Long-range models, built to be accurate early — not accurate late.
A technical overview of how HSAT builds, tests and operates the models behind CropGPT — written for traders and quants who want to audit the method before they act on the output. Five whitepapers are available to download at the bottom of this page.
We ingest every signal that has ever moved a crop — weather, satellite, survey, pricing, narrative — reconcile at field level, and train long-range models that are stable against adaptive bias and validated out-of-sample on twenty years of history. We publish the method. We don't switch it mid-season.
A long-range call that didn't need to move.
In November 2024, market predictions for Thailand's sugarcane crop ranged from 90 to 130 MMT — a 44% spread. HSAT's number, from day one, was ~105 MMT. It did not change. The crop came in at ~105 MMT. Here is what our competitor CZ did over the same window.
SOURCE — HSAT MODEL OUTPUT · COMPANY A & COMPANY B PUBLISHED ESTIMATES · THAI OCSB FINAL.
Being right at harvest is simple addition. Being right a year ahead is something else entirely.
HSAT's value is the stability of the number between the planting decision and the harvest — not the number that lands once everyone can count trucks at the mill.
- One number, held. The same model produced 105 in November, 105 in February, 105 all the way through. Nothing was retuned to land the call.
- The market spread was priced. A 90-130 MMT range is a tradeable dispersion. HSAT's flat 105 was a directional signal against both tails.
- Late accuracy is easy. By the time a competitor converges to the truth, most of the trade-relevant information window has closed. Early accuracy is where edge lives.
- This isn't one year. The same behaviour holds across 20 years of historical back-tests — documented in the in-sample vs out-of-sample paper below.
HSAT publishes one number and holds it. The consistency of a long-range model is what gives you the lower error margin early, when the information is worth something.
ACTUAL OUTCOME ~100 MMT · HSAT HOLDS 105 FROM DAY ONE · COMPETITORS CONVERGE BY NOVEMBER 2025.
What holds a long-range model steady.
Keeping a forecast flat from November to June isn't a property of the model architecture — it is a property of the four decisions below. Each one is documented in a whitepaper at the bottom of the page.
Signals chosen for the next twenty years, not this season.
Our inputs — satellite indices, weather ensembles, surveyor panels, pricing — were selected because they are recoverable, consistent and field-level across the entire twenty-year back-test. We don't chase whatever feed fits last quarter.
Variables earn their place through contribution, not correlation.
Every candidate feature is scored on its marginal, causal contribution to forecast stability across geographies and growing seasons — and pruned aggressively when it does not carry its weight. Large feature sets that work in-sample usually fail out-of-sample.
Accuracy is defined on the worst year, not the average year.
Our headline accuracy is the error band that covers every season in the back-test — the El Niño year, the drought year, the trade-policy year. Averaging tolerates catastrophic misses; we don't.
Out-of-sample or it doesn't exist.
Every model we ship is validated on seasons the model has not seen — a strict walk-forward protocol — with an explicit accounting for adaptive bias: the risk of fitting a model that looks good today because the world has not yet changed.
The method, documented.
Six short papers covering the pillars above. Pick one to read inline, or download the set. Each paper is self-contained and written for a technical audience.
HSAT — Approach to Data.
Why we chose the signals we chose, how we reconcile them at the field level, and what "ground truth" actually means when you collect it in forty-two countries every week.
HSAT — Input Variable Selection.
How we choose variables, the guardrails that stop feature-set bloat, and the discipline that keeps a twenty-year back-test honest when you have thousands of candidate signals.
HSAT — Ensuring Accurate Models.
The definition of "accurate" we hold ourselves to, why a 95% headline average isn't a model guarantee, and how we publish the worst-case error for every commodity we cover.
HSAT — Dealing with Adaptive Bias.
Why models that looked good yesterday can fail tomorrow, how we monitor for regime change in input distributions, and the explicit protocol for retraining when the underlying process shifts.
HSAT — In-sample vs Out-of-sample.
The validation protocol: walk-forward testing, how we hold out seasons, the statistics we publish alongside every forecast, and what we do when a model fails a held-out season.
HSAT — Forecast vs Nowcast.
Why HSAT's forecasts differ from the crop estimates conventional commodity analysts publish, what a nowcast can and cannot tell you, and why the distinction matters months before a harvest begins.
Thailand sugar, written up in full.
The case on this page is summarised. The full write-up — with monthly readouts, sourcing notes and the P&L a trader using the HSAT number would have run — lives in the case studies section.
Related reading.
We will run our method against your toughest commodity and hand you the history.
Give us a crop, a geography and a year. We will publish the forecast our current method would have produced, walk-forward, with the validation statistics we publish on every live model.