Monitoring US Treasury Supply-Demand with AI: A Practical Framework

Key Takeaways

• High Supply ≠ Crisis: 2024-2026 Treasury supply hit record highs, yet 10Y yield stayed stable at 4.0-4.5%, market absorbed smoothly
• Structural Optimization: Treasury adopted "more Bills, fewer Coupons" based on TBAC recommendations, Bills share rose from under 20% to over 25%
• Demand Differentiation: $7.8T+ MMF absorbed Bills, but foreign buyers shifted from sticky official demand to price-sensitive private demand—China holdings dropped to $683B (2008 low), private investors buying 2x+ what officials did
• Buffers Disappearing: ON RRP exhausted, basis trade leverage $1-2T, April Tax Day is the next liquidity stress test
• AI's Role: Swarm multi-agent system doesn't predict markets—it tracks marginal supply-demand changes, building contingency maps while consensus is still debating

A Counterintuitive Market Phenomenon

Since 2024, US Treasury supply has hit record highs—Q3 2025 alone saw net borrowing exceed $1 trillion ($1.058T), with H1 2026 estimated at over $1.2 trillion combined. By textbook logic: supply surge → price drop → yield spike.

But the market's answer was unexpected: The 10-year yield oscillated in a narrow 4.0-4.5% range, even with Middle East tensions pushing it to 4.48%, it quickly retreated. No 2022 UK pension-style liquidity crisis, digestion was surprisingly smooth.

Why?

Core Concept: Marginal Pricing

Here's something critical: Markets price based on marginal supply-demand balance, not absolute supply volume.

Think of an Apple Market

• 100 apples sold daily, 100 buyers want to buy → Price stable
• Suddenly 120 apples, but only 100 buyers → Price drops
• But what if 120 buyers show up too? → Price doesn't move

The Treasury market over the past two years is the third scenario. So the key questions are: Who's buying? What are they buying? Why are they buying?

Supply Side: Treasury's Duration Management

TBAC and Issuance Strategy

Understanding Treasury supply isn't just about how much is issued—it's about how issuance structure is determined. Each quarter's Quarterly Refunding is the core dialogue between Treasury and TBAC (Treasury Borrowing Advisory Committee)—TBAC comprises senior professionals from primary dealers and buy-side institutions, and their recommendations directly influence Coupon auction size decisions.

Throughout 2024-2025, TBAC recommended for consecutive quarters to "maintain nominal coupon and FRN auction sizes unchanged" (original: "Treasury anticipates maintaining nominal coupon and FRN auction sizes for at least the next several quarters"). This isn't passive inaction—it's an active choice: concentrating marginal supply increases at the Bills end, avoiding unnecessary pressure on long-end yields.

Structural Shift

Why This Adjustment?

“

The Treasury isn't "passively responding to markets"—it's actively managing duration structure. TBAC's optimal issuance model shows current composition is near the efficient frontier.

1. Strong Short-End Demand

Money market funds (MMFs) are the largest buyers of Bills:

• Early 2024 size: ~$5.5 trillion
• March 2026 size: $7.86 trillion (ICI weekly data, March 18, 2026)
• Growth driver: High-rate environment drove cash management demand surge, plus government fund expansion post-money fund reform
• Data sources: ICI, Fed H.4.1

2. Long-End Pressure Relief

Maintaining 10Y and 30Y auction sizes unchanged (10Y at $42B/auction, 30Y at $25B/auction):

• Avoids head-on collision with pension fund and insurer duration needs
• Preserves room to increase if fiscal outlook deteriorates

3. Fed Cooperation

October 2025 FOMC announced ending Treasury QT runoff December 1, while redirecting MBS maturing proceeds into T-bills:

• First batch of Reserve Management Purchases in December ~$40B, plus MBS reinvestment, monthly Bill purchases ~$55B
• Directly absorbs Treasury's new short-term supply
• Essentially a composition shift (from MBS to Treasury), not traditional QE—but in market effect, the Fed transformed from net seller to marginal Bills buyer

Demand Side: Who's Buying Treasuries?

Buyer Panorama

Structural Differentiation in Foreign Investors

Japan: Largest Holder, But Volatile

Japan is the largest foreign holder of US Treasuries. January 2026 TIC data shows $1,225.3B holdings, but the past two years haven't been one-way selling—there were increases and decreases, with volatility driven by yen exchange rates, BOJ YCC adjustments, and domestic rate environment. Overall trend oscillates around $1.1-1.2T range—not the "great exodus" markets feared.

China: Structural Selling, Down to 2008 Low

China's selling is the true trend change:

• December 2024: $759B
• November 2025: $682.6B (17-year low)
• Down over 48% from 2013 peak of $1.32T
• Drivers: FX reserve diversification, geopolitical risk hedging, gold accumulation (PBOC bought gold 15 consecutive months)

Who Filled the Gap?

• UK: Rose to second-largest holder ($877.9B), mainly reflecting London's role as global custody center
• Belgium/Luxembourg: Significant increases (Belgium from $374.6B to $477.3B), possibly indirect holdings by China via Euroclear
• Canada: Nearly tripled, reflecting energy export revenue and reserve expansion

Key trend: Foreign buyer composition is shifting from official institution-dominated to private investor-dominated. Latest TIC data shows private sector purchases of ~$158B, more than double official institutions' $64B. This means Treasury financing increasingly relies on yield-driven market capital rather than geopolitics-driven reserve recycling.

Hedge Fund Basis Trade: Double-Edged Sword

Basis trade is the "invisible buyer" in Treasury markets—it provides liquidity in normal times but can amplify volatility in stress scenarios. The March 2020 lesson was painful enough: margin call-triggered forced liquidation led to ~$100 billion in Treasury sales, ultimately forcing Fed intervention.

April 2025 Stress Test: Tariff shock caused violent Treasury market volatility, 10Y briefly broke 4.5%, 30Y hit 5%—but basis trade overall remained resilient. Dallas Fed research noted stability came from three "tailwinds": increased volatility added delivery option value, rate cut expectations lowered funding costs, Standing Repo Facility provided liquidity backstop. But these tailwinds won't always be there.

The Fed's Changing Role

From QT to Composition Shift

Why This Shift Matters

When markets no longer worry about "how much more the biggest seller will sell," term premium (ACM model) stabilized after rising from ~50bps to 80bps—not because risk disappeared, but because the biggest uncertainty vanished. This distinction matters: if you use Kim-Wright model, numbers differ, but direction is consistent.

How I Track These Changes

Pain Points of Traditional Methods

If you work in fixed income, traditional monitoring looks like this:

Daily checks of Treasury Direct for auction announcements, refreshing TIC database (foreign holdings, 2-month lag), waiting for FOMC meetings for wording changes, watching Bloomberg terminal for primary dealer inventory. The problems: information scattered across 10+ sources with different formats, update frequencies ranging from daily to quarterly, key marginal changes easily drowned in noise.

So I built a Swarm architecture multi-agent system to automate this process.

Monitoring Framework: Five Dimensions

Understanding Treasury markets requires watching both supply and demand:

Supply Side

1. Treasury issuance plans: Quarterly Refunding Announcements (QRA), TBAC recommendations, auction schedules and sizes
2. Market liquidity: Primary dealer inventory, repo rates (SOFR), SRF usage

Demand Side

1. Foreign investors: TIC monthly holdings (note 2-month lag), Japan/China/UK trend differentiation
2. The Fed: RMPs progress, MBS reinvestment pace, SOMA holdings changes
3. Domestic investors: MMF size (ICI weekly), pension allocation, hedge fund futures positions (CFTC COT)

Multi-Agent Architecture: Why Swarm?

Limitations of Traditional Pipeline

Traditional Agent architecture is serial—Supply Agent only watches supply, Demand Agent only watches demand, no communication.

But market analysis requires dynamic collaboration: When Supply Agent detects larger-than-expected auction sizes, Demand Agent should immediately assess recent foreign buyer TIC trends, Liquidity Agent should check primary dealer inventory space and repo rates, finally Risk Agent integrates to judge market digestion. These judgments need to happen simultaneously, not queue waiting for their turn.

Core Idea of Swarm Architecture

I use a Swarm architecture (similar to LangGraph Swarm), where agents can dynamically hand off control:

“

Agents can dynamically hand off control to each other, automatically switching focus based on market events, rather than executing along a preset pipeline.

Architecture and Data Flow

Data Layer                  Agent Collaboration Layer        Output Layer
─────────                   ────────────                     ──────
Treasury Direct ─┐
  (XML feed)     │
TIC Database ────┤       ┌──────────────┐
  (CSV, 2mo lag) ├──→    │  Orchestrator │──→  Brief / Alert
FRED API ────────┤       │  + Handoff    │
  (JSON)         │       │  + Shared Mem │──→  Telegram Bot
ICI MMF Data ────┤       └──────────────┘
  (Weekly)       │              ↕
CFTC COT ────────┘       Human-in-the-loop
  (Weekly)

Agent Collaboration Layer

Core Mechanisms

1. Dynamic Handoff

When Supply Agent detects auction size changes, proactively hands off control to Risk Agent:

Supply Agent: "Detected 10Y auction size increased 15%"
    │
    ▼ handoff_to("Risk Agent", context={auction_change: "+15%"})
    │
Risk Agent: Reads dealer inventory, recent TIC data from Shared Memory
    │
    ▼ "Primary dealer inventory elevated, indirect bid ratio needs attention"

2. Shared Memory

All agents share the same market state snapshot: historical auction data, holdings trends, risk indicators. Demand Agent doesn't need to ask Supply Agent "how much was issued recently"—reads directly from Memory. This Memory layer is also the persistence store for the MacroRAG system.

3. Human-in-the-loop

When the system can't decide, pause via Telegram Bot and request human confirmation:

Risk Agent: "30Y auction tail widened to 2bp, trigger alert?"
    │
    ▼ Telegram → Wait for human confirmation
    │
PM: "Yes, push alert + check 30Y basis spread"

Case Study: February 2026 Refunding

Take the Q1 Quarterly Refunding released February 4, 2026.

Swarm Collaboration Flow

T-3 days: Supply Agent detects QRA release
    │
    ▼ handoff_to("Demand Agent")
    │
Demand Agent: "Japan Jan holdings rose to $1,225B, China Nov dropped to $683B (17yr low)"
             "Foreign total holdings $9.31T, private buyers dominant"
    │
    ▼ handoff_to("Liquidity Agent")
    │
Liquidity Agent: "ON RRP near zero ($6B), but SRF operating normally"
                 "SOFR year-end spike faded, repo market stable"
    │
    ▼ handoff_to("Risk Agent", task="Comprehensive assessment")
    │
Risk Agent: "Risk level: Low-Medium.
            Key focus: ON RRP buffer exhausted, future liquidity shocks will transmit directly to reserves.
            30Y auction indirect bid ratio is the weathervane for foreign demand."
    │
    ▼ Generate brief → Telegram

System Output Brief

## Treasury Supply-Demand Brief - 2026-02-05

Supply Side
- Treasury Q1 QRA: Maintains Coupon auction sizes unchanged (per TBAC recommendation)
- Bills issuance continues expanding, 4-week Bill now at $101B/auction
- Q1-Q2 combined net borrowing estimate ~$1,261B

Demand Side
- Japan: Jan holdings rose to $1,225B (+$39.8B), stable amid volatility
- China: Nov dropped to $683B (lowest since 2008), structural selling continues
- Foreign total: $9.31T (new high), private investors main incremental buyers
- MMF: $7.8T+, continues absorbing Bills supply

Market Conditions
- 10Y yield: 4.35%, range-bound
- 30Y auction: Bid-to-cover healthy, indirect bid ratio needs tracking
- Liquidity: ON RRP ~$6B (essentially exhausted), repo market stable for now

Risk Alerts
- ON RRP buffer lost, liquidity shocks will act directly on bank reserves
- Basis trade leverage elevated, monitor MOVE index and repo funding conditions
- Whether China selling accelerates (watch Feb TIC data)

Current Market Risk Map

Calm doesn't mean problems disappear. Here are risk dimensions my system continuously monitors:

Immediate Risk: ON RRP Exhausted

This is the most notable change at time of writing. Overnight Reverse Repo (ON RRP) went from $2.5 trillion peak in 2023, to near zero mid-2025, to year-end spike to $106B on December 31, 2025, then back to $6B on January 2, 2026.

Why does this matter? ON RRP is the banking system's liquidity buffer. When it exists, shocks from TGA balance fluctuations, large Treasury issuance, quarter-end rebalancing get absorbed by ON RRP. Now this buffer is gone—every liquidity shock will act directly on bank reserves. The September 2019 repo market lesson (overnight rates spiked from 2% to 10%) happened in a similar buffer-exhausted environment.

Near-term Risks (3-6 months)

1. Basis Trade Fragility

Basis trade is a "stable marginal buyer" in normal times, but it's fundamentally a leverage-driven arbitrage strategy, whose stability highly depends on three conditions: repo funding stays cheap, dealers willing to intermediate, volatility doesn't breach margin call thresholds.

Dallas Fed July 2025 research noted basis trade stability is more sensitive to declining dealer intermediation capacity than rising funding costs. This means SLR constraints, G-SIB surcharge changes, or dealers' own balance sheet pressure could be more dangerous catalysts than rate moves.

2. Short-term Debt Refinancing Risk

If short-term rates spike again (say from re-anchored inflation expectations), Treasury faces "rollover risk"—cost of issuing new debt to pay old debt rises sharply, and higher Bills share means more concentrated risk.

Medium-term Risks (6-18 months)

1. Structural Changes in Foreign Buyers

China's selling isn't cyclical—it's structural: reserve diversification, gold accumulation, reduced concentration in dollar assets. $683B vs 2013 peak of $1.32T, down over 48%. The question isn't "will China buy more" but "can private investors sustainably replace official buyers?"

Private investors are price-sensitive—their purchases depend on yield attractiveness. When yields fall or dollar weakens, they may exit. This is fundamentally different from official reserve management's "sticky" demand.

2. Term Premium Direction

ACM model term premium rose from ~50bps to ~80bps. If it breaks 100bps:

• Long-end yields may passively rise
• Interaction with fiscal deficit expectations may form negative feedback loop
• Need to watch CBO updated forecasts and One Big Beautiful Bill's fiscal impact

Long-term Risks (18+ months)

Structural deficit won't disappear: Social Security, Medicare spending grows inexorably, interest expense now exceeds defense budget (CBO 2026 Budget Outlook). Moody's downgraded US sovereign rating from Aaa to Aa1 in May 2025, citing "deficit out of control." Tariff policy, geopolitics (especially Middle East transmission to oil prices and inflation expectations) add forecasting difficulty.

Positioning Implications: Where's the Asymmetry?

The framework's value isn't describing markets. Value lies in identifying asymmetric risks consensus hasn't fully priced. Here are three positioning-relevant observations flagged by the system as of end-March 2026.

1. Consensus Is Too Comfortable

Market base case is "high supply but digestible." Past 18 months basically correct—Coupon auction bid-to-cover stable at 2.3-2.6x, tails within 1bp. But this consensus implies an assumption: $7.8 trillion MMF will always be the infinitely elastic absorber at the Bills end.

MMF willingness to buy Bills depends on Bills yield vs ON RRP rate vs bank deposit rate spread structure. When Fed cuts (FFR 3.50-3.75%), Bills yield compresses, MMF marginal allocation may shift from Bills to repo or other short-term instruments. If this shift happens during Treasury's Bills issuance window—say post-debt ceiling TGA rebuild—marginal supply-demand balance may be more fragile than markets expect.

2. Term Premium Pricing Efficiency Declining

ACM 10Y term premium rose from ~0bps in 2024 to ~50-80bps now. Reasonable: deficit expansion, supply increase. But an overlooked nuance: divergence between ACM and Kim-Wright models is widening.

When two models disagree on the same quantity's estimate, it usually means yield curve dynamics are undergoing structural change—models' fitting assumptions are loosening. For curve trades: RV judgments based on term premium decomposition need wider confidence intervals. Belly (5-7Y) carry-adjusted richness/cheapness signals relative to wings are becoming less reliable.

3. Geopolitical Premium Is Regime-Change Catalyst

As of end-March, Iran situation pushed oil prices to 2022 highs, 10Y briefly touched 4.48%. Consensus debates one-time shock vs sustained inflation. But for rates positioning, the key isn't CPI path—it's whether term premium repricing of geopolitical uncertainty becomes persistent.

If Middle East conflict continues into Q2 and pushes oil higher: Even if Fed doesn't hike, markets will price out remaining 25bp cut. Long-end term premium path from 80bps toward 100-120bps may be shorter than most desk models imply.

Risk Scenario Matrix

System dynamically adjusts risk ratings based on incoming data. Here are currently active scenarios and triggers:

“

This table isn't prediction—it's a continuously updated contingency map. When left-column scenarios start appearing, you already know what to watch, how to judge.

System Current Status (As of 2026-03-31)

┌─────────────────────────────────────────────────┐
│  UST SUPPLY-DEMAND MONITOR    Status: AMBER     │
├─────────────────────────────────────────────────┤
│                                                 │
│  Supply:  STABLE                                │
│   · Coupon sizes unchanged (TBAC)               │
│   · Bills elevated, 4wk avg $101B               │
│   · Next QRA: May 6, 2026                       │
│                                                 │
│  Demand:  STABLE with WATCH items               │
│   · MMF: $7.86T (+$38.7B last week)             │
│   · Foreign total: $9.31T (Jan, +$34.8B)        │
│   · China: $683B (Nov, 17yr low) ⚠              │
│   · Basis trade: notional $1-2T  ⚠              │
│                                                 │
│  Liquidity:  WATCH                              │
│   · ON RRP: ~$6B (buffer exhausted)  ⚠          │
│   · SOFR: 3.75% (at target)                     │
│   · SRF: $0 (no stress usage)                   │
│   · Next test: Apr 15 Tax Day  →                │
│                                                 │
│  Regime:                                        │
│   · Geopolitical premium rising (Iran)          │
│   · Oil at 2022 highs                           │
│   · Fed cut probability declining               │
│                                                 │
│  ACTIVE ALERTS: 2                               │
│   1. ON RRP exhaustion (since 2025-09)          │
│   2. China sub-$700B (since 2025-07)            │
│                                                 │
│  NEXT EVENTS:                                   │
│   → Apr 15: Tax Day liquidity test              │
│   → May 6:  Q2 QRA                              │
│                                                 │
└─────────────────────────────────────────────────┘

This dashboard auto-updates daily before UTC 06:00, pushed via Telegram Bot. When indicators trigger alert thresholds, system generates context-enriched briefs—not "number exceeded" but complete briefs with historical comparison, cross-indicator correlation, and suggested actions.

Summary

The system's core assumption is simple: In an information-overloaded market, edge doesn't come from having more data, but from faster identification of marginal change direction from data. Swarm architecture multi-agent collaboration makes this process automated, traceable, reproducible—while human-in-the-loop ensures final judgment is always made by someone with market intuition.

About Me

I'm Quinn Liu, a FICC PM and Agentic AI enthusiast. I explore integrating multi-agent systems into investment research workflows—the monitoring system described here is a practical example of this exploration. If you're interested in agent-based market analysis frameworks, feel free to reach out.

• Email: quinn@quinnmacro.com
• LinkedIn: linkedin.com/in/liulu-math
• GitHub: github.com/quinnmacro

More research and project details at my website.

Frequently Asked Questions

Why hasn't record Treasury supply triggered a crisis?

Because markets price on marginal supply-demand balance, not absolute supply volume. Treasury adopted "more Bills, maintain Coupons" structural adjustment based on TBAC recommendations, combined with $7.8 trillion money fund demand and Fed MBS reinvestment into Bills, successfully digesting new supply. Meanwhile, foreign investor total holdings reached record $9.3 trillion, though internal composition shifted from official to private, from China to Europe/Canada.

What is Swarm multi-agent architecture?

Swarm is an agent collaboration pattern where agents can dynamically hand off control, automatically switching focus based on market events. Compared to traditional pipeline architecture, it's better suited for market analysis scenarios requiring multi-dimensional real-time collaboration—like an auction event requiring simultaneous assessment of supply, demand, liquidity, and risk dimensions.

What's the biggest risk in Treasury markets today?

Near-term most urgent is April Tax Day liquidity test—ON RRP buffer exhausted (~$6B), SOFR spike will act directly on bank reserves. Medium-term watch three dimensions: (1) basis trade notional $1-2T, leverage 15-20x, potential forced liquidation if dealer intermediation capacity declines; (2) China structural selling to $683B, foreign buyers shifting from "sticky" official to "price-sensitive" private demand; (3) if Middle East conflict persists, ACM term premium path from 80bps toward 100bp+ may be shorter than consensus expects.

How does this framework concretely help PM positioning?

System doesn't give trade recommendations, but does two things: (1) Advance-flag areas where consensus implicit assumptions may be loosening—like current "MMF infinite Bills absorption" assumption, monitorable via ICI weekly data government fund vs prime fund relative growth; (2) Maintain a real-time updated contingency map, so when scenarios start appearing you already know transmission paths and indicators to watch, rather than explaining afterwards.

How to start building a Treasury monitoring system?

Start with five core data sources: (1) Treasury Direct for QRA and auction results (XML feed); (2) TIC database for foreign holdings (monthly CSV, 2-month lag); (3) FRED/NY Fed for ON RRP, SOFR, Fed balance sheet data; (4) ICI for MMF weekly size; (5) CFTC COT report for futures positions (hedge fund basis trade proxy). Then gradually build agent system per this article's five-dimension framework.

Data Sources

Data in this article comes from these authoritative sources:

• Treasury Direct - Official Treasury auction announcements and data
• Quarterly Refunding - Quarterly Refunding announcements, TBAC recommendations and Discussion Charts
• TIC Data - Monthly foreign holdings of US Treasuries
• Federal Reserve FRED - Economic data and Fed balance sheet
• NY Fed Operations - RMPs, ON RRP, SRF operation data
• ICI Money Market Fund Data - Money market fund weekly size
• CFTC Commitments of Traders - Futures market positions report
• PGPF Quarterly Refunding Analysis - Peter G. Peterson Foundation independent analysis