How it works
Five moves, no black box: real market data flows in, a model learns which markets move together, a cone of possible futures comes out, the cone is graded against history it never saw — and your what-ifs bend it along couplings actually measured in the data.
01 · The promise
The dashed centre is the middle of the range — half of outcomes land above it, half below. Nobody trades a single line, so we never sell one.
One finished outcome in two should land inside the inner band. That is a promise with a number on it, so you can count whether we kept it.
Nine in ten should land inside the outer band. Anyone can draw a narrow cone; everything on this page exists to make these counts come true instead.
02 · Data in
Daily settlement prices flow in from real exchanges — energy, grains, metals, rates. No synthetic feeds, no hand-picked histories.
The markets arrive as one panel. Crude and diesel and corn sit side by side, because what one does often matters to another.
For every forecast the model reads a window of recent months across every market at once — level, momentum, and how rough the ride has been.
03 · Joint patterns
Diesel leans on crude. Soy watches corn. Gold mostly ignores both. The couplings a desk pays attention to are real, but few.
The model learns which markets move together from the data itself — a few weighted links, strong where the co-movement is strong, absent where it is not.
When crude jumps, the model lets that jump inform diesel the same day — inside the reading, not after it. That is what "joint patterns" buys.
04 · The cone
For each day ahead the model answers with a full range of levels, each carrying its own share — a level 1 in 10 outcomes finish below, the middle, a level 9 in 10 stay under.
Matching levels bracket a band with a counted share inside. Each day ahead gets its own column, and further out the columns stand taller — less is known, and the picture says so.
Connect each band’s edges across the horizon and the deliverable appears — a cone of possible futures, honest about how fast certainty fades.
05 · The proof
A 30-day forecast is scored 30 days later, on data the model never saw while learning. Nothing still in flight counts.
Every finished outcome either landed inside a band or escaped it. We count the escapes — above and below separately — against the promised share.
The counts are checked against held-out history, side by side with what was promised. When they drift, the cone is fixed before it ships — not explained after.
06 · What-ifs
Think crude ends the month higher than the centre? Pin it there. That is the whole input — a view, not a formula.
Markets coupled to your pin shift their centres by the co-movement actually measured in the data — diesel follows crude a lot, cattle barely at all.
A what-if moves the centre of the story. It does not shrink the uncertainty — the bands ride along and keep their counted shares.