◆ 2026-07-14·Sistema Nacional de Refinación · Mexico
Mexican gasoline production estimates - June / July 2026
Mexican refining just posted the strongest month our instruments have ever measured. That heat has a consequence: fewer American gasoline cargoes. We put a number and a date on it: 399 kb/d of U.S. finished-gasoline exports to Mexico in July.
PEMEX eBDI · EIA MGFEXMX1 · MGFEXMX2 · The Orbit Desk FCC uptime · logged 2026-07-13
7 sites tracked · 5 calibrated ≈72% of SNR gasoline · scored out-of-sample on exports
01 · Mexico ran hot in June
Fluid catalytic crackers make the gasoline. In June the system's crackers ran at levels we have not recorded before: Tula at 100%, Cadereyta at 97%, and Olmeca (Dos Bocas) at 97%, the first sustained run in that unit's history. Passed through per-site transfer functions calibrated against Pemex's own institutional database, June's thermal readings translate to 378 Mbd of gasoline production (±28), against 321 Mbd officially reported for May.
02 · Read it against the season, not last month
More Mexican production and rising imports can be true at once, and usually are in July: it is structurally a heavy import month. Mexican gasoline demand peaks with summer driving, and July cargoes averaged 435 kb/d over 2021-25.
03 · The call
U.S. finished-gasoline exports to Mexico, July 2026: 399 kb/d (±68, 1σ). That is −36 kb/d against July's seasonal norm and −20 kb/d against the counterfactual without June's cracker restart: roughly a 70% probability that July prints below its seasonal norm. EIA publishes the answer around September 30. The call is logged before the data exists; we will publish the verification either way.
04 · Would you have made money listening to us?
We simulated the question the only honest way: an expanding-window backtest in which the model is refit every month on prior data only, then made to forecast the next print. 40 months of one-step-ahead calls, 2023 through 2026, scored against the two benchmarks any desk would use instead.
| Metric | This model | Carry last month | Seasonal naive |
|---|---|---|---|
| Mean absolute error | 42 kb/d | 69 kb/d | 68 kb/d |
| Skill vs benchmark | · | +39% | +38% |
| ±1σ coverage (nominal 68%) | 68% | · | · |
| Directional accuracy, m/m | 69% | · | · |
The production leg tells the same story with a sharper edge: over 41 out-of-sample months, the nowcast of covered-site gasoline output beats persistence by 25% overall, and by 49% in the months where production actually moved. In calm months, last month is a fine forecast and we say so. The instrument earns its keep precisely when something happens, which is when a forecast is worth anything.
05 · Method, briefly
Key units activity, converted to barrels through per-site transfer functions calibrated against refinery-level production in Pemex's institutional database, then through a demand-and-seasonality-controlled displacement model estimated on 112 months of EIA trade data. One disclosure the backtest owes you: the uptime histories were labeled retrospectively; treat it as pseudo-out-of-sample. The forward record starts now: methodology frozen 2026-07-14, every call timestamped in a public verification log before the data exists, misses included.
One flow shown. The same chain, unit heat to production to trade, can be built for any refining system we monitor.
Imagery: modified Copernicus Sentinel-2 data; Landsat courtesy of the U.S. Geological Survey.
The Orbit Desk publishes observational analysis derived from satellite and public data. It is information, not investment advice, and not a solicitation to buy or sell any security, commodity, or financial instrument. We are independent and hold no positions in the assets we observe. Forecasts carry stated uncertainty and may be wrong; any decision taken on them is your own.