20 — The Forecaster Layer

20 — The Forecaster Layer

Authored 2026-09-15 · 200 resolvable forecasts · 25 six-scenario analyses · 150 scenario branches

20.1 What this layer adds

Phase 1's 500 trend records are descriptive: what is happening, how well evidenced, how it scores. Useful, but a trend platform that only describes the present is a well-sourced newspaper.

This layer makes the corpus predictive and accountable. It converts trends into propositions that will be unambiguously right or wrong on a known date, against a named source — and ships the machinery to score them when that date arrives.

The property that matters is resolvability. A forecast that cannot be settled is an opinion wearing a number. Every forecast here names one source, one date, one threshold. For example:

F-16-04 — The Trade Desk reports full-year 2026 revenue growth below 8%. Resolves 2027-02-28 against the company's Q4/FY2026 results release. p = 0.88. Two-thirds of the year has already printed at +3% with Q3 guided down sequentially.

Nobody has to interpret whether that came true.

20.2 How forecasts are derived from trends

The layer is not freehand. Every trend now carries a mechanical prior — a transparent heuristic estimate that it continues in its stated direction over its stated horizon:

base  = 0.35·persistence/5 + 0.20·technical_maturity/5 + 0.25·adoption/5 + 0.20·evidence_factor
prior = 0.30 + 0.55·base + stage_adjustment + direction_adjustment + horizon_adjustment

Mean priors by class behave as they should: current 0.833 · cooling 0.642 · emerging 0.554 · overhyped 0.497. The model has never seen an outcome, so this ordering is a property of the scoring dimensions, not evidence of predictive skill.

The prior's job is to be argued with. Analysts set their own probability and must record prior_delta with a written reason. Across 200 forecasts: mean |delta| 0.246, and not one forecast was left equal to its prior. The largest disagreements are where the layer earns its keep — the prior says continuity, the analyst says this specific thing will not be disclosed:

Four of the five largest adjustments are analysts saying "the data will not exist" — which is what the Phase 1 structural-gap findings should predict, and is a good sign the two layers agree.

20.3 Calibration discipline

A register of nothing but 85% forecasts cannot distinguish a good forecaster from a cautious one. The contract set floors; the register clears them comfortably.

First resolutions fall due 2026-10-31 — six weeks after authoring. That is deliberate: a forecasting layer that produces no track record for three years is unaccountable in practice whatever it promises in principle.

Analysts were also required to write forecasts they expect to lose. Those are marked in-record. A register where nothing resolves against you was a description of the present, not a forecast.

20.4 The register by sector

Eight forecasts per sector, 200 total. Δ is the mean absolute adjustment away from the prior.

Two patterns worth noting. Sector 03 (semiconductors) has the lowest mean probability (0.49) — its analyst is systematically betting against continuity, because memory has cycled four times since 2007 and that base rate dominates. Sector 11 (venture capital) has the highest (0.70) — its structural conditions (DPI drought, concentration) are slow-moving and unlikely to reverse inside the horizon. Neither is a bias; both are base rates doing their job.

20.5 Resolution gaps — 15 forecasts nothing can currently settle

These are the most useful records in the layer. Each names a question that matters and a source that does not exist, which is a direct, costed Phase 2 engineering and licensing backlog.

Several of these are engineering fixes, not licensing spend — CourtListener/RECAP is a free official API that was robots-blocked in this environment, and it gates the entire publishing/music litigation forecast set. Others are genuine voids: no audited public CPM index has existed for any advertising channel in 25 years, and no source consolidates NIS2 enforcement across 27 member states.

20.6 Six-scenario analyses

Twenty-five sets, one per sector, on that sector's most consequential trend. 150 branches, all six types present in every set, probabilities summing to 1.00 in all 25.

Each branch carries: a mechanism (the causal chain, not a mood), early indicators with the source that would show them and the month they would become visible, affected industries, and what a business should do. Each set names exactly one load-bearing assumption — the thing that, if wrong, invalidates the whole set. Naming it is the difference between scenario analysis and scenario theatre.

20.7 The accountability machinery

resolve.py ships with the layer. It is the half that most published forecasting omits.

python3 resolve.py due 2027-03-31     # what needs checking, and exactly where to look
python3 resolve.py resolve F-16-04 1  # record an outcome
python3 resolve.py resolve F-24-01 void  # unresolvable — excluded from scoring, not deleted
python3 resolve.py score              # Brier score, skill vs climatology, calibration buckets

due prints the proposition, the resolution source with its URL and the exact table to read, and the TRUE/FALSE criteria — so resolution requires no memory of what the forecaster meant.

score reports mean Brier score (0.00 perfect, 0.25 the coin-flip baseline), skill against climatology, directional accuracy, calibration by probability bucket (of everything called at 70%, how much happened?), the five worst calls, and — the test that matters for this layer's existence — whether analyst adjustment beat the mechanical prior. If it did not, the analysts are adding noise and the honest response is to say so and simplify.

Run today it prints exactly what it should:

No resolved forecasts yet. The register was authored 2026-09-15; the first resolutions fall due 2026-10-31. Until then this layer has NO track record, and its probabilities are unvalidated.

20.8 Editorial rules for this layer

• Never present a forecast as a certainty, in the proposition, the rationale, or any rendering.
• 89 of 200 forecasts carry review_required — health, financial, legal, political and safety adjacent. They inherit the review flags of the trends they derive from and must clear the human gate in 19-editorial-policy.md before publication. All eight sector-06 forecasts are flagged.
• void is a published state, never a deletion. A forecast that turned out to be unresolvable stays in the register, marked, and is excluded from scoring. Deleting inconvenient forecasts is the standard industry practice this layer exists to refuse.
• A resolved forecast is never edited. Corrections attach as amendments with their own dates.
• Publish the track record before publishing new forecasts, once one exists. A forecaster who shows their Brier score has earned the right to be read; one who does not, has not.

20.9 Honest limitations

1. Zero track record. Everything here is unvalidated. The mechanical prior is a heuristic, not a fitted model — there were no outcomes to fit on. Treat all probabilities as first drafts until the ledger has resolutions.
2. Single-forecaster-per-sector. No aggregation, no independent forecasts of the same proposition, so no wisdom-of-crowds effect and no inter-forecaster reliability estimate. Phase 2 should have at least two analysts forecast an overlapping subset.
3. Resolution-source risk. Sources change methodology (Nielsen's ARF DASH recalibration is a live example, handled explicitly in the sector-14 resolution criteria), stop publishing, or go private. Each is a route to an unresolvable forecast.
4. Selection effects. Forecasts were drawn from the highest-scoring trends, so the register over-represents well-evidenced subjects. That makes the layer look better calibrated than it would be on a random sample of questions.
5. The corpus is the ceiling. Thirteen sectors were researched with degraded source discovery (see the QA register); forecasts derived from them inherit that limitation.

More research from the library · Explore trend records