What didn't work
Most sites show only the analyses that worked. This page lists the ones that did not, the mistakes we made and the methods that lost. Each entry says what we expected, what happened and where the evidence is. If something is missing, tell us.
We got a headline wrong
4 November 2026
Our new-wells-by-vintage post said Alberta gas wells were starting 19% stronger. That was an artefact of how we sorted wells into oil and gas: the rule put many oil wells that also produce gas into the gas group. We replaced it with a rule based on which product is at least half of a well's output, reran the study and rewrote the post.
Corrected post · Corrections log
Our first explanation for low forecasts only went part of the way
3 to 4 November 2026
Our decline-curve backtest on about 1,900 wells found that the fitted curve runs low. We then checked whether wells that recover after the training window explain it: about 35 to 50% of wells do recover, and a low month just before the forecast multiplies the typical miss by two to five. That does not account for all of the gap, and we have not found the rest.
The backtest · The recovery study
Our seasonal forecast does not beat repeating last month
4 November 2026
We built a forecast that takes the same month last year and scales it by recent year-on-year growth. In the first backtest of the scorecard it was beaten, on average across the six series, by simply repeating last month: roughly 4.7% mean error against 5.1%. Our method won on some series and horizons, not overall. We publish it anyway, and the live log will show whether that holds on data we have not seen yet.
Informational only, not investment, legal or engineering advice. Entries are written by hand in posts/gen_didnt_work.py.