Prior work

Where the methods, shipped numbers and constraints came from

The rest of this documentation describes postforecast on its own terms. This page records where its methods, numbers and constraints came from. The academic background is on the literature page.

The buy-side consensus study

postforecast was extracted from the buy-side consensus study, an Exabel research repository that compared alternative-data revenue forecasts with sell-side consensus.

• Anchored estimate: Ported from docs/anchoring.qmd and docs/anchoring_signal.py, from one company-quarter of scalars to the panel API. The worked example, Zscaler’s quarter ending January 2026, and its self-check numbers are a golden test here.
• Shipped coefficients: ANCHORING_2025_01 (formerly STUDY_2025_01) and CORE_2025_01 were fitted in the study on reports before 2025-01. The derivation is in its docs/anchoring.qmd.
• Reading caps: The 21-day model and 60-day consensus age limits are the study’s research decisions, which is why they are arguments here.
• Bayesian posterior: BayesianPosterior is ported from docs/bayesian_signal.py and docs/bayesian.qmd. That tutorial follows the production signal described below, and its self-check is a golden test here.
• Learnings: Every item in the learnings was found while building and rearchitecting the study. The as-of join that changed 580 of 586 daily returns, the winsorisation look-ahead and the degenerate inputs all broke the study before they were fixed. Its parity check, which reruns the pre-migration scripts at a fixed commit, is the model for how anything ported here is proven.
• Teaching notebooks: The notebooks under notebooks/ are the panel-API ports of the study’s anchoring and point-in-time tutorials.
• Export layout: The study’s BSC_SIGNALS directory has the same layout as the ADC export the model walkthrough reads, so POSTFORECAST_EXPORT can point at it.

Bayesian source implementation

The posterior was ported from an Exabel production calculation. The separate workflow reference records its source snapshot, parity evidence and deliberate differences.

The ADC export analysis

The ADC models repository documents, in docs/export-vintage-analysis.md, that later exports can contain changed historical forecasts after model replacement. The model walkthrough’s caution about historical snapshots rests on that analysis.

Related packages

Read for design ideas, never vendored:

• forecastlens: A diagnostic layer over darts and neuralforecast outputs, with regime-aware calibration, leak-safe backtesting and decision-value evaluation. Its position is closest: it also acts on forecasts produced elsewhere.
• metaforecast: Meta-learning and data-centric forecasting; useful for how it structures families of methods behind one interface.
• statsforecast: The model-list and cross-validation shape worth copying for an evaluation harness.
• AutoGluon: Its long-format conversion API makes series identifiers explicit. The input panel keeps plain pandas frames and validates subject and source keys instead.
• sktime: Its estimator conformance checks test a common interface across implementations. The method contract here is tested for repeatability, non-mutation, row-order independence and empty inputs.
• skforecast: Its backtesting folds make chronological evaluation explicit. Here, future-observation invariance tests enforce the reading cutoff, and rolling evaluation stays with the caller.