DLI 2.0: Why We Rebuilt the Index Around Net Liquidity Instead of CISS

What broke in the first version

The original DLI was a CISS-style aggregation: four groups of indicators, each reduced to a rolling tightness percentile, combined through a rolling-correlation quadratic form, with the headline expressed as the trailing five-year percentile of that composite. It is a respectable design — the ECB uses it for systemic stress — and it was the wrong design for this index.

The evidence was a pair of correlations. The composite tracked the Chicago Fed National Financial Conditions Index at roughly +0.28, which is weak but not absurd for a financial-conditions proxy. Against the thing users actually mean when they say "dollar liquidity" — the net-liquidity flow out of the Fed and Treasury — it ran about −0.10. The index was a mediocre financial-conditions index wearing a dollar-liquidity name.

Every fix we shipped for two years was a patch on that mismatch rather than the mismatch itself. The Group A transform went from raw level, to trailing one-year change, to a median-referenced one-year change, each iteration solving the symptom that the previous one exposed. The June 2026 decision was to stop patching the aggregation and make the spine of the headline the net-liquidity flow itself.

The spine: net liquidity, smoothed

Net liquidity is the Fed balance sheet minus the two largest drains on bank reserves: the Treasury General Account and the overnight reverse repo facility. WALCL minus TGA minus ON RRP. Nothing in that expression is novel — it is the formula every macro desk quotes — and that is exactly why it belongs at the centre of an index named after dollar liquidity.

The raw series is unusable as-is. WALCL prints weekly and steps; TGA and ON RRP print daily and are noisy around tax dates, settlement dates and quarter-ends. So the spine is a ten-day exponential moving average of the level, which kills the weekly step function without introducing a lag long enough to matter at the horizon this index describes.

From level to impulse: the anchor-free flow estimator

The index does not score the level of net liquidity. It scores the flow — whether liquidity is being added or drained right now. The first version of the flow estimator used point anchors: two-thirds of the 182-day change plus one-third of twice the 91-day change. That is the obvious construction and it has a defect that took a diagnostic to find.

When you difference against a point 182 days ago, the estimate moves for two reasons: what happened today, and what happened 182 days ago sliding out of the window. We measured the split. Between 75% and 77% of daily changes in the flow estimate came from the anchor moving across old data, not from the current tape. The index was reacting to history scrolling past.

The replacement is an exponentially weighted daily slope: an EWMA of the one-day change in the smoothed level, with a 60-business-day half-life, scaled by 126 to express it as a six-month-equivalent flow in trillions. Every day enters the estimate once, with a weight that decays smoothly, and no single historical date can dominate. That number is then mapped to an impulse in the zero-to-one range: 0.5 minus the flow divided by 1.3 trillion, clamped. Zero is maximum loosening, 0.5 is flat, one is a maximum drain.

The funding override is a noisy-OR, not a weighted average

Macro flow is not the only way dollar liquidity goes wrong. The plumbing can seize while the balance sheet looks fine — September 2019 is the canonical case, when reserves were still counted in the trillions and overnight repo printed 10%. So the headline carries a second term built from four funding legs, each a ramp from a no-stress floor to a full-stress cap: SOFR minus IORB (2 to 15 basis points), the 99th-percentile SOFR minus IORB tail (which blows out at quarter-ends before the median moves), Standing Repo Facility usage (50 million to 20 billion dollars), and a rise-gated swap-lines leg (10 to 200 billion, measured over a 28-day rise so repayment tails do not charge it).

Those four are combined with a maximum, not a sum: any one of them at full stress means the plumbing is stressed, and three quiet legs should not dilute a fourth that is screaming. The result is smoothed with a ten-day EMA before it enters the headline, because a single quarter-end print is a technical, not a regime. It is then wrapped in a decay floor with a fifteen-day half-life: realized stress persists, so the override tracks the base upward immediately but is not allowed to fall faster than that half-life.

The combination with the flow impulse is a noisy-OR: tight equals one minus the product of the complements. Dollar liquidity is tight if the macro flow is draining, OR the plumbing is stressed, OR the cushion is scarce. A weighted average would let a calm plumbing reading cancel out a genuine drain, which is precisely backwards — these are alternative routes to the same outcome, not competing measurements of one quantity.

Absolute, not percentile: the change that stopped the jumps

The most consequential change is the least visible. The old headline was the trailing five-year percentile of the composite. The new one is simply the smoothed tightness value multiplied by one hundred and rounded — an absolute reading.

Ranking an already-bounded signal against its own history is doubly relative, and it fails in two specific ways. First, it drifts: the score moves when nothing in the data moved, because an old extreme aged out of the five-year window. Second, it jumps: when the current value sits in a dense region of the historical distribution, a small move in the underlying crosses many ranks at once. We measured single-day headline moves of up to 28 points on the percentile construction. On the absolute construction, the worst single day across ten years of history — including March 2020 — is 6.8 points.

The absolute reading also means the number means the same thing every year. A 70 in 2019 and a 70 in 2026 describe the same liquidity stance. A percentile 70 describes only "tighter than 70% of the last five years", which is a different statement each time the window slides. The regime cuts are fixed: below 33 is loose, above 67 is tight.

What it had to pass before it shipped

A methodology change to a published index is only as good as the gates it clears, so the rebuild was held to explicit ones. Direction agreement against benchmark financial-conditions indices came in at 78% against a gate of at least 75%. Phantom drift — how much the headline moves during windows where the underlying data is flat, the failure mode that killed the percentile version — had a median of 3.9 points against a gate of at most 8. The QE episode gate required the index to read genuinely loose during large-scale asset purchases: it scores 13.8. The QT episode gate required genuinely tight: 79.8. And the funding override had to pin at its maximum during both the September 2019 repo crisis and the March 2020 dash for cash. It does.

One scope note matters more than any of the numbers. This is deliberately a dollar-liquidity gauge, not a financial-conditions index. It does not fold in equity volatility, credit spreads or dollar strength — those stay as context panels. The consequence is that an equity-stress event with calm plumbing and an expanding balance sheet, such as the March 2023 regional-bank episode, reads neutral-to-loose here. For a financial-conditions index that would be a failure. For a liquidity index it is the correct answer, and pretending otherwise is how an index ends up meaning nothing in particular.