Why this lane
Working Paper No. 22 ended with an instruction to stop mining price. Of 129 price-derived candidates tested here, two survived, and both were plumbing rather than knowledge about firms. Perceived cost of capital is the opposite. Managers state an internal hurdle rate on earnings calls; Gormsen and Huber, at the University of Chicago, collect those statements and publish a firm-level panel free at costofcapital.org. It is not a price, it moves slowly, no vendor sells it, and no paper tests it cross-sectionally.
The published file, as downloaded.
| Measure | Value |
|---|---|
| Rows | 264,922 |
| Firms | 9,573 |
| Coverage | Quarterly, 2002 to 2021, fourteen country codes |
| Mean stated cost of capital | 8.72% |
| Mean hurdle rate | 11.36% |
| Mean gap, hurdle less cost of capital | 2.64pp |
| Documented in-sample fit of the Lasso, cost of capital / hurdle | 33% / 16% |
| Variables the Lasso selects | Eleven, led by CAPM beta, a Europe indicator, debt to market, firm age, market equity |
| US firms with at least twelve quarters, used for firm paths | 4,601 |
| Median correlation, a firm's own path against the market-wide path | 0.826 |
| Firms above 0.95 | 17% |
A fitted value, not a measurement
The value columns are named predicted_costcap and predicted_hurdle. On Gormsen and Huber's own documentation they hold not the collected statements but a Lasso fit of those statements onto the 153 firm characteristics of Jensen, Kelly and Pedersen, projected across Compustat, with time variation added through within-firm year effects. They say so; I did not measure it.
| Measurement | Cost of capital | Hurdle rate |
|---|---|---|
| Share of variance that is between firms, not over time | 0.871 | 0.825 |
| R-squared of a firm-plus-year fixed effects fit | 0.890 | 0.788 |
| Residual standard deviation after those fixed effects | 0.47pp | 0.80pp |
| Raw standard deviation | 1.41pp | 1.74pp |
Eighty-seven percent of the panel's variance is a near-static ranking of firms, so what it says about one firm against another in 2002 it is still saying in 2021. Fit a constant per firm and a constant per year and 89% of the variance is gone. Of the tenth that survives, the median firm's path tracks the market at 0.826.
Verdict
The tradable part is spanned by what everyone already holds, a linear combination of beta, leverage, age and size on a slower update schedule than the originals. The part that is not spanned is a market-wide discount rate move, which is not a cross-sectional signal at all. Rejected.
No criticism of the research is implied. Gormsen and Huber built the panel to study corporate investment, and for that a characteristic-projected estimate covering Compustat is right. Working Paper No. 22 found the price lanes empty because everyone mines them; this one is empty for a different reason. Check whether an uncrowded dataset is a measurement or a fitted value before checking anything else. A fitted value cannot carry information its inputs do not have, and no downstream gate catches that.
What this does not establish
| Not claimed | What the record says |
|---|---|
| Spanning, as a measurement | The linear-combination claim comes from Gormsen and Huber's documentation, not from a regression run here. |
| A tested signal | No backtest, no returns, no Sharpe ratio. The variance structure was decisive on its own. |
| A view on the July 2026 method | Gormsen, Huber and Jensen's market-based cost of capital, from machine-learned cash flow forecasts and the present value identity, is a separate object, unread here. It excludes financial firms. |
| Anything current | The panel ends in 2021. |
| A verdict on the hand-collected measure | The statements collected from calls are a measurement rather than a fit, and run to a few thousand firm-quarter observations on the published account. That is thin, and untested here. |
What would change it
Regress the panel on the Jensen-Kelly-Pedersen characteristics and report the R-squared. Near one, spanning goes from documented to measured; materially below one, the residual deserves a real test. The hand-collected observations are the other route, though at a few thousand rows the multiplicity bar would be brutal.
Method. One unauthenticated download from costofcapital.org, analysed locally by deterministic Python in pipeline/coc_panel_audit.py, with output in state/cache/coc_panel_audit.json. The two-way fixed effects R-squared is solved by alternating projections. Firm paths use only firms with at least twelve quarters. Every count and mean is reproduced from the raw rows. No licence is stated and the publisher reserves all rights, so no row of the data appears here.
Educational research only. Not investment advice.