The reserve-demand curve is convex, so a drain into a thin cash cushion is not the same event as a drain into an ample one. How DLI 2.0 prices that asymmetry without falling into the reserve-level trap that a percentile weight fell into.
Take a 300 billion dollar drain of net liquidity over six months. In 2021, with the system awash in reserves and the reverse repo facility absorbing more than two trillion dollars of genuinely surplus cash, that drain is close to a non-event: it removes money nobody was using. In mid-2019, with the cash cushion sitting at 8 to 9 percent of commercial-bank assets, the same drain is what produced a 10 percent overnight repo print.
A model that scores the flow alone cannot tell those two apart, and for a while ours could not. The reserve-demand curve is convex: near scarcity, removing reserves spikes funding rates sharply, while adding reserves merely relieves scarcity. The response is asymmetric, and any faithful model of it has to be asymmetric too.
The reserve-relative gain multiplies the draining part of the flow — and only the draining part — by a factor that depends on how thin the system cash cushion is. The cushion is measured as reserves plus the overnight reverse repo balance, divided by total commercial-bank assets, expressed as a percentage. That series is the reserve buffer, and it is on the site as its own indicator.
The multiplier ramps smoothly between two empirical anchors. Above 14 percent — the threshold at which a cushion stops being comfortable — the multiplier is exactly 1.0 and the model behaves identically to the pure-flow version. As the buffer falls toward 8 percent, the floor observed during the 2019 repo crisis, the multiplier ramps to 2.5. The ramp is a smoothstep, so there is no kink at either band edge. Neither 8 nor 14 was fitted to maximise a backtest statistic; both are readings taken off the historical record.
The asymmetry is the load-bearing part. An injection is never rescaled. A drain into a thin buffer is scaled up by as much as 2.5x; an injection into that same thin buffer counts exactly as much as it would in an ample regime.
There is an obvious alternative that we tested and rejected: put the reserve level into the score directly, as a percentile weight. It fails badly. Reserves-to-GDP correlates with the Chicago Fed NFCI at about −0.745 — that is, the reserve level runs strongly against financial conditions across the ample-reserve era. A level term makes the index print "tight" in exactly the periods when reserves are being rebuilt and conditions are easing.
The 2024 to 2026 window shows why the drains-only asymmetry is the escape hatch. The buffer thinned toward 12 to 14 percent over that stretch — a level term would have been screaming tightness the whole way. But net liquidity was rebounding, because QT was ending, so the flow was positive. The gain multiplies only the draining part of the flow, and there was no draining part. It stayed dormant and contributed nothing.
The gain only bites when a drain resumes into a thin buffer. That is precisely the early-warning window the funding override misses. Through mid-2019 the override sat at roughly zero for months while the buffer sat at 8 to 9 percent and drained — the plumbing had not seized yet, so the override had nothing to report, right up until the morning it did. In backtest, adding the gain raises the mean 2019 repo-episode score from 72.9 to 78.8, while 2021 QE, 2022 QT, the March 2023 bank episode and the current reading are all unchanged, and the worst single-day headline move stays at 3.6 points.
The gain answers "how much does this drain hurt". It does not answer "how scarce is the cushion right now, drain or no drain". Mid-2019 sat at an 8 to 9 percent buffer for months with a roughly flat flow and a silent override, and the headline read neutral until the September 17 repo blowup landed. That is a real gap, and the fix is a third term.
The stock leg is a bounded penalty on a thin cushion, reusing the same 8-to-14-percent band as the gain, folded into the noisy-OR as a third factor. Its ceiling is 0.25, and that number is a semantic anchor rather than a fitted parameter: full scarcity alone lifts a neutral 0.5 headline to at most 0.625, which sits below the 0.67 risk-off line. Scarcity by itself can therefore never print risk-off. It takes an actual drain or actual funding stress on top.
That cap is what makes this term categorically different from the reserve-level weight we rejected. The level weight was unbounded and active across the entire ample regime, where it anti-tracked conditions. The stock leg is exactly zero everywhere the buffer is at or above 14 percent — which is all of 2020 through 2024 — and prices only cushions that are genuinely thin. In backtest it lifts the June-to-September 2019 pre-repo warning window from a mean of 58 to 75 and the repo episode itself from 88 to 91, leaves the QE and QT episodes untouched at 13 and 80, and moves the current reading from 44 to 47 on a buffer near 12 percent.
The headline answers which way dollar liquidity is moving. It does not answer how much cushion is left if a drain arrives, and those are genuinely different questions — August 2019 had a middling flow and a fragile cushion. So the state carries a second, orthogonal axis derived from the buffer and from whether the plumbing has started to twitch.
Low means the buffer is at or above 14 percent: a drain gets absorbed. Elevated means the buffer is below 14 percent but the plumbing is calm — a thin cushion with no trigger yet, which is where the system sits at the time of writing, with a buffer near 12 percent and SOFR running a couple of basis points under IORB. High requires the conjunction: a buffer below 14 percent AND a funding override that is lifting. That is the fragile cell, and it is what late-2025 quarter-ends looked like.
The conjunction is deliberate. Requiring both conditions is what keeps the vulnerability axis from re-creating the level trap: 2022 through 2024 looked thin-ish on some measures, but the plumbing was dead calm and the buffer was still above the threshold, so it reads low or elevated and never high.
Bank reserves plus the overnight reverse repo balance, divided by total commercial-bank assets, expressed as a percentage. It is a relative measure on purpose: a fixed dollar figure for "enough reserves" is meaningless when the banking system it has to support keeps growing.
Both are read off the record rather than fitted. Eight percent is where the buffer sat during the September 2019 repo crisis — the observed scarcity floor. Fourteen percent is the level below which a cushion stops being comfortable. The multiplier is 1.0 above 14 and ramps smoothly to 2.5 as the buffer approaches 8.
Because the reserve-demand curve is convex, so the response is genuinely asymmetric: near scarcity a drain spikes funding rates while an injection merely relieves scarcity. It is also what keeps the term out of the reserve-level trap. In 2024 to 2026 the buffer thinned while net liquidity was rebounding, so a symmetric term would have printed false tightness; a drains-only term stayed dormant because there was no drain to scale.
No. The scarcity stock leg is capped at 0.25, so full scarcity on its own lifts a neutral 0.5 headline to at most 0.625 — below the 0.67 risk-off cut. Printing risk-off requires an actual drain or actual funding stress on top of the thin cushion.
Deep Dive
The full methodology behind the June 2026 rebuild: the net-liquidity spine, the anchor-free flow estimator, the noisy-OR funding override, and why the headline became an absolute 0-100 reading instead of a rolling percentile.
Deep Dive
A deep dive into the Fed's ON RRP facility — how it works, why it surged to $2.5 trillion, how its drainage to near-zero fueled the rally, and what its depletion means for 2026 markets.
Deep Dive
Updated analysis of the Net Liquidity formula — its R² of 0.91 with SPX, component breakdown for 2026, what changed with ONRRP depletion, and a practical positioning playbook.