AUDIT · Reproducible records

A forecast record only means something if its history stays intact.

Graham Says records model versions, observations, forecasts and scored outcomes separately. The purpose is to preserve what was known, what was predicted and what happened afterwards.

Each part of the experiment is stored separately.

Model artifactWhat rule was frozen?The model version identifies the forecasting specification used for the prospective run.
Observation recordWhat was known?The pre-open futures signal records the information set available at forecast time.
Forecast recordWhat was predicted?The target-date closing-level forecast is stored before the outcome becomes known.
Result recordWhat happened?The realised close and forecast error are added after the market session finishes.

The order of the record matters.

01Version exists

The forecasting rule is identified before the target result.

02Input recorded

The permitted market information is written to the observation log.

03Forecast locked

The target-date prediction becomes a dated record.

04Outcome arrives

The market closes and the realised index level becomes known.

05Score added

The outcome is joined to the existing forecast rather than replacing it.

Later runs should not silently change earlier evidence.

Write-once and immutability checks are used so an existing dated forecast or scored result cannot simply be replaced by a different value on a later execution. If the research specification changes, the clean response is a new model version or a new prospective record—not revision of history.

This is why the site distinguishes retrospective backtests from genuine prospective forecasts. They are different forms of evidence and should never be merged into one performance history.

Enough information to inspect the experiment without publishing the private data warehouse.

The public site shows forecasts, actual closes, errors, model-race summaries and selected observation fields. Raw market-data operations and credentials remain private.

The record should be inspectable without relying on a screenshot.

The design goal is for a published number to trace back through a target date, recorded input, model version and resulting score. Reproducibility is therefore an engineering property of the workflow rather than a paragraph added to the write-up afterwards.