Live Site | Empirical decomposition of Treasury General Account rebuild funding channels using auction-schedule surprise identification.

When the U.S. Treasury rebuilds its cash balance by issuing securities, where does the money come from? This project identifies 51 rapid rebuild episodes (2005-2026) and uses a bill-size surprise instrument to trace the causal funding channels.

Unexpected bill issuance that rebuilds the TGA is primarily funded by money market funds (+$41B per 1-std shock, t=3.4), with ~75% of that absorption sourced from ON RRP runoff (-$31B, t=-5.0). Bank deposits, reserves, and bank Treasury holdings show no significant causal response once the predictable component of auction supply is removed.

The bill-size surprise eliminates most pre-trends (12 significant placebo coefficients → 1), confirming that the binary event indicator used in prior analysis was contaminated by predictable auction-schedule repositioning.

All values normalized to millions of USD. H.8 series converted from billions. OFR series from raw USD.

Weeks are flagged as rapid rebuilds when the weekly ΔTGA or rolling 4w/8w cumulative change exceeds the 90th percentile of positive changes only. Contiguous flagged weeks (max 9-day gap) are grouped into events. The positive-only threshold is critical — computing quantiles on the full distribution inflates the event count by ~50%.

For each non-CMB bill auction:

surprise = offering_amount − trailing_median(last 8 same-term × reopening-status auctions, min 4)

Aggregated to weekly (Wednesday-ending). This strips ~94% of weekly auction supply that was already announced by the prior Wednesday, isolating unexpected variation in Treasury financing. A tax-receipt surprise control (from DTS deposit categories) absorbs non-issuance TGA movements. A same-term-only grouping is tracked separately as a robustness comparison.

Jordà (2005) specification:

Δy_{t+h} = α + β(h) · shock_t + γ · controls_t + ε_{t+h}

Baseline (binary shock): 2 shock lags, 1 response lag, 11 month dummies. Bill-surprise specification adds a tax-receipt surprise control (from DTS deposit categories) to absorb non-issuance TGA movements. Newey-West HAC standard errors (bandwidth = horizon). Placebo tests at h=-4 to h=-1.

ON RRP abundant (≥$100B) vs. scarce (<$100B). Pre-facility weeks (before 2013, ON RRP = NaN) are excluded from regime estimation, not lumped into "scarce."

* Significant at 5% with Newey-West HAC standard errors.

1. Issuance-specific. The shock identifies unexpected bill supply, not all TGA changes.
2. h=-1 ON RRP. The one remaining pre-trend (t=-2.1) likely reflects same-week announcement effects.
3. Shock persistence. Bill surprise autocorrelation = 0.86. Interpret h=4-8 as primary; h>8 is contaminated.
4. NSA baseline. Month dummies control for seasonality. SA sensitivity confirms: swapping NSA for SA deposits and bank T&A leaves all results unchanged.
5. Accounting overlap. Proxy channels are not mutually exclusive.

Activate the environment you want make to use, install the package, then run the canonical MVP pipeline. make will use the active shell environment, or VIRTUAL_ENV / UV_PROJECT_ENVIRONMENT if those are set.

# Example with uv
uv venv ~/venvs/tgarefill --python 3.11
source ~/venvs/tgarefill/bin/activate
uv pip install -e ".[dev]"

# Canonical MVP path
make mvp

# Extended site build
make site

The MVP command creates the canonical artifacts:

data/processed/master_weekly_panel.parquet
data/processed/event_candidates.parquet
outputs/tables/attribution_baseline.csv

The extended site build extends that with additional tables, figures, and the repo-root site/ bundle used for GitHub Pages:

outputs/
├── figures/
│   ├── tga_timeline_events.png        # TGA with 51 events highlighted
│   ├── irf_binary_vs_bill_surprise.png # Central comparison figure
│   ├── attribution_stacked_top20.png   # Top 20 event decomposition
│   ├── era_dominant_source.png         # Funding channel evolution
│   ├── onrrp_era.png                   # TGA + ON RRP dual panel
│   ├── event_size_over_time.png        # Event scatter
│   ├── irf_pooled.png                  # Binary shock IRFs
│   ├── irf_by_regime.png              # Regime-split IRFs
│   └── irf_continuous_regime.png      # Continuous shock regime IRFs
├── tables/
│   ├── event_candidates.csv
│   ├── attribution_baseline.csv
│   ├── local_projections.csv
│   ├── auction_shock_lp.csv
│   └── auction_shock_grouping_comparison.csv
site/
├── data/                              # GitHub Pages JSON
├── img/                               # GitHub Pages figures
└── index.html                         # GitHub Pages entrypoint

All data from free, official U.S. government sources:

• FiscalData API — DTS, MTS, Auctions
• FRED — H.4.1, H.8, ON RRP
• OFR STFM — MMF, Primary Dealer, Repo
• Treasury.gov — TIC, Investor-Class, Refunding

MIT