Version 2026-08-01. Data through 2026-07-31. Window 2003-01-31 to 2026-07-31, 283 months.
CONSTRUCTION PIT ONLY These retrospective statistics are not validated-backtest evidence. chronological transforms use final/current-vintage history and do not reconstruct what was publicly knowable on each historical date.
The firm that coined "global liquidity" as a macro category publishes three quantitative claims for its index family: liquidity leads asset prices by 3 to 6 months, liquidity leads the business cycle by 12 to 15 months, and global liquidity runs a cycle of roughly 65 months. The index is proprietary. The claims ship without error bars, out of sample tests, or a revision record. So this page does what any referee does: it takes the largest component everyone agrees on, G3 central bank balance sheets in dollar terms, rebuilds it from public data, and tests the claims under the same standards this board applies to itself.
Scope, stated above the fold. This is not a test of the proprietary index. It is a test of its publicly reconstructible layer: Fed, ECB and Bank of Japan total assets at monthly average spot, month end, 2003 onward. The full index adds private credit and cross border flows across some 80 countries back to 1974. If the full index performs better than its public layer, that performance lives entirely in the unpublished layers, and that is a claim its publisher could settle tomorrow by releasing an error distribution. We would run it the day they do.
The reconstruction today: 17.49 trillion dollars (Fed 6.74, ECB 6.79, BoJ 3.97), -6.6 percent year on year.
Year on year growth of the G3 aggregate, correlated with forward equity returns (Nasdaq Composite, the only broad index with full window public history). Intervals are circular moving block bootstrap.
| horizon | correlation with forward return |
|---|---|
| 1 months ahead | -0.01 in [-0.17, +0.14], n 271 |
| 3 months ahead | -0.01 in [-0.23, +0.21], n 269 |
| 6 months ahead | +0.00 in [-0.26, +0.25], n 266 |
| 9 months ahead | -0.02 in [-0.31, +0.24], n 263 |
| 12 months ahead | -0.07 in [-0.34, +0.18], n 260 |
| 18 months ahead | -0.14 in [-0.42, +0.13], n 254 |
The walk forward test (2011-01-31 to 2026-02-28, signal from trailing data only) puts the 6 month forward log return after high liquidity months at +0.082 against +0.066 after low liquidity months: a spread of +0.016 with a 95 percent interval of [-0.033, +0.065]. An interval that straddles zero is the test refusing to certify the claim.
Causality, both directions: liquidity to returns F 0.56 (p 0.76), returns to liquidity F 1.19 (p 0.31), 6 lags. Neither direction clears conventional significance on this layer.
The transform is not the story: rerunning the forward correlations on 6 month annualized growth gives the same flat zeros.
| horizon (6m growth) | correlation |
|---|---|
| 3 months ahead | +0.01 in [-0.19, +0.20], n 275 |
| 6 months ahead | +0.04 in [-0.26, +0.26], n 272 |
| 12 months ahead | +0.00 in [-0.35, +0.31], n 266 |
Neither is the definition. The strongest known variant of the claim is net liquidity: central bank assets minus the Treasury cash balance minus reverse repo take up (today 5.84 trillion dollars, window 2003-02-28 to 2026-08-31). The level chart of that series against equities is the famous exhibit. The honest question is whether its growth predicts forward returns, and it reads the same way:
| horizon (net liquidity) | correlation |
|---|---|
| 3 months ahead | -0.04 in [-0.24, +0.25], n 166 |
| 6 months ahead | +0.02 in [-0.16, +0.23], n 163 |
| 12 months ahead | +0.02 in [-0.26, +0.20], n 157 |
| 6 months ahead, 2015 onward | +0.09 in [-0.42, +0.41], n 134 |
Its walk forward evaluation (2019-08-31 to 2026-02-28) lands on exactly the era the exhibit is drawn from, and puts the 6 month spread at +0.001 with a 95 percent interval of [-0.142, +0.115].
Stability check, forward 6 month horizon:
| sample | correlation |
|---|---|
| full sample | +0.00 in [-0.26, +0.25], n 266 |
| 2003 to 2014 | -0.01 in [-0.46, +0.38], n 132 |
| 2015 onward | +0.07 in [-0.31, +0.33], n 134 |
Whatever the 3 to 6 month lead is, it is not in the central bank layer, gross or net.
Against US industrial production growth, the contemporaneous correlation is -0.41: central banks expand when the economy contracts. The correlation turns positive only far out, peaking at a lead of 22 months (+0.23 in [+0.05, +0.41], n 248). At the claimed 13 month lead it reads +0.15 in [-0.14, +0.46], n 257, an interval that straddles zero.
Two honest readings. Either the lead is real but sits well past the claimed window, or the pattern is a policy reaction function: recessions trigger balance sheet expansion, recoveries follow recessions, and the apparent lead is partly the economy forecasting the central bank rather than the reverse. Separating those requires exactly the private credit layer that is not public.
On 271 months of data the periodogram's dominant period is 68 months, but the spectral resolution at a 65 month period is plus or minus 16 months: the sample cannot separate 65 from its neighbours. If the claim is true this window holds about 4.2 cycles. Peak to peak spacings of the smoothed series run 25, 27, 35, 29, 28, 51, 34, 27 months.
This deserves saying with respect: the long homogeneous history behind the proprietary index is the one genuinely irreplaceable thing its publisher owns, and it is the only instrument that could settle their own headline claim. We would rather they published the test than the assertion.
Verdict tokens are cheap; the evidence matrix behind them is the standard. Eight coded rows, adapted from the execution assumption matrix of arXiv:2606.08285: the two rows a non trading terminal cannot honestly self grade (cost and turnover treatment, execution semantics) are re coded as vintage and revision handling and threshold provenance, and a verdict revision policy row is added. The ordering rule is the feature: Seiche grades itself first, on the same rows, and publishes its own PARTIAL rows with the gaps named.
4 PASS, 4 PARTIAL, 0 FAIL, 0 not applicable.
| row | grade | evidence |
|---|---|---|
| point-in-time controls | PASS | Expanding window discipline is enforced by unit test: test_history_has_no_look_ahead appends future data and requires every already computed index value to stay identical, and the same invariant is pinned on undertow, tidetables, swell and bathymetry. The Leak Audit then rebuilds the same index with the two dominant leak classes deliberately switched on (full sample standardization, a centered smoother) and publishes the Leakage Gain each break would buy. |
| train/eval split transparency | PARTIAL | Nothing in the composite is fitted, so there is no training set to disclose: the weights and the alert percentile are frozen in config.py and PROOF prints its warmup slice and its caveat list. The gap: those editorial choices were made by an operator who had already seen the full history, and no document separates the data that shaped the design from the data that now grades it. Disclosed, but not separated. |
| held-out evaluation | PARTIAL | The genuinely held out sample is the live record since first publication: the PIT ledger commits one record per day per stream, and the attest layer signs each committed record (Ed25519) and anchors it to Bitcoin through OpenTimestamps, so a post publication reading is sealed before its outcome. The backtest window itself is not held out: the design saw it. The true out of sample record is the attested one, and it is still short. |
| universe and series definition | PASS | Every input is a registered SeriesSpec in config.py with mnemonic, source ID, fetch start and native lag. The funding event is defined once (backtest.pop_bp: SOFR minus IORB against its trailing 5 business day median, 10bp threshold) and every consumer imports that definition. Registry tests pin fetch starts so a retune cannot silently starve an engine. |
| artifact release | PASS | The whole pipeline is public AGPL-3.0 code; the board snapshot is baked to frontend/public/data/overview.json on every publish; the Time Machine endpoint (/api/asof/{date}) replays the full light board as of any historical date; and the PROOF scoreboard has its own signed and Bitcoin anchored stream (attest prove-scoreboard). |
| vintage and revision handling | PARTIAL | PROOF consumes final vintage values: the FRED collector reads fredgraph.csv, which serves the current revised history, and no ALFRED as published replay exists in this codebase. The PROOF page prints the caveat (weekly H.4.1 aggregates are lightly revised; daily market prints effectively are not) and the PIT ledger accrues true as published readings from publication day forward, but the historical backtest itself rides revised values. That is the gap, named. |
| threshold provenance | PARTIAL | Editorial, frozen, disclosed: the alert percentile (80) and the composite weights sit in config.py under a TUNING POINT banner that calls them the tool's editorial voice, and nothing in the pipeline refits them. The Leak Audit's THRESH_FIT row publishes what an in sample fitted threshold would have bought, and PROOF's AUROC is threshold free, so the cost of the editorial choice is measured. But measured is not derived: the numbers come from judgment, not from a walk forward derivation, and this rubric does not round that up. |
| verdict revision policy | PASS | Published readings are append only: the notary hash chains every reading so editing any past day breaks every link after it, and the attest layer makes a rewrite attributable (signature) and undatable (Bitcoin anchor). The referee, methodology and skeptic pages carry dated changelogs, and the PROOF page's stated contract is that unimpressive numbers publish anyway. |
Four rows PASS, four are PARTIAL with the gaps named: no as published vintage replay, a still short attested live record, thresholds measured but not derived, and a design that saw its own evaluation sample. Each PARTIAL is locked by a regression test and flips only when the missing evidence ships.
0 PASS, 1 PARTIAL, 7 FAIL, 0 not applicable.
| row | grade | evidence |
|---|---|---|
| point-in-time controls | FAIL | The claims ship with no stated look ahead protections. Rebuilt with them (trailing median signal, no look ahead), the public layer's 6 month spread is +0.0158 with a 95 percent interval of [-0.0327, +0.0653]. |
| train/eval split transparency | FAIL | No design or evaluation split ships with the claims: no error bars, no out of sample tests, no design sample named. The referee page states that absence above the fold. |
| held-out evaluation | FAIL | Nothing is held out because nothing is released to hold out. The referee's own held out test on the public layer refuses to certify the asset price lead; at the claimed 13 month business cycle lead the correlation reads +0.15 in [-0.14, +0.46], n 257. |
| universe and series definition | PARTIAL | The largest agreed component is public and is reconstructed here: G3 central bank assets at monthly average spot, 283 months from 2003-01-31 to 2026-07-31. The remaining layers, private credit and cross border flows across some 80 countries back to 1974, are defined in prose and reproducible by nobody outside the firm. |
| artifact release | FAIL | The index is proprietary: no series, no code, no error distribution. The referee page commits to rerunning every test the day an error distribution is released. |
| vintage and revision handling | FAIL | No revision record ships with the claims, so as published values cannot even be distinguished from today's history. An index family whose vintages are unpublished is ungradeable on this row, and ungradeable by choice is a FAIL, not an N/A. |
| threshold provenance | FAIL | The claimed windows, 3 to 6 months, 12 to 15 months and a 65 month cycle, arrive with no stated derivation. Tested on the public layer, the business cycle correlation peaks at a 22 month lead. The sample's spectral resolution at 65 months is plus or minus 15.6 months: too coarse to certify the cycle either way. |
| verdict revision policy | FAIL | The claims are undated and carry no changelog and no stated policy for revising or retiring a claim that fails. A claim that cannot fail in public is copy, not a forecast; the referee page's dated changelog is what the alternative looks like. |
Seven rows FAIL and one is PARTIAL, all from one root cause: nothing is released that would let an outsider grade the claims any better. The rubric grades the published record, and the published record is assertion.
A published lead lag distribution on the full index, with revision errors and out of sample splits. That is the standard this board holds itself to: every reading sealed before outcomes, every miss on the public record. Until then, the reproducible layer of the global liquidity story contains no tradable asset price lead, and the business cycle lead sits far from where the marketing puts it.
Fed total assets (H.4.1 via FRED WALCL), ECB total assets (ECBASSETSW),
BoJ total assets (JPNASSETS), converted at monthly average spot (DEXUSEU,
DEXJPUS); Nasdaq Composite (NASDAQCOM); US industrial production (INDPRO).
Derived statistics only are republished here. G3 central bank assets are converted to dollars at monthly average spot and summed month end from 2003. Year on year growth of that aggregate is correlated with forward equity returns at horizons of 1 to 18 months and with forward industrial production growth at 0 to 24 months; every correlation carries a circular moving block bootstrap interval. The predictive claim is additionally tested walk forward with a trailing median signal and no look ahead, and with lagged OLS F tests in both causal directions. Cycle length is read from a linearly detrended periodogram with its resolution limit stated. The transform choice is stress tested by rerunning the forward correlations on six month annualized growth, and the strongest known variant of the claim, net liquidity defined as central bank assets minus the Treasury cash balance minus reverse repo take up, gets the same forward correlations and the same walk forward test. Historical inputs are final/current-vintage unless the attached historical_evidence manifest proves otherwise; chronological alignment alone is not a validated-backtest claim.
The G3 reconstruction is deliberately incomplete: no keyless PBoC or Bank of
England balance sheet feed exists, and the gap is stated rather than
interpolated over. The engine behind this page is
engines/refereegli.py; its method paragraph is on the
methodology page and the code is at
https://github.com/beepboop2025/seiche under AGPL-3.0. The two questions a
skeptic asks first are answered in the
skeptic pack.
Free public data with native lags. Not investment advice. Seiche is free open source software (AGPL-3.0) and a public good.