A spreadsheet of historical observations can look complete while hiding the most important detail: when each number became available. For a fundamentals bot, that detail determines whether an experiment is testing a real decision or one made with hindsight.
Three dates, three different meanings
An observation period identifies the month, quarter or trading session a value describes. A publication timestamp identifies when it was released. A vintage identifies which version of the observation was available in a particular snapshot.
These fields should remain separate. A quarterly observation is not automatically usable from the first day of its quarter. Its publication may occur after the period ends, and a later release may change the reported value.
Keep the revision history
Our research design treats a revised observation as a new version rather than silently overwriting an old research snapshot. That makes it possible to reconstruct the inputs available to a model at a given decision time.
FRED’s real-time-period documentation distinguishes information available today from information known at a past point in time. Its vintage-date endpoint identifies dates when new or revised values became available. These are useful starting points for understanding a dataset’s revision history.
Define the decision cutoff
Every experiment needs a decision time. Inputs released after that cutoff belong to a later decision, even if they describe an earlier observation period. The cutoff should also include practical assumptions about ingestion and processing delays.
When a source supplies a date but no dependable intraday timestamp, the experiment should record that limitation. Guessing an exact release time introduces precision the evidence does not support.
Name the prediction target
Predicting an initial economic release and predicting its eventual revised value are different tasks. A research record should identify the exact target, units, horizon and release definition before the model is evaluated.
Our preferred research contract uses the first official release for the primary training label and evaluation target. Later revisions can be studied separately. This is a design principle for the research workflow, not a claim that every source offers complete first-release history.
Make missing history visible
Eighteen years of historical data does not mean every series has eighteen years of usable vintages. Publication histories, licensing, changes in definitions and missing observations can shorten the eligible period for an individual experiment.
Record the available range for each input, document missing periods and distinguish actual observations from any imputation. If the required information is unavailable, reduce the experiment’s coverage rather than filling the gap with later knowledge.
Leave a research record
An experiment should retain its source identifiers, retrieval date, snapshot reference, transformations and decision cutoff. The useful question is whether another researcher can reproduce the inputs and understand why each observation was eligible.
A traceable dataset makes disagreements easier to investigate. It allows the research team to separate a changed model from a changed source, and a revised observation from an original release.
The historical period tells you what a value describes. The publication history tells you whether a model was allowed to know it.
This note describes research principles and development objectives. It is not a report of a validated model or live trading performance.