How head to head works

DataSnack's mission is to make forecasts attainable for everyone.

Three ingredients, all public

• Public data: EIA's monthly figures for each price and the daily or weekly prints it publishes, free from eia.gov. They fit the equation and set its starting point.
• Prediction markets: the prediction markets DataSnack follows in the price's domain (shipping straits, supply agreements, record prices), each with its number and DataSnack's seven-day forecast of it, free to read on the markets' own pages and on datasnack.ai.
• The news: the sources of DataSnack's research on the price's model, each linked.

Once a week, and the day before each EIA edition, DataSnack's agents read the markets and the news and may move the equation's path by at most 15% either way, citing what they read. A move that cites no market or news item is refused in code.

The equation

In natural logs of monthly averages: ln ŷ(k) = ln a + β(k) · (μ − ln a). The anchor a is the mean of the last five daily prints known on the day (the last weekly print for pump prices); μ is the mean of the log monthly figure over the last 60 months; β(k) is fitted for each month ahead k by least squares through the origin on up to 30 years of past months. Standing still is a, held flat.

Do it yourself

The workbook for each price (no account) holds the equation as live formulas, every fitted value, the rows the fit used, and the record. The feeds and the workbook reproduce the public-data row exactly; every adjustment lists the markets and news it rests on, with links.

The rules

• DataSnack's number is made the day before. An outside number published on a day is paired with the path DataSnack made on the latest day before it. A number published before DataSnack's first path is late: shown, never counted.
• The actual is EIA's own monthly figure, read from its outlook when it prints.
• Closer is the smaller absolute error. A tie is a difference under half the display step: $0.01 per barrel and per MMBtu, $0.001 per gallon.
• Below 30 graded numbers the boards show counts, never percentages. Numbers from one edition move together, so an edition is closer to one observation than to eighteen.
• Numbers from forecasters other than EIA are filed by hand with the source link and the sentence they were read from; a correction is listed on the page.

Replay, 2014 to 2026

The replay computes public data alone for 154 archived EIA outlooks, January 2014 to October 2026, each anchored on prints dated before EIA's own cut-off and fitted only on months before it. Prediction markets and the news as they stood on past days cannot be replayed honestly, so the replay has no DataSnack row. It is labelled replay everywhere and never added to the live record.

So far: public data alone closer on 6307 of 12270 graded numbers, EIA on 5913, 50 ties.

What may be said from it: over twelve years a one-line equation fitted on public data was about as close as EIA's outlook beyond three months, and EIA was closer for the near months. Nothing stronger.

Canonical: https://datasnack.ai/versus/method · DataSnack forecasts are analysis, not investment advice.