Climate risk as an equity signal
← All projects · Ownership and prices
Companies own physical things. Power stations, pipelines, mines, cement works. Those things sit in places, and places have weather. The question was whether investors price the risk that weather damages those assets or interrupts what they produce.
The hard part was never the question, it was the join. Which company owns which asset, and what that company’s shares did, are recorded in different places under different identifiers, and neither contains the other. So most of the work went into building that bridge and then trying to prove it wrong. The ownership graph is built from Global Energy Monitor’s free tracker. 328 ownership entities resolve to the US and developed-Europe listed universe, 302 of them priceable, and all 328 together hold 5,115 tracked assets.
One return test was specified and set aside before it ran, on a power calculation published in the repository. A second, pre-registered, was run, and its result and the minimum detectable effect that makes it readable are reported in the data-quality note below.
Most of what a vintage difference records is the file being edited. I drew 19 apparent ownership changes at random from the 2,543 that two releases of the same database disagree on, and looked for the transaction behind each one in exchange filings, company statements and press releases. Twelve had none. The six that were real arrived 12 and 71 days late for the two recent and heavily covered deals, and 670 to 1,674 days late for the four that were older or smaller, so the recording error is selective, and a constant offset cannot repair it. Two undocumented choices, one the vendor’s and one mine, each move a verdict I had committed to in writing before any script ran. What that costs a book: on the 191 firms with any coal or gas exposure, the rank correlation between the two releases is 0.73 under one blank-share convention and 0.77 under the other. The note, its code and its provenance records are public.
The data-acquisition layer is public and deliberately independent of any test. It was built to be reusable by whatever the analysis turned out to be, which is why setting that test aside required no change to it at all.