Inputs are commodities. The verdict isn’t.

Four principles

Everything below follows from four rules we do not bend.

The pipeline

Between a market event and a paragraph you read, there are ten steps. Six are automated checks, one is a human, and the last is an archive that does not let us forget what we said.

1. Collect the signal

Through the session and overnight, we collect and timestamp:

• Macroeconomic releases — official series for inflation, labour, growth, housing, trade and energy, taken from the primary statistical agencies via the Federal Reserve’s FRED database (BLS, BEA, Census, EIA, the Federal Reserve itself).
• Central banks — policy decisions, target ranges, effective rates and the language around them.
• Equities — prices, volumes and moves across our coverage: the S&P 500 and NASDAQ universe in the US, plus large-cap Europe.
• Crypto — prices and flows for the major assets, treated as a separate asset class with its own analytical frame.
• Company news and filings — earnings, guidance, single-name news and insider transactions.
• Calendars — the official release schedule, with exact dates, for what is coming this week.

Collection is deliberately broad. Almost all of it is discarded later; that is the point of collecting it.

2. Fix the clock before anything else

Most published market commentary is wrong about time before it is wrong about anything else. A quote taken at 07:00 in Lisbon carries a change figure from the previous US regular session — describing it as “today” is simply false.

So the first thing every piece of generated research receives is a session context: whether the US market is in pre-market, regular hours, after-hours or closed, computed in New York time against the exchange calendar, including holidays and early closes. From that:

• a change figure is always attributed to the session it actually belongs to;
• before the open, the piece is framed as a pre-market snapshot — no “today’s winners”, no “today’s breadth”, because the session has not happened;
• a scheduled release is described with its real date, never as a vague “if it comes out this week”.

3. Verify the numbers

Macro data passes a validation layer before it can be used in a sentence. It checks plausibility, units and freshness — a figure with an implausible value, the wrong unit, or one that is too stale to describe as current is rejected, and the generation stops rather than proceeding on bad input.

What survives becomes a block of verified facts, and narrative numbers may only come from it. Three specific rules exist because these are the most common errors in market writing:

Policy rates are stated as they exist

The Fed sets a target range, and the effective rate trades within it. We publish the range and the effective rate, never a single invented “the Fed funds rate is X%”.

Periods are never mixed

Month-on-month is never compared with year-on-year, and a monthly figure is never annualised to make it sound larger.

Commodity claims follow the actual price

Statements about oil are conditioned on the real benchmark price at the time of writing, not on a narrative about where it “should” be.

4. Classify honestly

Sector labels are the easiest place to quietly mislead. We use the official GICS classification — eleven sectors — rather than an intuitive grouping. Meta and Alphabet are Communication Services, not Technology. Amazon and Tesla are Consumer Discretionary. Visa and Mastercard are Financials.

When we aggregate performance across our coverage, we call it what it is: an equal-weighted coverage basket. It is not “sector performance”, because our coverage is not the sector, and a handful of names is not an index.

5. Filter and rank

Each surviving event is weighed against comparable historical prints: what equities, the dollar, gold and crypto typically did over the following one day, five days and one month. Events are ranked by expected impact rather than by how loud the headline was.

Market breadth is computed separately from index performance: advances versus declines, and new 52-week highs and lows across the coverage basket. Index ETFs are a measure of the index, not a proxy for breadth, and we never present one as the other.

Most of what was collected is cut here. A brief that mentions everything has told you nothing.

6. Attach the cause

A list of names that moved is data, not research. Every mover carries a candidate explanation drawn from real news, classified with a confidence level, a source and a date — and that classification travels with it all the way to the published sentence.

Where no relevant news is found, the published reason says so plainly. “No specific catalyst identified” is a legitimate output. Inventing a plausible-sounding reason is not.

7. Write the read

Analysis is generated into a fixed structure rather than free prose, so that a piece cannot quietly omit the inconvenient parts. A company read carries a thesis, the evidence for and against it, risks, and a verdict on a five-point scale: Buy · Accumulate · Hold · Reduce · Sell.

Where a price target is published, every one of these is mandatory and none may be blank:

• the target and the implied upside and downside;
• the horizon it applies to;
• the method used to derive it;
• the assumptions it depends on;
• the invalidation level — what would prove the thesis wrong;
• a stated confidence.

Crypto is analysed as its own asset class. Equity valuation metrics — P/E, EPS, dividend yield, beta, discounted cash flow — never appear in a crypto read, because they mean nothing there.

Ticker references produced during generation are validated against the market-data provider before publication. Anything that does not resolve to a real, covered instrument is removed and logged.

8. Human editorial review

This is the step that makes the rest of it accountable. Research is produced with automated systems, and nothing reaches the site without human editorial review: a person reads the output against the evidence behind it, and answers for what appears. Editorial standards, review and corrections sit with one editorial owner, reachable at norwicfinance@gmail.com.

Editorial review is not a formality and it is not a spell-check. Before publication, the editor checks that:

• the conclusion is actually supported by the evidence presented, and the two have not drifted apart;
• the framing matches the market session — nothing is described as happening now that has not happened;
• every number in the narrative traces back to a verified source;
• attributed causes are real and correctly weighted, and weak attributions are labelled as weak;
• the tone stays research-grade: no hype, no urgency, no language that reads as a personal recommendation;
• the risks and the invalidation conditions are stated as prominently as the thesis;
• where the model has produced confident language on thin evidence, the confidence is brought back down.

The editor can amend a piece, send it back, or hold it. Holding is a normal outcome, not a failure — a read we cannot stand behind is not published at all.

9. The publication gate

Every automated publication runs through a validator that sits between generation and the live site. It blocks anything that shows:

• language we have banned as misleading or promotional;
• temporal inconsistency against the session state;
• sector labels outside the official GICS set;
• tickers outside our published coverage;
• empty or missing mandatory fields.

If the gate fails, nothing new is published: the previous valid version stays up and the failure is logged for review. We would rather show you yesterday’s honest read than today’s broken one.

10. Keep the record

Everything we publish is archived with its date and its original wording. Predictions are evaluated against what actually happened, and the results — hits, partials and misses — stay public in the track record and the archive.

We do not delete a bad call. A research record that only contains wins is marketing, and it is the single clearest signal that a publisher should not be trusted.

Where AI fits — and where it does not

We are explicit about this because you deserve to know how what you read was made.

Norwic Finance research is AI-assisted research produced under human editorial supervision. Large language models do the work that scales: reading thousands of releases, filings and headlines overnight, extracting structure, drafting into a fixed schema, and holding far more context than a person can before breakfast.

What the models do not do:

• they do not supply facts from memory — market and macro numbers are injected from verified sources and the model may only use those;
• they do not decide what is publishable — the validator and the editor do;
• they do not set editorial standards, and they cannot override a rule on this page;
• they do not carry the responsibility for what appears. A person does.

We do not claim our research is written by hand, and we do not hide the machinery. Both would be a lie, and the second is the more corrosive one.

What this method cannot do

An honest methodology page has to include its own limits.

• It cannot make anything certain. Verified inputs and a disciplined process reduce error; they do not produce foresight. Markets stay probabilistic.
• It cannot fix the data. Official statistics get revised. Providers have outages. A read built on a figure that is later restated was still the best available read at the time — and it stays in the archive as published.
• It cannot know you. Everything is written for a general audience. It takes no account of your objectives, your horizon or your risk tolerance, and it is not personalised advice — see the Financial Disclaimer.
• It cannot catch every model error. Layered checks catch most, not all. If you find one, tell us at norwicfinance@gmail.com and we will correct it visibly.

---

Informational research content. It is not personalised investment advice. Decisions are the sole responsibility of the user.

← Back to home