Decision Geometry

Most forecasting systems ask: given what we know about the structure of this market, what is the likely next price? That question requires a correct model of how the market works — something that is difficult to build, hard to validate, and often wrong when markets behave unexpectedly.

We ask a different question: given today's market state, which historical periods were structurally similar — and what followed from those situations? This is a retrieval question. It requires no parametric model of market dynamics. It requires only that history contains patterns worth finding, and that we can represent market states in a way that makes similarity measurable.

The second requirement is where the methodology lives.

A market state is not a price. A price is a single number with no context. A market state is the structure of prices — their distribution across grades and species, their recent trajectory, volume patterns, seasonal position, and environmental conditions.

We encode market states in rolling percentile rank: each feature is expressed as its position within a recent historical window rather than as an absolute value. This removes level and trend, preserving structural patterns that recur across market cycles. A state of “price in the 80th percentile of its recent range, volume declining, late-season” is comparable across different years even when absolute price levels differ by 40%.

The encoded states form a geometric space. Each historical auction day is a point in that space. The current market state is another point. The task is proximity search: which historical points are closest to today?

The K nearest historical states — the analogues — are retrieved from the geometry. Each analogue carries its subsequent trajectory: what prices did in the days and weeks after that historical point.

These trajectories are transported through the current market's price distribution, producing a probabilistic forecast grounded in historical evidence rather than parametric assumptions. The result is not a point prediction. It is a distribution of possible outcomes, weighted by how closely each historical analogue matches the current state.

The analogues are auditable. You can inspect exactly which historical dates were retrieved, how similar they were to today, and what happened after each one. The forecast is not a black box — it is a structured evidence review.

We publish our failure modes alongside our results. A methodology that only shows successes is not validated — it is curated.

Calibration is a design goal: prediction intervals are validated against out-of-sample outcomes before deployment. When the model says 80%, it should be right approximately 80% of the time on data it has never seen.

Scope conditions are documented explicitly. The beef module degrades in post-2021 regimes because rolling percentile encoding is level-agnostic — it cannot distinguish pre-inflation and post-inflation market structure when levels shift but percentile patterns do not. This is a known limitation, documented in our research, and addressed only partially by the methods available. We say so.

Epistemic honesty is not a constraint on the methodology. It is part of it. A validated scope condition tells you more than a suppressed failure.

Decision Geometry is not specific to any one market. The same methodology has been applied across structurally different domains. Each application validates the method independently.