Explanation

Why the methods and their boundaries work this way

Position in the workflow

postforecast acts on forecasts already produced elsewhere. Its reference point is sell-side consensus. Model training, acquisition and evaluation remain with the caller. The overview explains the distinction between accuracy and information relative to consensus.

Point-in-time discipline

  • Availability: A period label says what a number describes; known_at says when the number could be used. Read the point-in-time walkthrough.
  • Staleness: Choosing a current observation and explaining a refused estimate need different treatment of stale rows. The staleness discussion explains why age must survive preparation.
  • Training boundaries: A learned parameter must not depend on the row being evaluated. The look-ahead discussion connects this constraint to fit and apply.

Method assumptions

  • Anchored estimate: Consensus is the baseline, moved by a versioned regression on the gap and revision. The anchoring tutorial explains the attenuation argument and the role of consensus revision.
  • Independent Bayesian method: Relative precision determines the contribution of prior, model and consensus. The Bayesian tutorial explains growth space, heavy-tailed intervals and how this relates to anchored estimates.
  • Correlation adjustment: Joint covariance can produce negative coefficients and an intercept. The worked example shows why independent precision weights cannot express the same relationship.
  • Method choice: Consult the comparison and the correlation-adjusted limits alongside the available history and the intended published quantity.

Refusals and auditability

  • Refusals: Missing output is an explicit result, with a reason. The anchoring walkthrough explains why substituting a value or silently dropping a row would lose information.
  • Declared behavior: The independent Bayesian method can omit an unavailable model precision while preserving its reference-defined prediction behavior. The mathematical reference specifies the limits.
  • Versioning: A result must remain attributable to the parameters that produced it. The audit recipe describes what to retain.
  • Design history: Learnings that shaped this package records the failures behind the point-in-time, purity and validation constraints.

The existing notebooks retain their full explanations alongside the worked calculations. These links provide direct routes to that material without duplicating its derivations.