Methodology

Why this isn't another AI summary site

Wrapping a general-purpose language model around a news feed has become trivial. Type a prompt, paste a press release, get a paragraph that reads competently and means little. The wrapper has no view on whether an 8-K is a $50M acquisition or a going-concern warning, or whether a Form 4 is a routine vesting sale or a CEO unloading 10% before a guidance cut.

Wiseek is built the opposite way. The model stack is proprietary — designed and fine-tuned in-house by a data scientist with a finance background, not assembled from off-the-shelf prompts. The trainable layers were tuned on SEC disclosure data. The scoring model was calibrated against historical filing-by-filing market reaction, so a $50M acquisition and a $50B acquisition don't score the same. The sentiment classifier was tuned on the language gap between a beat-and-raise 8-K and a going-concern 8-K — both lawyer-edited, both neutral on the surface, opposite implications. The extraction layer enforces a ticker-aware structured schema so every article is comparable across the corpus.

What Wiseek produces

Every filing or news event Wiseek surfaces carries four pieces of analysis:

• Importance score (1–10) — a single number estimating how market-moving the event is, considering filing type, deal size, parties involved, and historical patterns. Wiseek displays only events scoring 7 or higher to keep the feed signal-dense.
• Sentiment label — positive, negative, or neutral, reflecting the tone of the source content and the likely directional implication for the named ticker. This is a tone label, not a price forecast.
• Plain-English summary — a short rewrite of the headline plus a paragraph explaining what happened, why it matters, and which specific numbers, parties, or terms appear in the underlying filing or report.
• Key events — a structured list of discrete factual claims (e.g., "Director acquired 10,000 shares at $42.10," "Company guides full-year revenue $1.2–1.3B"). Extracted verbatim where possible.

Where the data comes from

Wiseek processes two source streams:

• SEC EDGAR — every public 8-K, 10-K, 10-Q, Form 4, Form 144, DEF 14A, Schedule 13D/A, and related disclosure filed by U.S.-listed companies. Wiseek does not edit the underlying filing; the original document is always linked from each article.
• Licensed financial news — wire-service reporting (Reuters, Dow Jones Newswires, Moneycontrol, and similar) covering ticker-attributable events. Wiseek aggregates and scores headlines; the source publisher retains the underlying reporting and is attributed on every article.

Wiseek's models

Wiseek operates its own financial-event analysis stack — branded as Wiseek AI. The pipeline combines language models running Wiseek's prompt and scoring rubric, classifiers calibrated against historical filing-and-market-reaction data, and structured-output schemas that enforce ticker-aware extraction. The system is domain-scoped to SEC EDGAR + licensed news — it does not summarize the open web, and it does not exist to answer arbitrary questions. Its only job is to score and structure financial-event content the same way for every covered ticker.

Three Wiseek-tuned layers run per item:

• Importance scorer — a 1–10 score produced by Wiseek's scoring model, calibrated on filing type, deal magnitude, insider-participation patterns, and historical market reaction across comparable filings. Items below 7 are dropped from the feed.
• Summary and key-event extractor — Wiseek-tuned passes that produce the plain-English headline, summary paragraph, and verbatim key-event list. Output schemas are enforced so every article is structurally comparable across the corpus.
• Sentiment classifier — calibrated on corporate-disclosure language to distinguish, for example, a beat-and-raise 8-K from a going-concern 8-K when both share neutral phrasing.

What it takes to do this right

Going beyond a single generic LLM call requires per-filing-type domain work. The specifics generic models miss:

• Filing-type baselines. Each form has its own market-impact prior. An 8-K covers material events. A 10-K covers earnings. A Form 4 covers insider trades. A Form 144 covers proposed insider sales. These are not interchangeable, and the rubric weights each accordingly.
• Deal-magnitude weighting. A $10M acquisition and a $10B acquisition arrive in the same form with the same template language. The rubric scales score with dollar amount, share count, and percentage-of-market-cap rather than treating them as equivalent events.
• Insider-participation patterns. A 10% beneficial owner, a CEO, and an independent director each acquiring shares produce different signals. Form 4 scoring weights the filer's role and ownership concentration, not just the transaction count.
• Recency and clustering. Multiple related filings on a single ticker within a short window — for example, a Schedule 13D followed by a Form 4 followed by an 8-K — get score amplification together. That pattern is invisible to single-event summarizers.
• Tone trained on disclosure language. Corporate filings are written in lawyer-edited neutral phrasing. A beat-and-raise 8-K and a going-concern 8-K both read calmly to a general-purpose sentiment model. Wiseek's classifier was tuned specifically on the disclosure-language gap between them.
• Coverage uniformity. Every U.S.-listed ticker filing with SEC EDGAR runs through the same model stack with the same rubric. Wiseek does not operate a separate "premium-tier" pipeline for paying subscribers — paying users get different surface access (real-time alerts, scanner filters, dashboard features), never different model output.

How the scoring works

Importance scores are produced by Wiseek's in-house scoring model under a fixed rubric. The rubric weights factors such as:

• Filing type and historical market-impact baseline (10-K covers earnings; Form 4 covers insider trades; etc.)
• Magnitude — dollar amounts, percentage changes, share counts, deal valuations
• Material-event flags — guidance changes, executive transitions, M&A, going-concern warnings, restatements
• Insider participation and ownership concentration on Form 4 filings
• Recency and clustering — multiple related filings on the same ticker within a short window

Scoring runs autonomously the moment a filing or news item is ingested, typically 30–90 seconds after EDGAR or wire publication. Items are not re-scored once published unless Wiseek pushes a methodology or model revision.

What Wiseek does not claim

• Not financial advice. Importance scores and sentiment labels are descriptive, not prescriptive. Nothing on Wiseek is a recommendation to buy, sell, or hold a security.
• Not real-time at the millisecond level. Wiseek processes filings and news with a typical latency of 30–90 seconds from EDGAR or wire publication. Trading decisions requiring deterministic millisecond latency should use a direct exchange feed.
• Not exhaustive. Wiseek covers U.S.-listed equities and the majority of SEC filing types. Some niche filings, foreign issuers, and OTC tickers are out of scope.
• Not a substitute for the source. Wiseek's summary is an interpretation. Investors making decisions should read the linked source document.

Limitations

Three honest constraints worth knowing:

• Model output can have errors. Summaries and scores are produced autonomously by Wiseek's models. Errors, omissions, and edge-case mis-classifications occur. Report mistakes to contact and we will correct them.
• English is the primary surface. Source filings and reporting are processed in English. Translations into other languages are machine-generated and best-effort.
• Score thresholds are calibrated, not perfect. The 1–10 scale reflects relative impact within Wiseek's universe, not an absolute measure of market reaction. A score of 9 does not guarantee price movement.