From a forecast to a defensible estimate

Terminology, purpose, and a guided route through the two methods

A model already predicts a company’s revenue. Analysts publish a consensus for that same period. postforecast asks how those readings should be combined, using only information available at a stated moment.

What can be done

Inspect a disagreement: Measure a forecast’s gap from consensus and distinguish a new view from an old, stale observation.
Publish an anchored estimate: Combine that gap with the consensus revision and a customary beat, using a frozen coefficient set.
Publish a Bayesian posterior: Combine historical growth, a model reading, and consensus using their estimated precisions; inspect the resulting interval.
Explain a result: Retain its inputs, reading moment, parameter version, contributions or weights, and eligibility reason.
Decline to publish: Preserve a refusal when required information is missing or a method’s eligibility conditions fail.

The package consumes forecasts. Model training, vendor acquisition, and portfolio backtesting remain in the calling application. The real-data notebook performs its own file reads before passing a data frame to the pure library.

The vocabulary

Describe the data

Time series: Values indexed through time for an identified quantity. Multiple series can be held together using identifying keys and a time index, as described in Forecasting: Principles and Practice.
Estimate panel: This project’s name for the collection of forecasts, consensus, and actuals across entities and periods. The exact phrase is a project convention; it is not presented as an established forecasting term.
Tidy data frame: The representation: a variable per column and an observation per row, following Wickham’s Tidy Data. Here a row holds one source’s value for one subject at one availability time.
Subject: One entity, target, and period, such as Apple revenue for a fiscal period ending in September. Forecast and consensus must refer to that same subject.
Target period: The period being estimated. period is an opaque label; period_end orders periods. A calendar quarter label need not be a company’s fiscal quarter label.
Availability time: known_at, the timestamp supplied by the data source. It answers when a value may enter a calculation, separately from its target period.
Vintage: A historical version of information. Real-time databases such as ALFRED preserve changes to observations across vintages. An export’s model vintage is additional provenance: an old known_at does not prove the selected model existed on that date.

Describe the readings

Forecast: An externally produced estimate, identified by its source name.
Consensus: The aggregate analyst estimate used as the reference reading.
Actual: An observed outcome. The first actual present in an export is not necessarily the original reported value if the export starts late or backfills history.
As-of moment: The explicit cutoff for reading the panel. The default includes observations strictly before it, excluding observations at the cutoff itself.
Point-in-time anchoring: Selecting each value at that cutoff. This is a data operation shared by both method families.
Consensus anchoring: The specific method that adjusts consensus using a gap and revision. This is a different use of the word “anchoring”.

Describe the arithmetic

Gap: Forecast divided by consensus, minus one. A gap of 0.05 means 5% above consensus; it is not yet a prediction of the eventual surprise.
Revision: Current consensus divided by its earlier reading, minus one. The anchored method uses a 30-day lookback by default.
Expected surprise: The anchored method’s predicted fractional deviation from consensus. The anchored level is consensus multiplied by one plus that surprise.
Growth: A level divided by the previous available actual, minus one. These methods use sequential period growth, not automatically year-on-year growth.
Prior: The mean and variance of the company’s realised growth history.
Precision: Inverse uncertainty. Model and consensus precisions use inverse mean squared growth error; a systematic bias therefore counts as error.
Posterior: The combined growth estimate, converted back to a level.
Interval: The independent method’s configured Student-t uncertainty band. Its nominal coverage is a modelling choice, not measured calibration on this example.
Eligibility: Whether a method can publish, together with a reason when it cannot. Missing diagnostics or a refused result must not be silently replaced with zero.

Why the distinction matters

Suppose a forecast for a September period changes in August. Both observations have the same target period, but different availability times. A calculation as of the start of August must use the earlier observation. A September export may contain both observations while still reflecting a model selected in September.

Timestamp discipline protects the calculation; source provenance determines what can be claimed about the history. The walkthrough makes both visible.

flowchart LR
    A[Forecasts and consensus<br>with actual history] --> B[Validate the tidy panel]
    B --> C[Read strictly before<br>the as-of moment]
    C --> D[Consensus anchoring<br>frozen coefficients]
    C --> E[Bayesian posterior<br>historical precisions]
    D --> F[Estimate and explanation<br>or refusal reason]
    E --> F

Follow the model methodology

Improve an existing forecast: Place post-processing after an existing model and compare errors with later outcomes.
Real model walkthrough: Load an ADC export, inspect the timestamps, and compute both methods for one company with explanatory figures.
Anchoring: Study the small worked example, contributions, parameter versions, and refusal rules.
Bayesian posterior: Study the smaller teaching interface and the effect of changing each source’s precision.
Independent Bayesian reference: Follow the complete production-compatible mathematics used by the real-model notebook.

For the input mechanics first, start with the panel and point-in-time reads.