Methodology

Methodology Deep Dives

Full write-ups of how each layer of the terminal works.

Indicator Universe

17 displayed series: 10 scored indicators across 4 tiers plus 7 context/reference series. Each scored indicator's direction is normalized so a positive z-score always means "tighter" liquidity.

Context (not scored)

• Net Liquidity = Fed BS - TGA - ON RRP
• M2 Money Supply
• Fed Central Bank Liquidity Swaps
• 90D Financial CP - 3M T-Bill Spread
• Foreign Holdings of US Federal Debt
• ECB Total Assets
• Bank of Japan Total Assets

Scoring

1. Robust z-score

For each indicator, we compute a robust z-score using the median and median absolute deviation (MAD) over a rolling 10-year window:

z = (value - median) / (MAD x 1.4826)

The 1.4826 constant makes MAD comparable to standard deviation for normally distributed data. Z-scores are winsorized to [-4, +4] to prevent extreme observations from dominating the composite. We use median/MAD rather than mean/std for resilience to structural breaks (regime changes, COVID, etc.).

2. Direction normalization

Some indicators rise when liquidity tightens (TGA, ON RRP, VIX, HY Spread, etc.); others fall (Fed BS, Bank Cash Buffer). For "falling-is-tightening" indicators we negate the raw z-score so a positive z always means "tighter" across every indicator.

3. Composite

composite = sum (z_i x weight_i)

  Tier A (Policy/Reserves):       3 x (13/60) ~= 0.65
  Tier B (Funding/Plumbing):      2 x 0.05 = 0.10
  Tier C (Credit/Intermediation): 2 x 0.025 = 0.05
  Tier D (Risk/Price):            3 x (1/15) ~= 0.20
                                              ----
                                              1.00

When indicators are missing on a given day, the remaining weights are renormalized so a partial-data day still produces a comparable composite.

4. Regime classification

We rank today's composite within the rolling 5-year history of composites and assign a regime:

• Loose - <= P20
• Mildly Loose - P20-P50
• Mildly Tight - P50-P80
• Tight - >= P80

5. Momentum & concentration

• Momentum: composite_today - composite_7d_ago. >+0.15 = deteriorating, <-0.15 = improving, else stable.
• Concentration: top-3 |contributions| / total |contributions|. High concentration = signal driven by a few drivers.

Asset Impact Lens

For SPX, QQQ, BTC, and GOLD we precompute three relationship windows on every cron cycle (SPX/QQQ/BTC from FRED; XAU/USD from Twelve Data after FRED retired the LBMA daily series):

• Rolling correlations (30D / 90D / 180D) between the daily Liquidity Index composite and daily log returns.
• Lead-lag correlations (the Liquidity Index leads asset returns by 10D / 20D / 60D).
• Conditional state outcomes: average forward 20-day cumulative log return + win-rate when the Liquidity Index was in the loose (<= P20) or tight (>= P80) state.

Liquidity Index v1 — served headline since 2026-09-09

v1 — a relative reading. Ten inputs (Fed balance sheet, TGA, ON RRP; SOFR–IORB, SRF take-up; bank cash buffer, HY spread; VIX, dollar index, 10y real yield) are each ranked against their own trailing five years, combined in four tiers weighted 65/10/5/20, and the composite is ranked again against its own five-year history. A reading of 80 means tighter than 80% of the past five years, not a fixed level. Known limitation: once the ON RRP balance is exhausted its level percentile stays pinned at the bottom of its range, which keeps pulling the policy tier toward loose; the grid is calendar-day and the Fed balance sheet is scored by level percentile only.

Liquidity Index Methodology v2 (headline 2026-07-11 → 2026-09-09)

The Liquidity Index headline was computed under Methodology v2, adopted 2026-07-11 after a pre-registered anchor-based selection process. Two defects of the previous methodology were corrected: (1) a depleted, flatlined indicator (e.g. ON RRP near zero) could pin at the bottom of its percentile range and permanently dilute its tier toward "loose"; (2) the Fed balance sheet was scored by its level percentile alone, which misread the 2022 QT onset as loose.

v2 introduces a uniform information gate — an indicator whose percentile is pinned at a distribution boundary while its raw values flatline carries no marginal information and is ramped out of the active set (and ramped back in when it becomes informative) — and scores the Fed balance sheet as the mean of its level percentile and its 13-week-flow percentile, capturing both stock scarcity and the pace of change. All other pipeline stages are unchanged.

The full v2 history was recomputed under one contract and is stamped methodology_version = 2. Because v2 removes the dilution and level-percentile defects, its readings sit structurally higher than v1 in depleted-buffer regimes — the step from v1 to v2 at cutover is a documented methodology correction, not a market move. On 2026-09-09 the owner withdrew v2 as the served headline and restored v1 as the default; v2 computation was stopped on 2026-09-09; its stored history is frozen and remains queryable via ?track=v2 until 2026-12-08, and the disposition after that date is a separate owner decision (see docs/dli-v2/DLI_OWNER_RULING_V2_HEADLINE_WITHDRAWN_2026-09-09.md and DLI_OWNER_RULING_V2_COMPUTE_STOPPED_2026-09-09.md).

Liquidity Index v3.1

v3.1 is a parallel, independently-built liquidity gauge shown beside the official v1 reading as a labelled v3.1 reading. It is not a forecast, not a second opinion on the headline, and not a candidate that has been accepted.

• Scale: an absolute 0-100 with fixed bands (loose < 33, neutral, tight > 67). It is NOT a percentile, so it is not comparable to the official v1 percentile reading number-for-number.
• Construction: a net-liquidity flow spine (Fed balance sheet - TGA - ON RRP, six-month equivalent, calendar-day parameterised), a reserve-buffer stock leg that amplifies drains when the buffer is thin, and a funding-stress override built from SOFR-IORB and Standing Repo Facility take-up.
• Key contributions: the v3.1 card decomposes the raw score exactly. raw = spineGain × amplified impulse + overrideWeight × override, where the amplified impulse = impulse + (buffer gain − 1) × rectified(impulse). The three displayed parts are the net-liquidity flow term (spineGain × impulse), the buffer-thinness amplification term (spineGain × (amplified − impulse); the rectifier passes drains only, so it reads 0 during injections even when the gain is above 1) and the funding-stress override term (overrideWeight × override). They sum to the raw score by construction; the score is the logistic map of raw onto 0–100.
• Record-run verdict: the pre-registered acceptance battery was run once, on 2026-08-09, and 4 of its 14 blocking gates failed. v3.1 was stopped in research state and the failures were recorded as findings rather than repaired. Nine further gates flipped somewhere inside a plus/minus 50% perturbation of their pinned constants and are marked fragile.
• Why some gates fail: three of the four failures are registered conflicts between the acceptance battery and the construct v3.1 implements - a pure liquidity-stance gauge, in which rate-driven parking at the Fed is not treated as a drain on the level. That is a disagreement about what is being measured, and it was recorded rather than resolved by moving a threshold.
• Status of the official reading: v1 is the headline on every surface and every default; v3.1 is displayed alongside it under its own label and is not a default anywhere.

Data Sources

• FRED - Federal Reserve Bank of St. Louis, public economic time series
• US Treasury Fiscal Data - Daily Treasury Statement (TGA balance)
• NY Fed Markets - Standing Repo Facility usage