Use Forbes 400 data to improve $100M+ AGI bracket synthesis

[MaxGhenis]

Summary

Use Forbes 400 wealth data to improve the $100M+ AGI bracket in the disaggregated PUF. This builds on #606 (aggregate record disaggregation) by adding a public data source for the extreme top tail.

Motivation

The $100M+ AGI bucket (RECID 999999) has 349 returns representing ~$135.7B in AGI. Our current approach generates synthetic AGI from a truncated Pareto distribution, but Forbes 400 provides individual-level wealth data for ~400 households that heavily overlap with this bucket. Using Forbes as a backbone would:

• Ground the AGI distribution in observable (public) data rather than parametric assumptions
• Enable industry-based archetype assignment (tech → LTCG-heavy, finance → mixed, real estate → partnership-heavy)
• Provide a validation target: does our synthetic income distribution produce plausible wealth accumulation?

Approach

Step 1: Wealth-to-income model

Map Forbes net worth to expected income components:

• LTCG: net_worth * realization_rate (calibrate to SOI realization rates by wealth class, ~2-5%)
• Qualified dividends: net_worth * dividend_yield (~1-2% for diversified portfolios)
• Partnership/S-corp: Industry-dependent; real estate and finance allocate more here
• Wages: Known from proxy filings for public company founders; assume minimal otherwise
• Interest: net_worth * risk_free_rate * bond_allocation

Step 2: Industry-to-archetype mapping

Forbes reports primary industry. Map to existing archetypes:

• Technology → LTCG-heavy (founder stock sales)
• Finance/Investments → mixed-financial
• Real estate → partnership-heavy
• Diversified → mixed-financial
• Food/Beverage, Retail → wage-heavy or LTCG-heavy depending on active vs. passive

Step 3: Generate Forbes-derived synthetic records

For the $100M+ bucket:

1. Use Forbes 400 as backbone for top ~349 records (or top N by estimated income)
2. Apply wealth-to-income model to get AGI and component shares
3. Rescale so weighted totals match the aggregate record's known totals
4. Fill secondary variables using archetype donors (same as current approach)

Step 4: Validation

• Compare synthetic AGI distribution against ProPublica tax data (where available)
• Check total Forbes 400 wealth vs. Fed SCF+ top wealth estimates
• Verify income composition shares against SOI Table 1.4 for $10M+ bracket

Data sources

• Forbes 400 list (annual, public): net worth + industry + source of wealth
• Forbes real-time billionaires API (if available)
• SOI realization rates by AGI class
• Fed SCF+ for wealth-income relationship calibration

Challenges

• Wealth ≠ income: Buy-Borrow-Die strategies mean many billionaires report low taxable income
• Forbes estimates have ~20% error bars
• Need to handle unrealized gains carefully (they're wealth but not income until sold)
• Some Forbes 400 members may not file US returns
• Realization rate is the key parameter and varies enormously year-to-year

Depends on

• Improve top-tail income representation in enhanced CPS #606 (aggregate record disaggregation — in progress)

[MaxGhenis]

Update: wealth-to-income modeling from literature

Key parameters (from BSYZ 2025, Fox & Liscow 2025, ProPublica)

Average AGI as % of wealth: 1.7-2.2% for Forbes 400 (BSYZ 2025, matched IRS returns)

But heterogeneity is critical. A single realization rate is inadequate. Proposed mixture distribution for annual AGI/wealth ratio:

Regime	Share of Forbes 400	AGI/wealth	Example
Near-zero	~30%	<0.5%	Founders holding stock, Buy-Borrow-Die
Moderate	~50%	1-3%	Diversified portfolios, regular dividends
Active	~15%	5-15%	PE/hedge fund managers, active traders
Liquidity event	~5%	20%+	IPO, block sale, divorce, estate

Sources for calibrating the mixture:

• BSYZ 2025 (NBER WP 34170): matched Forbes 400 → IRS returns 2010-2020
• Fox & Liscow 2025 (J. Public Econ): SCF panel waves 1995-2016, economic income by wealth class
• ProPublica Secret IRS Files: individual-level tax/wealth ratios for top 25
• IRS SOI Top 400 (through TY 2014): income composition shares

Industry-conditional realization rates (from Forbes industry codes):

• Technology: bimodal — very low (holding) or very high (IPO/sale)
• Finance/Investments: moderate-high (carried interest, management fees)
• Real estate: moderate (partnership K-1 income is regular)
• Diversified/inherited: low-moderate (dividend income)

Key insight from Fox & Liscow: Buy-Borrow-Die is only ~1% of economic income. The main driver of low AGI is simply unrealized appreciation, not borrowing strategies. This simplifies modeling — we mainly need a realization rate model, not a borrowing model.

[MaxGhenis]

Two useful data sources discovered while researching the CA wealth tax:

1. rtb-api (https://github.com/komed3/rtb-api) — MIT-licensed Forbes real-time billionaires data going back to 2020, updated monthly. Structured JSON with name, net worth, country, industry. This is the best free source for the Forbes backbone in Step 1.

2. Rauh et al. replication data (https://github.com/bjaros20/wealth_tax) — MIT-licensed. Contains Raw_Data_Collection.xlsx with 214 CA billionaires including Forbes net worth, city of residence, departure status, and individually researched residential real estate valuations. Also has R scripts fitting a Pareto (α=1.44) to FTB filer counts above AGI thresholds to estimate the income tax distribution among top earners. Their approach to mapping wealth → income tax liability (Section 5.1 of their paper) is relevant to our Step 1 wealth-to-income model.

[MaxGhenis]

Scoping proposal for this Forbes exercise, separate from the aggregate-record rewrite:

1. Treat Forbes as a fixed backbone, not just a calibration prior.

• Create one synthetic tax unit per eligible Forbes record.
• Keep each Forbes record at weight 1.
• Do not reweight the Forbes backbone itself unless we find a clear reason some records should be excluded.

2. Learn wealth-to-income priors from SCF, but anchor taxable composition to PUF/SOI.

• Use SCF to learn wealth -> portfolio mix / income-regime relationships.
• Use PUF/SOI to anchor taxable component composition (wages, LTCG, dividends, interest, partnership/S-corp, etc.).
• Use Forbes industry/source-of-wealth metadata to shift regime priors within the $100M+ bucket.

3. For the $100M+ bucket, subtract the fixed Forbes contribution from the admin targets and calibrate the remainder.

• For each target variable v in the bucket, compute:
  remainder_v = admin_total_v - Forbes_total_v
• For the count target, compute:
  remainder_n = admin_count - n_Forbes
• Then generate/reweight only the non-Forbes residual records to hit the remainder.

4. Add eligibility gates before fixing weights.

• Not every Forbes entry necessarily maps to a US filing tax unit in the target year.
• Need explicit rules for: US return likelihood, household/tax-unit deduplication, and handling missing/uncertain metadata.
• If a Forbes record fails these gates, it should not enter the fixed-weight backbone.

5. Suggested implementation sequence.

• Phase 1: eligibility + canonical Forbes schema (name, family/tax-unit grouping, wealth, industry, source, year)
• Phase 2: SCF-based wealth -> regime model
• Phase 3: PUF/SOI-based taxable income composition model conditional on regime
• Phase 4: fixed Forbes backbone + residual calibration for the remainder
• Phase 5: validation against top-tail SOI targets and internal consistency checks

6. Main risks / open questions.

• Forbes records are households/families, not necessarily 1:1 with filing units.
• Wealth is much smoother than realized taxable income, so the regime model matters more than a single realization-rate formula.
• Need to ensure Forbes totals do not oversubscribe any admin target before residual calibration.

I think this should stay separate from the immediate aggregate-record PRs:

• PR A: land safe top-tail calibration improvements
• PR B: replace aggregate-record synthesis using non-agg PUF donors + selection/calibration
• Forbes: fixed unit-weight backbone plus residual calibration