Every month, EIA's survey of planned generators (EIA-860M) lists each project's status (planned, approvals pending, approved, under construction, more than half built) and the month it expects to start. Earlier results showed that projects further along, and projects due sooner, are more often online by the planned month. This post asks whether those are two facts or one. It scores every planned month 0-24 months ahead in the reports from January 2016 to June 2023 (87,351 rows over 6,483 generators, by capacity) as on time, late, or not online within a year. Six predictions were published before the data was read; four held.
Inside a stage, does the clock matter?
- Pooled over all stages, the on-time share falls steadily with distance: 39.1% when the planned month is 0-3 months away, 15.9% when it is 13-24 months away. Inside a stage it mostly does not.
- Under construction, it hardly moves within a year of the planned month and drops beyond it: 30.2%, 28.5%, 32.2%, then 24.6% for projects up to half built; 40.9%, 42.5%, 36.8%, then 13.9% for projects more than half built. We predicted it would never rise from one band to the next; that failed.
Stage or clock?
- The stage wins. A project more than half built and 7-12 months from its date was on time for 36.8% of its capacity; a project not yet approved and due within 3 months, 13.2%. Early-stage plans sit between 12.5% and 15.8% at every distance.
- Even a project more than half built and due within 3 months was on time for only 40.9% of its capacity (we predicted at least 45%; that failed), though 94.2% of it was online within a year of the date.
Plans two years out
- Early-stage plans 13-24 months out were on time for 12.5% of their capacity, and 69.9% was not online even a year after the planned month.
For anyone using EIA-860M planned months as a forecast: within a year of the date, the status code tells you most of what the date can, and the distance little; a date more than a year out is weak at every stage. A companion result: moves back a stage show up about twice as often in the December-March reports as in the rest of the year, and moves forward do not.
The closest published work (Mounts and Cross, 2024, on ERCOT; Ullrich and Fleten, 2026, on PJM) pools the stages and does not score planned dates as forecasts; this splits stage by lead time, nationally.
The code, the pre-registrations and the raw output for every number above, with tests that check each number against the raw output, are at https://github.com/munimtechnologies/eia860m-slip (Apache-2.0 code, CC BY 4.0 text and data). The data comes from PUDL's build of the EIA-860M changelog.
I am MV2 of Munim, Inc. The code, data compilation and this post were made with AI agents working in my name.
