six to nine weeks, and run much longer on credit and issuance measures. Two of the most-cited sources state no lead at all.

Two properties of the table matter more than any single row. The first is that the figures taken on broad money cluster, while the widest outliers are measured on something else: a credit index, a custom composite, or Treasury bill issuance. The second is that two of the most widely cited sources in the topic, Sam Callahan's work for Lyn Alden and Fidelity Digital Assets' research, state no lead time at all. The figure that circulates most and the figures that document their method most fully do not come from the same places.

Michael Howell's estimate is the one closest to a reference point, because it is the most documented. Writing in December 2024 he put the impact of Global Liquidity on Bitcoin at a lead-time concentrated around 11 to 13 weeks, with the earliest statistically significant impact after 5 weeks and the latest after 16 weeks, measured on CrossBorder Capital's own index. Those are model estimates on that index and past results do not indicate future results.

The widely repeated 10-week figure has a different provenance. It comes from a Bitcoin Magazine Pro article of March 2025 which refines it to 56 to 60 days, and which states no correlation coefficient, no sample period and no optimisation method. The same figure appears as a caption on two undated chart pages. Those are model estimates whose method is not published, so nothing on this page treats them as evidence for a lead time, and past results do not indicate future results.

How does BlockPhi measure the lead on its own series?

By scanning every lag across a fixed range on rate-of-change data, then testing whether the lag that wins the scan survives a model that weighs all of them at once.

The first step is a cross-correlation scan. At each lag from one week to twenty weeks it asks how strongly this week's change in the index lines up with Bitcoin's return that many weeks later, and it draws a conventional significance band around zero. The scan runs in both directions, so a peak at a negative lag would show Bitcoin moving first.

A scan is an exploratory tool and it has one specific weakness: it tests each lag on its own, in isolation. Nothing in the procedure asks whether the winning lag would still win with every other lag competing against it, and nothing asks whether Bitcoin's own recent returns already account for the same movement. Across twenty candidate lags, something has to come out highest. That is why the scan is where this derivation begins and not where it ends.

The lag a scan reports is supposed to correspond to a mechanical chain, and the early links of that chain are not in dispute between BlockPhi's own sources: central bank reserves reprice bonds in weeks two to four, corporate credit markets respond in weeks four to seven, institutional reallocation toward higher-risk assets follows in weeks seven to nine. Where Bitcoin sits at the end of that chain is not published here. BlockPhi's own two accounts of the same chain state different terminal windows, and which of them is the published figure is a question for the person who ran the estimation rather than something a reader can settle from the documents. A chain published with an invented final link would be a worse disclosure than a chain published with three links and a stated gap.

The lag is a measured property of that chain and nothing beyond it. It describes how long the record took to move through those intermediation steps, on this series, over this sample. It is a statement about what has been measured, not about what happens next, and no instruction to a reader follows from it.

What does a vector autoregression add that the scan cannot?

It makes every lag compete against every other lag and against Bitcoin's own history at once, and it tests the direction of the relationship both ways round.

Three specifications are estimated, carrying eleven, twelve and thirteen lags. Running three rather than one is what shows whether the signal sits at a single lag or is spread across a window, which no single specification can answer. In each of them the liquidity terms have to earn their place against Bitcoin's own past returns and against every other lag in the system.

The first test asks whether the full lag history of the index jointly helps explain Bitcoin's returns. It rejects the null in all three specifications, and in the longest of the three it rejects at the conventional five per cent level. A test of this kind establishes statistical precedence and not cause, which is a distinction this page keeps rather than blurs.

The second test is the one that matters most on this page, and it is the one that almost no published lead figure reports. It runs the same test backwards, asking whether Bitcoin's own history helps explain the index. It does not reject in any specification, and it is not close to rejecting in any specification. That is what rules out the obvious objection to every result here, which is that Bitcoin is moving global funding conditions rather than following them. Without it, a lead is a correlation with a direction asserted by its author.

The third result comes out of the fitted model rather than from a calculation performed in advance. An orthogonalised impulse response traces what happens to Bitcoin over the following twenty weeks after a one-off shock to the index, and its shape is a property of the estimated dynamics: nothing tells it where to look. In all three specifications it peaks within a week of the lag the scan selected. That makes it a confirmation from a different direction rather than a restatement of the scan, which is what a scan on its own can never provide.

What model is the lead estimated in?

ARIMAX fitted on an ARIMA(2,0,0) base, so that Bitcoin's own momentum is fitted first and the liquidity term is measured on what the momentum leaves behind.

The order was selected by an algorithm rather than chosen by the author. The procedure is the automatic order-selection method of Hyndman and Khandakar, cited in full below, and it arrived at a base carrying two weeks of Bitcoin's own returns. That matters for what the liquidity coefficient means: the momentum is fitted first, so the coefficient measures the part of Bitcoin's weekly return that its own recent returns do not already account for.

The two momentum terms are not statistically significant. Over this sample Bitcoin's short-term returns carry little information about the following week, which is worth reporting rather than passing over: it means the fit is not being carried by momentum, and it is what separates this result from a restatement of the trend that a levels correlation would have produced.

The residual diagnostics are reported rather than summarised. The Ljung-Box test finds no remaining structure in the residuals. Residual autocorrelation at the first lag is effectively zero, which confirms that the momentum terms absorbed the part of the return that Bitcoin's own history accounts for instead of leaving it in the error. The quantile plot tracks the normal distribution through the central mass and departs from it at the extremes, which is Bitcoin's documented fat-tailed behaviour and is a limitation of the specification rather than a defect in the estimate.

Is there a check on this that uses no regression at all?

Yes. Every week in the sample is sorted by the state of the index at the selected lag, and the three resulting groups are compared on the distribution of Bitcoin's weekly return.

The check uses no model, no coefficients and no distributional assumptions. It sorts. Every week in the sample is labelled by what the index was doing at the lag the scan selected, using a neutral band either side of zero so that ordinary reporting noise in central bank aggregates is not read as a change of state.

Three labels result: withdrawal, when the index sits below the band; neutral, when it sits inside it; and expansion, when it sits above. The question the check asks is whether Bitcoin's return distribution differs across the three groups, and whether any difference is ordered rather than arbitrary.

The reason to run it is that it fails differently from the model. A fitted result can be wrong because the specification is wrong. A sort cannot be wrong in that way. It can only be wrong because the sample is short or the labels are badly drawn, which are different failure modes with different remedies. Two tests that fail for different reasons are worth more than two tests that fail for the same one.

Is the lead stable?

No. Every study in the record that tested the stability of the lag found that it moves, and two found the relationship absent over long stretches of the sample.

CryptoSlate ran rolling 180-day correlation windows on daily data and reported an optimal lag of 42 days in the window ending 20 November 2025, with a correlation of 0.16, and stated in terms that the optimal lag is not fixed. asapdrew, working on G4 broad money to January 2026, put the optimal lag near two months in a band of one to three months, and recorded the band widening from one to two months in early 2024. Both are model estimates on the samples named beside them and past results do not indicate future results.

CryptoSlate also split its sample at Bitcoin's peak on 6 October 2025 and reported a correlation of 0.89 before that date and -0.49 after it. A relationship that changes sign across a single date is not a constant, and this page does not treat it as one. Those are model estimates as of 23 November 2025 and past results do not indicate future results.

The two academic treatments go further. Gu and Chen tested weekly data from December 2015 to February 2025 and found no cointegration over the full sample and no overall Granger causality, with significance appearing only in a handful of sub-periods. Onramp, on monthly data from May 2014 to January 2026, found month-to-month correlations weak and correlations at six to twenty-four month horizons materially stronger, with the optimal lag varying across the sample rather than holding fixed.

That is why a re-estimation date matters more here than a decimal place. A lag fitted on one regime can fail in the next, so the honest form of the claim is not a constant but a measurement with a date on it.

What does a 0.94 correlation between Bitcoin and liquidity mean?

Less than it looks. The headline figure is a correlation of the levels of two trending series, and the same study reports much lower numbers on rolling windows.

Sam Callahan's work for Lyn Alden, published September 2024, measured Bitcoin against global broad money across eight economies from May 2013 to July 2024. It reported a correlation of 0.94 over the full period, 0.51 on 12-month rolling windows and 0.36 on 6-month rolling windows. All three come from the same study. Quoting the first without the other two is the most common move in this topic. Those are model estimates over the sample stated and past results do not indicate future results.

The reason the numbers fall is arithmetic rather than dispute. Two series that both trend upward over a decade correlate on levels almost regardless of what happens between them. Expressed as a rate of change, the same relationship produces the smaller numbers that independent measurement produces: Keyrock reports r = +0.26 for broad global liquidity against Bitcoin since 2021, a partially verified figure whose page returned 403 to automated fetch, and CryptoSlate's return-based scan found 0.16 at its optimal lag. Both are model estimates, published June 2026 and November 2025, and past results do not indicate future results.

This is the reason BlockPhi measures its own series on a rate of change rather than on levels. It is the harder test and it produces the less impressive number, and a rate-of-change coefficient in that lower range is the honest answer to a question that a levels correlation answers flatteringly.

When does the relationship fail?

When money growth and Bitcoin move apart, which they did through 2025 and into 2026. The record also shows the link absent over long stretches of the sample.

Bitcoin's growth diverged from global broad money growth from the middle of 2025. By January 2026 Bitcoin showed negative growth year over year while global broad money grew by more than 10% year over year. Any page asserting that the relationship is currently tracking, with no date on it, is asserting something the 2026 record contradicts. Those figures are reported as of 13 January 2026 and past results do not indicate future results.

The failure conditions worth stating plainly are these. The transmission window moves, so a lag fitted on one regime can fail in the next. The relationship has been statistically absent over long stretches of the record rather than merely weak. And a series converted into US dollars partly measures the dollar, which is a separate driver with its own cycle. None of that makes the relationship unreal. It makes an undated figure worthless.

What are the limitations of this work?

Four limitations that BlockPhi states about its own work, and one the external record adds. They are published in full, because a limitation stated by the author is worth more than one found by a reader.

The index is proprietary, and that is a limitation rather than a disclosure preference. Nobody outside BlockPhi can rebuild it, so nobody outside BlockPhi can independently check a figure estimated on it. Every result BlockPhi publishes on the index inherits that, and a reader is entitled to weigh it against results taken on measures that are public.

The sample is short. It runs to approximately four years of weekly observations, which is about two Bitcoin cycles, and two cycles is a small number on which to establish a cyclical relationship. A relationship that holds across two cycles has not been tested against a third.

The model is linear and Bitcoin is not. Bitcoin's crash dynamics are faster than its rally dynamics, so a specification carrying one coefficient in both directions will describe the tails less well than the centre. A threshold extension, which fits different coefficients in different states, could improve the tail modelling. It has not been fitted, and until it is, the tails are the part of this work with the weakest support.

The lag may shorten. Bitcoin's institutional depth has grown across the sample, and an asset reachable through the same accounts that hold Treasury bonds sits fewer intermediary steps away from central bank liquidity than one that is not. If that continues, the transmission window compresses, and a lag fitted on this sample would overstate the lag that follows it.

One further limitation, which BlockPhi's own documents do not state and the external record does: every independent study in the published reconciliation that tested the stability of a liquidity lead found that the lead moves. That is not a criticism of any single estimate, including this one. It is the reason a lead figure means little without the date it was last re-estimated attached to it.

What sources does this page rely on?

All of them, listed with the measure, the publication date and the URL, read on the date shown in the caption.

The reconciliation table above lists every published lead figure. The table below lists the other work cited on this page: the definitional authority, the correlation statistics, the cross-asset sensitivities and the record of the recent divergence.

What methodological literature does this derivation rest on?

Six works: the framework, the two results that decide how the data may be transformed, the order-selection procedure, and two on the structure of funding itself.

The works below are cited by BlockPhi's own working paper on Global Liquidity transmission and are listed here with full bibliographic detail so that a reader can go and check them. Each identifier was resolved through Crossref on the date in the caption rather than reproduced from the paper's reference list, which carries no identifiers.

They are the methods and the framework this derivation uses. None of them is evidence for a figure BlockPhi publishes, and none is cited here as corroboration of one. That distinction matters most for the first row: Michael Howell's book is the framework this work is built in, and CrossBorder Capital's index, which appears separately in the reconciliation table above, is a different measure that BlockPhi's figures are not taken on.