Evidence before a forecast.

Wheats connects wheat-market evidence with its original sources. Public EBM quotes, licensed price history, forecast probabilities and numerical factor contributions are currently unavailable. The standards below define what must be established before a forecast can be published; they are not a claim that a model has passed validation.

Define the contract before the outlook

Euronext Milling Wheat No. 2 futures trades in euros per tonne, with 50 tonnes per contract and a minimum movement of €0.25 per tonne (€12.50 per contract). Delivery months are March, May, September and December; settlement is by physical delivery. Euronext contract specifications.

A forecast must identify the delivery month, expiry, settlement target, issue cutoff and exchange calendar. A nearby continuous series changes contracts. Its roll and adjustment rules must be documented before a change is interpreted as a market return. No delivery-month contract or public forecast roll policy is currently approved here.

1 session. The next exchange trading session after the issue cutoff.

5 sessions. Five exchange trading sessions after the issue cutoff; overlapping forecast targets are not independent samples.

30 sessions. Thirty exchange trading sessions, approximately six weeks rather than one calendar month.

365 days. A 365-calendar-day research horizon; the target contract and non-trading-day convention must be fixed before forecasting.

A historical return is an observation, not a forecast. A direction probability is not an expected price path. The annual horizon needs its own contract or rolling-series target; it cannot assume that today’s nearby contract will still trade a year later.

The option expiry is the research horizon

The terminal’s selected target is the direction of the specified EBM futures contract at the corresponding OBM option’s last trading date. OBM is the option, EBM its futures underlying. A December option can stop trading in November while its underlying futures continues into December.

Every forecast must preserve the same actual underlying delivery month at both endpoints and state its issue-time benchmark. Official daily settlement, last trade and a nearby continuation are not interchangeable. Up, unchanged and down outcomes must be defined before evaluation; direction is distinct from an option’s premium or profitability.

Daily forecasts sharing one expiry have overlapping targets and are not independent cases. Evaluation must group by expiry, purge target-window overlap at split boundaries and keep calibration separate from unseen evaluation. Market-only, macro and news additions must be compared on matched targets and dates.

Option trading closes at 18:30 Europe/Paris; the detailed expiry procedure separately lists indicative reference publication at 19:30 and exercise/abandonment cutoff at 19:45. An actual provider publication/receipt timestamp is still required. See the official calendar and technical specification.

A risk-neutral probability inferred from an authorized option chain is a different measure from a real-world forecast probability. American exercise, bid/ask, actual strikes and stale quotes require explicit treatment. Delta is not an exact real-world probability.

Ask what was known at the time

Publication time, first receipt, usable processing time, revision availability and model generation time describe different events. An input is eligible only when its actual version was available by the issue cutoff. A publication date without a verified time cannot establish intraday availability.

A replayable forecast needs source identifiers, version hashes and an immutable issue record. Later corrections must remain distinct from the original issue. A successful source check does not make an old report new; a newly published edition can contain unchanged source material. When an individual receipt time is not supplied, Wheats labels that gap rather than substituting the edition timestamp.

Keep the economic meaning intact

Production, export availability, stocks, crop quality and policy restrictions are separate quantities. A change in exports is not automatically a change in harvest tonnage. Dated originals take priority for source tracing: EU cereals statistics, JRC MARS, FranceAgriMer and USDA WASDE. These links do not imply that every dataset is connected or licensed for reuse.

Map links locate evidence. Record counts describe coverage and may include reports about the same event. Neither is a price-impact weight. Statistical association, contribution to a specified model and an identified causal effect require different evidence. SHAP’s explanation of this distinction describes why model explanations alone do not establish causality.

Validate each horizon independently

Our publication standard requires time-ordered training, tuning, calibration and unseen evaluation, with every transformation fitted using past data only. Targets whose outcome windows overlap must be separated at split boundaries and accounted for in uncertainty estimates. Random splits and treating overlapping forecasts as independent observations are unsuitable for this purpose. Time-ordered evaluation reference.

Frozen forecasts must be compared on the same dates and targets with predeclared simple baselines. Probability review requires Brier score, log loss, reliability checks, sample sizes and uncertainty; no single score proves calibration or profitability. Model selection, calibration and final evaluation must remain separate. Probability calibration reference.

A release also needs a model version, input cutoff, validation period, limitations and a rule for withholding output when inputs are stale, a contract expires or monitoring fails. Browser refreshes do not create new evidence or justify a new prediction.

Current publication requirements

All of the following remain required for public forecast publication:

Commercial display and derived-data rights. Public quotes, history and derived forecasts require rights for their actual use. Public access or a delay does not establish redistribution permission.

Explicit contract and target. Select a delivery month or approve a documented continuation and roll policy, with an exchange calendar and issue cutoff.

Point-in-time source versions. Record release, first-seen, processing and revision availability before admitting an input to a historical information set.

Independent horizon validation. Compare frozen, time-ordered predictions with simple baselines, check probability calibration and account for overlapping targets.

Approved model and monitoring. Register the model version and limitations, define stale-input abstention and complete review of the actual public product.

Commercial permission must cover the actual use, including any public display or derived output. Transforming data into a probability does not itself establish permission. Euronext’s commercial framework. Wheats currently provides information aggregation; it does not publish trading recommendations, price targets or investment strategies.