Skip to content

long-horizon-forecasting.md

Latest commit

 

History

History
168 lines (110 loc) · 10.7 KB

long-horizon-forecasting.md

File metadata and controls

168 lines (110 loc) · 10.7 KB

ext — Long-Horizon Forecasting (Quarter / Year / Election-Cycle)

Bounded-context module. Loaded by news-quarter-ahead.md, news-year-ahead.md, news-election-cycle.md, and (additively) by news-week-ahead.md + news-month-ahead.md. Adds rules on top of ext/tier-c-aggregation.md. Never replaces the 23-artifact contract from 04-analysis-pipeline.md — it raises the bar.

This module exists because forward-looking analysis at horizons ≥ 90 days has structurally different evidence rules from same-day or week/month-ahead reporting:

  • WEP language must degrade as the horizon lengthens (you cannot say "very likely" about an event 4 years out without ≥ 3 cycle-aged sources).
  • Scenario trees must branch deeper (year-ahead = 4 + 5 wildcards; cycle = 4 × 3 = 12 leaves).
  • Counterfactuals are no longer optional — they are the safeguard against narrative momentum.
  • Cross-horizon citation prevents the analysis from drifting away from its own recent shorter-horizon work.
  • IMF projection-year stamps (T+1, T+2, T+5) become first-class metadata, not a footnote.

The single source of truth for which workflows trigger which rules is analysis/article-types.json (consumed via scripts/horizon-context.ts).

1 — Horizon stratification (every Family-C/D artifact)

Every probabilistic judgement in synthesis-summary.md, intelligence-assessment.md, scenario-analysis.md, risk-assessment.md, threat-analysis.md, forward-indicators.md, and cross-reference-map.md MUST be tagged with one of the canonical horizon bands declared in analysis/article-types.json → horizonBands:

Band Days WEP floor / review expectation
72h 3 very likely / very unlikely permitted
week 7 likely / unlikely permitted
month 30 likely / unlikely permitted
quarter 90 roughly even / about even permitted
year 365 roughly even / about even expected; stronger confidence requires explicit corroboration in analysis (≥ 3 cycle-aged sources)
cycle 1460 roughly even / unlikely / very unlikely expected; avoid likely / very likely unless explicitly corroborated by ≥ 3 cycle-aged sources
election 1460 scenario-driven; coalition-formation outcomes never above "likely"

Operational rule. A judgement at the year band that uses "very likely" without three corroborated cycle-aged sources should trigger a Pass-2 rewrite during author review. The gate at 05-analysis-gate.md does not currently verify corroboration counts; it only checks that WEP terms carry a horizon-band tag.

Tag format (machine-grep-able). Inline marker [horizon:<band>] immediately after the WEP term, e.g.: … is **likely** [horizon:quarter] to pass before recess …. The gate uses grep -n -E '\b(very likely|likely|roughly even|about even|unlikely|very unlikely)\b' artifact.md and asserts every match line carries a [horizon: token within ± 80 characters; this validates horizon tagging proximity only, not source-count corroboration.

2 — Scenario-tree depth

Every long-horizon scenario-analysis.md MUST declare scenarios at the depth specified in the registry's longHorizonRules:

Article type scenarioCount wildcardCount (in wildcards-blackswans.md) Branches per scenario
quarter-ahead 4 0 1
year-ahead 4 5 1
election-cycle 4 5 3 governing-coalition branches per scenario = 12 leaves

Scenario probabilities at every level MUST sum to 100 %. Branch probabilities under each scenario MUST also sum to 100 %. These are authoring and review requirements for the canonical scenario summary table. Current long-horizon gate enforcement does not yet automatically audit ## Scenario / ### Branch header counts or probability-sum arithmetic — those checks are planned follow-up work; until they land, scenario depth and probability-sum compliance are validated only at Pass-2 review. The gate currently enforces the surrounding long-horizon checks that are implemented in 05-analysis-gate.md (horizon-band tags on WEP terms, IMF T+N projection-year stamps, counterfactual paragraph minimums via grep, PESTLE artifact presence for year/cycle, the 24th-artifact requirement for election-cycle, and predecessor-horizon citation in cross-reference-map.md).

3 — Counterfactual reasoning

Long-horizon runs MUST include explicit counterfactual paragraphs in devils-advocate.md:

Article type counterfactualParagraphs minimum
week-ahead 1
month-ahead 1
quarter-ahead 2
year-ahead 2
election-cycle 3

Each counterfactual paragraph is structured as **Counterfactual N — <name>:** followed by a 4–8 sentence narrative (a) restating the assumption being challenged, (b) constructing the alternative world, (c) listing the falsification trigger that would update the analysis. Anchored on a primary-source dok_id or URL.

4 — IMF projection-year stamps

Every WEO / FM / GFS_COFOG citation in long-horizon artifacts MUST carry the projection-year stamp in the inline citation:

[IMF WEO Apr-2026 vintage; SWE NGDP_RPCH 2.1% T+1, 2.4% T+2, 2.0% T+5]

The projectionYear field MUST also appear in every economicProvenance block in machine-readable artifacts (economic-data.json v2.0+).

The gate at 05-analysis-gate.md audits via:

grep -nE '\bIMF (WEO|FM|GFS_COFOG)\b.*\bT\+[0-9]+\b' \
  synthesis-summary.md scenario-analysis.md risk-assessment.md \
  intelligence-assessment.md

Missing projection-year stamps are a Pass-2 rewrite trigger, not a soft note.

5 — PESTLE mandatory thresholds

pestle-analysis.md is normally an analytical-supplementary template (never blocking — see analytical-supplementary-methodology.md). For year-ahead and cycle workflows it becomes blocking:

Article type PESTLE mandatory?
quarter-ahead optional (encouraged at analysis_depth=comprehensive)
year-ahead YES — blocking
election-cycle YES — blocking

Floors (when blocking): ≥ 4 dimension tables, ≥ 3 cross-dimension interactions, line-floor per reference-quality-thresholds.json.

6 — Cross-horizon citation rule

Long-horizon analyses MUST cite shorter-horizon analyses to prevent narrative drift. The registry's longHorizonRules.crossHorizonCitations declares the required predecessors:

Article type Required citations in cross-reference-map.md
quarter-ahead most recent week-ahead AND month-ahead
year-ahead ≥ 2 most recent quarter-ahead + ≥ 4 most recent monthly-review
election-cycle ≥ 2 most recent year-ahead + ≥ 12 most recent monthly-review

Citation format in cross-reference-map.md: each cited predecessor appears in a row with its analysis/daily/YYYY-MM-DD/<subfolder>/synthesis-summary.md GitHub blob URL plus the delta finding that has emerged since (one-sentence note on what changed). The current automated gate verifies presence of at least one configured predecessor path; it does not currently count rows by predecessor type or enforce the numeric monthly-review minima above. Treat the table as the required analysis standard, with per-type counts confirmed during review until the gate is expanded.

7 — PIR roll-forward

Every prior-cycle PIR from the most recent matching horizon folder MUST be ingested into intelligence-assessment.md and tagged in pir-status.json. New PIRs created during this run carry an explicit obsolescenceDate field. Roll-forward is recorded in the supplementary horizon-pir-rollforward.md (never blocking on its own — but if missing, a <!-- horizon-pir-rollforward: skipped — <reason> --> annotation MUST appear in methodology-reflection.md).

scripts/roll-forward-pirs.ts (extended for this module) consumes the most recent same-type predecessor folder and emits the genealogy table.

8 — Forward-indicator horizon bands

forward-indicators.md baseline is ≥ 10 indicators across 4 bands (72 h / week / month / election). Long-horizon overrides:

Article type Horizon bands required Indicator floor
quarter-ahead 5 — week / month / quarter / year / election 12
year-ahead 5 — month / quarter / year / cycle / election 12
election-cycle 4 — quarter / year / cycle / election 15

Each indicator carries: name, threshold, source URL/dok_id, current observed value, watch-trigger condition, owning PIR.

9 — Pass-2 audit additions

Append the following items to the Pass-2 self-audit checklist in methodology-reflection.md:

  • Horizon stratification — every WEP term carries a [horizon:<band>] tag.
  • WEP-band degradation — no very likely/very unlikely at year/cycle bands without ≥ 3 corroborated sources.
  • Scenario tree depth — count matches longHorizonRules.scenarioCount × branches.
  • Counterfactuals — ≥ N explicit **Counterfactual N — <name>** paragraphs in devils-advocate.md.
  • IMF projection-year stamp — every WEO/FM citation includes T+N.
  • PESTLE mandatory — blocking artifact present (year-ahead / cycle).
  • Cross-horizon citation — predecessors counted per registry rule.
  • PIR roll-forward — predecessor PIRs ingested or annotated as skipped.

10 — Determinism + reproducibility

Long-horizon analyses are designed to be trend-aware — they ingest their own predecessors and roll PIRs forward. To keep this auditable:

  • Every long-horizon methodology-reflection.md includes a ## Predecessor manifest block listing the exact predecessor folder paths consumed.
  • Every long-horizon data-download-manifest.md includes a ## IMF vintage pin block with vintage, retrieved_at, payload_sha256.
  • Every long-horizon run that fails to ingest a required predecessor records a <!-- predecessor-ingestion: skipped — <subfolder> not found --> note in methodology-reflection.md AND lowers the analysis-depth label to standard for that run.

11 — Out of scope (for this module)

  • election-cycle cycle-rollover at the 2026-09-13 election — handled in ext/cycle-rollover.md.
  • WEP / Admiralty / ICD 203 base rules — already in 04-analysis-pipeline.md + political-style-guide.md.
  • Tier-C period multipliers — already in ext/tier-c-aggregation.md. This module adds rules on top; it does not duplicate them.