Empirical Calibration: Backtesting 2,303 Public Companies across the Russell 3000
Before you trust any valuation model, you should ask how it performs when tested against thousands of real companies out-of-sample. We ran our 12-Step PME valuation engine across 2,303 public companies from the SEC EDGAR bulk archive using leave-one-out validation.
The Leave-One-Out Testing Protocol
To prove that our model does not rely on target leakage or hand-picked comps, we conducted a rigorous leave-one-out out-of-sample backtest:
- For every company in the database, the target firm was completely removed from its own peer reference set.
- Its raw fundamentals (TTM Revenue, Normalized EBITDA, 1y Growth, ROIC, Debt, and Cash) were fed into the 12-Step PME engine.
- The predicted Enterprise Value and calibrated conformal uncertainty intervals ($P_{10}-P_{90}$) were compared against its actual market Enterprise Value.
Global Benchmark Results
95.1% cross-sectional variance explained
median across all 2,303 companies
hits theoretical 80% calibration target
Index Cohort Breakdown
One of our most interesting empirical discoveries is that small caps were actually more predictable from fundamental line items than large caps:
| Index Cohort | Constituents | MedAPE (%) | Log R² | P10–P90 Coverage |
|---|---|---|---|---|
| Russell 2000 (Small Cap) | 1,418 | 24.0% | 0.931 | 81.5% |
| Russell 1000 (Large & Mid Cap) | 772 | 27.5% | 0.915 | 83.7% |
| S&P 500 (Large Cap) | 488 | 29.4% | 0.885 | 84.6% |
| Russell 3000 (Full Market) | 2,190 | 25.1% | 0.949 | 82.2% |
Large-cap leaders often trade at premiums reflecting intangible assets that pure financial statements under-represent (brand equity, distribution network moats). For small and mid-cap operating businesses, the fundamental financial mechanics explain nearly all cross-sectional variance.
Sector-Level Performance
Industrials (19.7% MedAPE), Consumer Goods (20.9% MedAPE), and Software (21.8% MedAPE, $R^2 = 0.98$) all exhibited exceptionally tight fundamental clustering.
Reproduction: