Research framework

From reported event to company exposure

A reproducible hypothesis-generation process for identifying plausible operating effects. The methodology is not yet a measured or validated result. The system ranks research leads; it does not estimate causal effects or forecast stock returns.

Part of a larger financial supermodel project

NordicStockExchange.ai is one research component of a broader financial supermodel project that uses artificial intelligence and large-scale compute clusters. This website develops the event-to-company exposure layer: it structures news, company reports and strategic transmission hypotheses for later integration with wider financial models.

The term “supermodel” describes the intended combined system, not a claim that this website has achieved superior predictive performance. Its outputs remain experimental until they are evaluated against independently labelled data and reported with measured accuracy, calibration and error rates.

Unit of analysis

Each assessment is an event–company pair. The hypothesis is that a reported change in an economic variable reaches a specific company through a named business exposure. Publication requires both a Nordic-relevant event and a defensible transmission mechanism.

1. Observe and represent the event

Collect publisher headlines and supplied summaries, preserve the original source link and text, and translate to English. Extract the changing variable, affected product or activity, actor and geography. A reported allegation, proposal or forecast retains that uncertainty; it is not converted into an established fact. A headline-only source provides less context than a full report.

2. Apply the Nordic relevance test

Require a direct Nordic exposure or a systemic channel: benchmark commodities, financing conditions, EU regulation, major trade routes or cross-border supply chains. General politics, human severity, a shared sector or a shared country do not alone establish relevance. The owner-editable interpretation playbook supplies explicit inclusion, exclusion and evaluation rules.

3. Index companies

Evaluate the companies in the configured active universe. Compare event terms with strategic drivers and expand related economic concepts. A semantic AI pass also examines the company maps for customer, supplier, substitution and second-order channels that keyword overlap may miss. Combine its candidates with the lexical ranking and retain a bounded shortlist for detailed assessment.

The index is defined by the active company registry, not by a permanent company count. The current implementation uses a fixed selected universe; expanding coverage requires updating the registry and its consistency checks.

4. Test the transmission mechanism

For each shortlisted company, require an exact named driver in its exposure map and a nonempty explanation of the event-to-business chain. Consider revenue and cost channels, possible offsets, the direction of the effect and uncertainty. Reject unsupported links. These checks establish internal consistency with an analyst map, not independent verification against a filing or proof of causation.

Heuristic ranking

Strategic impact score

S = M × G × C × (1 − O)

M is the map driver's materiality weight: high = 10, medium = 7, low = 4. G is assessed shock magnitude and C is assessed certainty, each between 0 and 1. O is the offset factor, between 0 and 0.7. Round S to one decimal place.

Publication rule

≥ 5

Require an accepted verdict, medium or high confidence, a stored map reference and an active company. Original English stories and stories translated into English enter the public feed only when they describe an economic event with at least one qualifying company match.

Evidence and inference are different

Reported observation: What the linked publisher reports. Translation does not independently corroborate the report.
Company disclosure: An annual or interim filing linked in the company directory. A link alone does not prove that a particular exposure was extracted or verified.
Strategic inference: The proposed operating effect based on an analyst-built map. Current matches belong to this category unless separately substantiated by an exact disclosure.

The score is an ordinal prioritisation heuristic, not a probability, expected return, revenue estimate or statistically calibrated measure. Confidence labels are model judgements, not confidence intervals. The current threshold means low-materiality drivers cannot qualify on their own.

Traceability and reassessment

The processing records retain original and English text, source URLs, model identifiers, analysis times, map versions, playbook versions, shortlisted candidates, accepted and rejected assessments, and their rationales. This supports inspection and comparison, but model updates and nondeterministic generation mean identical outputs are not guaranteed on rerun.

Validation protocol — not yet a measured result

A scientific evaluation should use a time-separated, independently labelled set of event–company pairs, include unmatched events and hard negative examples, and adjudicate disagreements between reviewers. Measure candidate recall separately from final-match precision and recall; report results by source, language, country and transmission type. Compare against a keyword-only baseline and test sensitivity to the score threshold.

Known edge cases belong in a regression set, not the held-out evaluation set. No held-out accuracy, calibration or causal validation is claimed here. Missing disclosures, source selection, translation errors, stale maps and model judgement can all bias results. Assessments require human review and are not investment advice.