skills/horizon-forecast/references/research-procedure.md

# Full research procedure

## 1. Frame an answerable decision

Set an as-of date, information cutoff, global/regional scope, and 5-/10-year endpoints. Define economic size using annual producer revenue or value added in constant currency, with its base year. Do not mix these metrics in one numeric comparison. If value added is unavailable, state the revenue proxy and discuss substitution and intermediate inputs. Separate annual flows from installed stock and capital investment.

Define what “explode” means before assigning a probability. The starting convention for a globally scoped segment is at least a doubling of real annual economic size AND at least USD 10 billion of added annual size at the ten-year endpoint. These are editable analysis conventions, not natural laws. Change them for scale and geography before looking at rankings, and disclose the change. An unknown baseline or unsuitable common threshold calls for a different, explicit resolvable event or no numerical probability.

## 2. Scan broadly, then allocate depth

For open-ended discovery, aim for 12–20 sufficiently distinct candidates, with 5–8 finalists when evidence supports them. This is a coverage target, not a quota. Include mature-industry transformations, industrial services, physical infrastructure, and non-US adoption. Include at least one established comparison and one appealing hypothesis that may fail screening. Decompose mega-themes into buyer/product segments rather than ranking “AI” alongside a narrowly defined component.

Map needs and constraints through multiple lenses: demographics, labor, energy/materials, climate/adaptation, health, infrastructure, computation/information, trade/security, institutions, and financing. Document exclusions. Rank research priority by plausible economic significance, uncertainty, evidence access, and likelihood that more information changes the shortlist.

## 3. Build evidence and reference classes

Use [evidence method](evidence-method.md). Search primary records before market-report aggregators: national statistics and customs data; regulator decisions; procurement; audited filings; scientific trials and engineering results; deployment, pricing and utilization records. Supplement with clearly attributed expert analysis and reporting.

For each finalist, seek evidence for the baseline, paid demand, feasibility/cost, adoption/deployment constraints, and strongest downside. Critical facts should have independent corroboration where available. If there is one authoritative original dataset, record that concentration rather than inventing a second source. Multiple articles repeating one estimate count as one lineage. Freshness depends on the variable: a five-year-old physical constraint may remain useful while last year's subsidy or price may be obsolete.

Choose historical analogues by adoption bottleneck, capital intensity, customer type, replacement cycle, regulatory path, and distribution—not narrative resemblance alone. Include failures and delayed successes. Avoid precise priors from tiny, cherry-picked samples. Label subjective priors and broad ranges explicitly.

## 4. Trace causes and adjacent effects

For each finalist, build a compact dependency map: driver → enabling change → adoption → paid output → economic capture. For every important link, record evidence, timescale, uncertainty, and a failure condition.

Inspect first- and second-order effects: a bottleneck relieved can create a new bottleneck; cheaper output can increase usage while reducing revenue; a substitute can transfer spending rather than create new demand; regulations can create markets and constrain them. Include complements and maintenance/compliance needs that persist across technology winners.

Use a tangent register: connection, causal path, direction, horizon, magnitude range, evidence, and disposition (research now / scenario stress / watch / discard). Research now if it could reverse a finalist ranking, move the adopted growth threshold across the horizon, or alter a major buyer/cost constraint. Park unsupported multi-hop speculation. Reopen it if new evidence supplies a meaningful link.

## 5. Reconcile market size and commercial timing

Build a bottom-up estimate (eligible buyers × paid adoption × usage × realized price) and a top-down check (relevant budget, activity statistics, or industry totals). Keep all factors in compatible units and periods. Separate technology potential from serviceable demand and feasible deployed capacity.

Model replacement cycles, utilization, installation rate, distribution, reimbursement/procurement, working capital, and required complementary investment. Stress-test the critical path. Funding, patents, pilots, reservations, and nonbinding announcements are leading signals, not equivalent to recurring sales or delivered capacity.

Do not add parent and child markets or count supplier revenue as entirely new final demand. Distinguish value created, producer revenue, value added, margins, and investable return. Assess regional differences; do not call a US-only estimate global.

## 6. Forecast, score, and challenge

Use scenario and scoring references. Quantitative scenarios need a common definition, explicit inputs and unit basis, and coherent demand/supply constraints. If defensible probabilities are unavailable, provide unweighted exploratory scenarios and say so. Do not assign thirds just to fill a table.

Rate both growth potential and overlooked opportunity, with confidence shown separately. A high-upside low-evidence field belongs in a speculative watchlist, even if its qualitative score is high. Keep score ranges when key components are missing. Examine weight sensitivity, key-input sensitivity, and correlated assumptions across candidates.

Seek an independent challenge, then resolve each finding by verifying, revising, explaining a justified rejection, or retaining it as an unresolved caveat. Preserve initial and revised forecasts and the evidence that caused changes.

## 7. Finish a decision-useful dossier

Lead with the ranking, uncertainty, and implications; include enough field-level depth for a reader to assess each thesis. Show scenario ranges and dates, overlooked subsectors, signposts, failure conditions, and practical research next steps. Explain rejected popular themes. Provide the full source and assumption ledger separately if needed for readability.

Completion gate: plan status reconciled; leading evidence verified; arithmetic reproducible; comparable boundaries; probability/confidence separated; major contradictions handled; ratings explained; source gaps stated; dated forecasts registered. Stop after two targeted follow-up passes add no material evidence to remaining answer-changing questions, or earlier once those questions are sufficiently resolved. This is a stopping convention, not a claim of exhaustive knowledge.