9 — Forecasting and Scenario Analysis

9 — Forecasting and Scenario Analysis

Phase 1 deliverable · research date 2026-09-15

9.0 The rule this section exists to enforce

Forecasts are never presented as certainties. Not in the database, not in the API response, not in a headline, not in an export. This is not a tone preference; it is a structural constraint implemented in the record schema (§9.3), in the rendering contract (§9.10), and in an append-only scoring ledger that makes a wrong call permanently visible (§9.9).

The distinction the platform is built on comes from §1.3 and is restated here because everything below depends on it:

A forecast without a falsifier is a prediction wearing a hedge. The seed database contains 130 forecast claims out of 2,580 evidence items (5%), against 1,952 fact items (76%). That ratio is deliberate: forecasting is a small, expensive, high-accountability part of the product, not its bulk.

9.1 Why six scenarios rather than three

The conventional bull/base/bear triad has one axis: magnitude. It asks "how much of the base case do you get?" and answers "more, the same, or less." That is a parameter sweep, not a scenario set. It cannot represent the branch that actually matters in most sectors, which is a change in the mechanism rather than in the quantity.

Phase 1 research produced repeated, dated evidence that the consequential branches in 2026 were mechanism changes that a three-scenario model has no slot for:

• The Supreme Court voided the IEEPA tariffs and the tariff architecture was rebuilt on sectoral authorities. Not a bigger or smaller tariff — a different legal instrument.
• The EU AI Act's high-risk employment obligations were deferred from 2026-08-02 to 2027-12-02. The most-marketed compliance deadline of the year simply did not bind.
• The Strait of Hormuz closed, resetting jet fuel to 31.4% of airline operating cost from 25.4% and invalidating the entire February 2026 container-shipping bear case by event rather than by modelling error.
• EPA rescinded vehicle GHG standards effective 2026-04-20, destroying a regulatory-credit profit pool that cross-subsidised several EV businesses.

None of those is "the downside case." Each is a different causal model. Six scenarios split the space along two axes instead of one:

• Base, upside and downside share a mechanism and differ in parameter values: adoption rate, price, build rate, conversion. They are a sensitivity analysis and should be written as one.
• Disruption, regulatory and failure each replace part of the mechanism: a substitute technology or business model arrives (disruption), rule-making changes what is permitted or economic (regulatory), or a constraint or shock breaks the thesis outright (failure).

Six is the minimum that separates those two classes while keeping each branch distinct enough to be falsified by a different observable. Fewer collapses mechanism into magnitude. More produces branches that cannot be told apart by any indicator we could actually poll, which is the failure mode of nine- and twelve-scenario corporate planning exercises.

How the six relate

1. Base is the modal path, not the average path. It is the single most likely coherent world, not the mean of the others. Writing base as an average produces a world nobody believes in.
2. Upside and downside are the base mechanism run at different parameter values. If you cannot express upside as "the base case, with parameter P at value V instead of W," it is not the upside — it is a disruption case that has been mislabelled.
3. Disruption, regulatory and failure are not ordered by severity. Disruption can be commercially positive for the sector and catastrophic for the incumbent; regulatory can run in either direction; failure is the one branch defined by the thesis being wrong rather than by the world being different.
4. The six must partition the outcome space. They overlap in reality — a regulatory change can trigger a failure. The partition is enforced by a precedence rule stated in the record: assign an outcome to the branch whose distinguishing indicator fires first. That rule is written down at forecast time, not after.
5. Midpoints of the six probability bands must sum to 100 ± 10. A set that sums to 150 is not a scenario set, it is a list of worries.

9.2 Required fields for every scenario

Each of the six scenarios in a six-scenario block carries the same record. A scenario missing any required field is invalid and does not render.

What belongs in each scenario type

Base — the modal world. Contains: the continuation of currently-observable rates, with the friction that is already documented. Must contain at least one decelerating element; a base case in which everything compounds is an upside case mislabelled. Must contain the indicator that confirms you are on this branch rather than merely failing to observe the others. Must not contain: a new actor, a new technology, or a rule change.

Upside — the same mechanism, unblocked. Contains: the named constraint that releases, who releases it, and by when. Upside is earned by a specific blockage clearing, not by optimism. Must name the constraint in the preconditions. Must not contain: demand appearing from nowhere. If the upside requires new demand, state where it comes from and who pays.

Downside — the same mechanism, throttled. Contains: the parameter that deteriorates, the rate, and the first place it shows up in public data. Downside is not failure: the thesis still holds, it delivers less. Must not contain: the word "recession" as a mechanism. Name the transmission channel.

Disruption — a different mechanism wins. Contains: the substitute (technology, business model, entrant, or geography), the cost or capability delta that lets it win, and the incumbent asset that becomes worth less. Must name at least two independent parties pursuing it — one company is a project; two is a sector, a rule taken directly from the sector 05 dossier's enhanced-geothermal branch. Must not contain: a technology with no disclosed deployment and no second adopter.

Regulatory — the rules change the economics. Contains: the specific instrument (docket number, bill, rule, directive, proclamation, court docket), the issuing body, the four-stage position (proposed / enacted / in force / enforced — vendor marketing routinely collapses these), and the date the stage changes. Must state the direction: liberalising and restricting are both regulatory scenarios. Must not contain: "regulators may act." Name the docket.

Failure — the thesis is wrong. Contains: the mechanism by which the trend stops rather than slows, who is left holding the asset, and the recovery period of that asset. Failure scenarios are the ones most often omitted and are therefore mandatory. Must include the historical analogue where one exists, with its date and how it resolved. Must not contain: a probability of zero. If failure is genuinely below 2%, say so and show the base rate that supports it.

9.3 What must be true for a trend to grow

Before any scenario is written, the trend is run through a six-gate precondition ladder. Every gate must be answerable with an observation, not an assertion. A gate that cannot be evidenced is recorded as a gap and caps confidence at medium.

A trend that clears all six is a candidate for a narrow base case. A trend that clears three is a candidate for a wide one. A trend that clears one — capability — is a speculative trend and must be scenario-analysed with judgement_only calibration and bands of 25pp or wider.

9.4 What could stop it

The stopper taxonomy. Every six-scenario block must map its downside, disruption and failure branches to at least one named stopper from this list, and must populate the trend's counter_trends field with the corresponding force.

Measurement revision deserves particular weight. In a platform whose product is calibration, the most common way a forecast is falsified is not that the world changed but that the measurement did. Any forecast built on a series that has been revised by more than 20% in the last two years must say so in assumptions[] and mark that assumption load_bearing: true unless it demonstrably is not.

9.5 Marking which assumptions are uncertain

Every scenario carries an assumption register. The register exists to make the difference between decorative and load-bearing assumptions explicit, because a scenario is only as good as its weakest necessary condition.

"assumptions": [
  {"text": "PJM's January 2027 load forecast does not fall more than 10% below the 2026 vintage",
   "confidence": "medium",
   "load_bearing": true,
   "basis": "S-05-05 — PJM publishes annually in January; two prior vintages available",
   "if_wrong": "Base and downside swap probability mass; the downside indicator has already fired"}
]

Rules:

1. Exactly one assumption per scenario may be marked load_bearing: true. Forcing the choice is the point. If an analyst cannot name the single assumption on which the branch rests, the branch is not understood well enough to publish.
2. confidence uses the same three-value scale as the trend record (high / medium / low) so it is comparable across the database. Seed distribution for trends: 227 high, 224 medium, 49 low.
3. if_wrong is mandatory and must name a consequence, not a feeling. "This would be bad" is rejected; "base and downside swap probability mass" is accepted.
4. An assumption whose basis is another of our own forecasts must say so. Forecast stacking is permitted but must be visible — a two-deep stack halves the effective confidence of the outer forecast and the band widens accordingly.
5. Assumptions inherited from the macro context are named and dated, because they go stale: the 2026-09-16 FOMC decision fell one day after the research date and was unresolved; any scenario that assumed an outcome is invalid on its face.

9.6 Writing indicators that are actually falsifiable

An indicator is falsifiable if a competent stranger, given only the indicator text, could determine on the check date whether it fired — without asking us what we meant. That is the whole test, and most published "leading indicators" fail it.

Six required components. An indicator without all six does not validate:

1. A named publisher and series — resolvable to a registered source_id
2. A specific observable — the field or figure, not the topic
3. A threshold — a number or a discrete event
4. A direction — above, below, crosses, occurs, fails to occur
5. A check date or window — bounded, and set at issue time
6. A resolution statement — what firing means for which branch, pre-committed

Rewrites

Four hard rules

• Never use our own analysis as an indicator. An indicator must be observable in a third-party source registered in §6. Self-referential indicators are how forecasting becomes unfalsifiable.
• Announcements are not indicators; contracted or metered quantities are. From the sector 05 upside branch: "Contracted megawatts, not announcements." An announcement indicator can be satisfied by a press release, which any interested party can produce on demand.
• One company is a project; two is a sector. Any indicator that turns on a single firm's behaviour must either name a second firm or be explicitly scoped as company-specific.
• Divergence beats level. The strongest indicators in the Phase 1 dossiers are divergences between two series that normally move together, because they are much harder to manufacture: "SEMI quarterly equipment billings posting a sequential decline of more than 10% while TSMC's monthly revenue is still growing — the divergence is the tell." Likewise the downside signature for robotics: "IFR unit growth remains positive" while Western vendors restructure — volume up, value down.

9.7 Calibration

Probability bands and their verbal anchors

Verbal probability terms are used inconsistently by everyone, so the mapping is fixed and published. The band is the primary object; the words exist only so the output can be read aloud without being misread.

Rules that bind the analyst:

• Band width is a function of evidence, not of comfort. Minimum width 10pp; minimum 20pp unless the trend's evidence_quality ≥ 4; minimum 25pp when calibration_basis: judgement_only. A narrow band on a thin trend is a validation error, not an editorial choice.
• Midpoints across the six scenarios sum to 100 ± 10.
• A band is fixed at issue time and timestamped. Revisions create a new version; they do not edit the old one (§9.9).
• calibration_basis must be declared. base_rate requires a named historical frequency (memory semiconductors have run four boom-bust cycles since 2007 — that is a base rate). analogue requires a named prior episode and an argument for comparability. judgement_only is permitted and is not a defect; concealing it is.

Why point estimates are usually dishonest

1. A point estimate cannot be scored on a single event. "A 37% chance" and "a 42% chance" make identical claims about one outcome. The apparent precision does no work and cannot be audited.
2. The precision is almost always unearned. Two decimal places on an unmodelled judgement is a claim about the analyst's resolution that no analyst can support. Where the underlying data itself moves 69% in a single revision (NERC's ten-year peak-demand figure), a two-significant-figure probability is theatre.
3. Points anchor readers; bands force them to hold uncertainty. A reader given "45%" remembers 45. A reader given "35–55%, roughly even" remembers that nobody knows.
4. Points let you quietly move. A point can drift 5pp a quarter without ever being wrong. A band that moves has visibly moved, and the move is a scored event.

The one legitimate use of a point is the midpoint for arithmetic — summing the block to 100, or computing a Brier score at resolution. It is computed, stored, and never rendered as the headline.

9.8 Forecast accountability

The standard industry practice is to publish confident forecasts, let the wrong ones expire from the website, and re-publish the same confidence next year. Phase 1 found the artefacts of that practice directly — MSCI's Net-Zero Tracker page still showing data as of 2024-08-31 while being sold as current; "cumulative paid to creators" figures reported at the same rounded $100bn a year apart. The platform's answer is mechanical, not cultural.

9.8.1 Every forecast is a resolvable object

At issue time a forecast record is frozen with: forecast_id, trend_id, scenario_type, probability_band, midpoint, resolution_criteria (the indicator text, verbatim), resolution_date, resolver (the registered source_id that will settle it), issued_at, issued_by, and assumptions[]. Resolution criteria written after the fact are worthless, so the criteria field is immutable from the moment of publication.

9.8.2 Scoring

• Brier score on the midpoint for each resolved binary forecast: (p − o)², where o ∈ {0,1}. Lower is better; 0.25 is the score of a permanent 50%.
• Calibration curve: bucket all resolved forecasts by issued band and plot realised frequency against band. A platform that says "likely" and is right 55% of the time is miscalibrated and should be told so in public.
• Resolution rate: the share of issued forecasts that actually resolved by their stated date. Chronically unresolvable forecasts are a design failure and are reported as such.
• Scoring by segment: by sector, by time_horizon, by evidence_quality, by calibration_basis. The expected and publishable finding is that judgement_only forecasts on evidence_quality ≤ 2 trends score close to chance. If they do, the honest product move is to issue fewer of them, not to hide the score.
• Unresolved is a state, not a gap. Open forecasts are shown as open with their check dates, so the denominator cannot be manipulated by leaving losers unresolved.

9.8.3 Publishing the track record

A permanent, public scoreboard covering every forecast ever issued, resolved and unresolved, sorted by default with the misses first. It carries: the aggregate Brier score, the calibration curve, the resolution rate, and a per-sector breakdown. It is regenerated on resolution, not on demand, and it is linked from every forecast rendering — not from a footer.

9.8.4 How wrong calls are prevented from disappearing

Five mechanisms, all structural:

1. Append-only ledger. Forecast records are immutable. A changed view creates forecast_id v2 with a supersedes pointer; v1 remains addressable at its original URL and continues to be scored. There is no update path that overwrites a band.
2. No delete. The only terminal states are resolved_correct, resolved_incorrect, resolved_ambiguous and withdrawn. withdrawn is a published state, carries a mandatory reason, and still counts against the resolution rate. There is no state that removes a forecast from the denominator.
3. Stable citable identifiers. Every forecast has a permanent URL and is exported in the API with its full history. Third parties can archive the ledger and diff it; the ledger is designed on the assumption that someone will.
4. Resolution by registered source, not by us. The resolver field names the source_id that settles the question. Where the resolver is ambiguous, the forecast resolves resolved_ambiguous and is scored as a miss, not excused. This removes the temptation to litigate the meaning of a forecast after the outcome.
5. The annual calibration report leads with the failures. A fixed annual publication naming the largest misses by Brier contribution, what the load-bearing assumption was, and which stopper from §9.4 actually fired. A year with no named misses is treated as evidence of a reporting failure, not of skill.

Corollary for opinion records. Third-party predictions are recorded with attribution and date precisely so their track record accumulates too: "Welcome to the AGI era" (Greg Brockman, 2026-09-03) sits in the database next to the independent benchmarks that contradicted it. The track record of predictors is itself a dataset, and it is one nobody else maintains.

9.9 The output format enforces the rule

Formatting is where "we don't present forecasts as certainties" either holds or quietly fails. The rendering contract:

1. A scenario cannot render without its probability band and at least one indicator. Missing either, the API returns the record with renderable: false and an error.
2. The six-scenario block is atomic. Single scenarios are not exposed as standalone endpoints and the export format carries all six. Quoting the upside alone requires deliberately discarding five siblings, which is visible in the citation.
3. Every forecast string carries its badge inline, in the same run of text, never in a footnote: [forecast · 35–50% · resolves 2027-01-31 · S-05-05].
4. Grammatical mood is constrained. Forecast text uses conditional or probabilistic construction. "will", "is set to", "is poised to" are blocked in the forecast field at validation. The permitted form is the one from §1.3: "If gas turbine lead times remain above 30 months and interconnection queues do not clear, then X by 2029. The indicator that would falsify this is Y."
5. Headlines may not contain a forecast. A headline may name a trend and its uncertainty; it may not state a future outcome.
6. Every rendered scenario links to the trend's evidence and to the accountability ledger. The link to our own track record sits beside the forecast, not behind it.

Rendered form:

T-05-01 · System-operator load forecasts re-rated upward by an order of magnitude
Scenario 1 of 6 · BASE · "Constrained growth"
[forecast · 30–50% (midpoint 40) · calibration: base_rate · resolves 2027-01-31]

If roughly half of queued large load materialises and equipment lead times hold,
then decade-ahead peak demand grows well below the 224 GW NERC assessment while
retail prices continue rising 3–5% annually.

Falsifier → PJM 2027 load forecast (S-05-05, published January 2027): a second
consecutive near-term trim with a stable long-term rate confirms this branch;
a further upward revision of >10% moves mass to upside.

Load-bearing assumption (medium confidence): queue-completion rates stay near
the 22% Exelon discloses for its own 65 GW pipeline.

Our record on 12-month energy-sector forecasts → [track record]

9.10 Worked example 1 — a high-confidence, well-evidenced trend

T-05-01 — System-operator load forecasts re-rated upward by an order of magnitude industry_id: 05 · classification: current · stage: growing · confidence: high · verification_status: triangulated · evidence_quality: 5 · source_diversity: 5 · composite_score: 90.2 (uncapped) · last_verified: 2026-09-15

Why this trend can carry a narrow forecast. Seven evidence items across seven source organisations, four of them Tier A primaries (PJM Inside Lines, the 2026 PJM Load Forecast Report, ERCOT, IEA). The core facts are not in dispute: NERC's 2026 LTRA projects a 224 GW ten-year summer peak increase, 69% above the prior year's 132 GW; PJM forecasts 3.6% annual summer peak growth against 0.3% in its 2021 forecast; ERCOT's preliminary 2026–2032 forecast reaches 367,790 MW against an all-time actual peak of 85,508 MW. A recorded contradiction (NERC vs Grid Strategies, 2026-07-06) is carried rather than resolved, and ERCOT's own characterisation of its number as "a preliminary snapshot" it expects "to be higher than expected future load growth" is recorded as opinion.

Precondition ladder: capability ✓ (load is metered), economics ✓ (utility capex is rate-based), supply ✗ (turbines sold out through 2030, T-05-03), demand ✓ but speculative at the margin (Exelon: 22% of its 65 GW pipeline likely to materialise), permission ✗ (interconnection is the binding gate), capital ✓ (~$1.4T of announced US IOU capex through 2030, EEI 2026-05-27). Four of six clear; the two that fail are the two that generate the branch structure.

Midpoints: 40 + 12 + 20 + 8 + 15 + 5 = 100. ✓

9.11 Worked example 2 — a low-evidence, speculative trend

T-05-15 — Fusion financed as infrastructure rather than research industry_id: 05 · classification: emerging_signal · stage: speculative · confidence: low · verification_status: single_source · evidence_quality: 2 · source_diversity: 2 · composite_score: 42.3 against an evidence cap of 64.0 · time_horizon: long_7y_plus · last_verified: 2026-09-15

Why the same format must produce visibly weaker output. Five evidence items, of which four trace to a single Tier B article (Utility Dive, 2026-08-24); the only Tier A source (EIA STEO, S-05-01) supports the negative claim that fusion contributes zero to US generation. adoption is scored 0 and revenue 0 because no fusion device anywhere has delivered a megawatt-hour. What is documented is real and specific: Commonwealth Fusion Systems raised $1bn on 2026-07-30 taking total capital to $4bn, with ENI and Google named; SPARC is reported about 80% complete with net energy gain targeted for 2027; ARC is planned as a 400 MW plant at Chesterfield County, Virginia on Dominion land for the early 2030s, with Google and ENI contracted for clean-energy credits and output selling into PJM. The trend is that fusion has acquired a named grid interconnection point and contracted offtake — infrastructure attributes, not physics attributes.

Precondition ladder: capability ✗ (net energy gain unachieved), economics ✗ (no cost curve exists), supply ~ (HTS magnet supply is a real constraint, unquantified here), demand ~ (offtake contracted on paper for the 2030s), permission ~ (NRC has indicated byproduct-materials regulation — a material advantage, untested at scale), capital ✓ ($4bn raised). One gate clears cleanly. Under §9.3 that mandates judgement_only calibration and bands ≥25pp on every branch, and it is why the block below is far less useful than the one above — which is the honest result, not a defect in the format.

Midpoints: 48 + 12 + 22 + 7 + 6 + 5 = 100. ✓

Indicators — deliberately few, because few exist. Each is checkable by a third party, which is the one respect in which this block is as strong as the first:

What must NOT be said about this trend, and would be blocked at validation:

• Any band narrower than 25pp on any branch.
• Any statement that fusion "will" be on the grid in the 2030s.
• Treating the Google and ENI offtake contracts as demand evidence. They are contracts for a product that has never been produced; customer_adoption remains "not yet measurable — zero megawatt-hours delivered by any fusion device anywhere."
• Treating the $4bn raised as validation. Capital scores 4; adoption scores 0; the gap between those two numbers is the finding.
• Any use of the four Utility Dive–derived claims as independent corroboration of each other. Four claims from one article is one source.

The contrast, stated plainly

Both blocks have the same shape. Only one of them is worth acting on, and the format makes which one visible without the reader having to take our word for it. That is the entire purpose of running the same six-scenario template across both.

9.12 Prohibited forms

Blocked at validation, not at review:

1. A scenario without a band, or with a point estimate in the band field.
2. An indicator without a registered source_id.
3. An indicator whose threshold is an announcement, a partnership, or a "growing number."
4. A six-scenario block whose midpoints sum outside 100 ± 10.
5. A band narrower than the floor set by the trend's evidence_quality and calibration_basis.
6. will, is set to, is poised to, inevitable, no doubt in any forecast field.
7. A failure scenario with probability 0, or omitted.
8. More than one load_bearing: true assumption per scenario, or none.
9. Editing a published band in place rather than superseding it.
10. Any forecast rendered without a link to the accountability ledger.

More research from the library · Explore trend records