Methodology

1. Where our data comes from

Market data is not real time. Quotes are pulled on a schedule and may be delayed relative to the exchange. Do not use anything on this site as an execution price.

2. How a story gets chosen

The desk does not write about whatever arrived most recently. Incoming news items are grouped into clusters by subject, and each cluster is scored for significance on a 0–100 scale using signals such as how many independent outlets are carrying the story, how concentrated it is on a single instrument, and how large the associated price move is. The highest-scoring clusters are what get covered.

Three articles are commissioned per publishing day, assigned to the reporter whose beat matches the cluster. If nothing clears the significance threshold, the desk publishes broader thematic analysis instead of manufacturing urgency.

3. How an article is written

Each reporter persona receives a strictly bounded package: the verified market data for the instruments in scope, the scored news cluster, 90 days of price history where we have it, the reporter's own recent articles (for stylistic consistency), and the reporter's currently open predictions so the coverage stays continuous rather than contradicting itself week to week.

The rules attached to that package are the important part. The reporter may only state figures that appear in the supplied data; it may not fill gaps from memory; it must express predictions as a level and a deadline; and prediction levels must fall inside a precomputed band around the current price so that a "call" cannot be a fantasy number.

4. How a draft is validated before publication

No article is published on the strength of the model's first attempt. Each draft is parsed and machine-checked. A draft is rejected and regenerated — with the specific failures fed back to the reporter — when the validator finds:

• Unsupported figures: any price or percentage that cannot be reconciled with the supplied data package.
• Broken arithmetic: inline calculations are recomputed and must hold.
• Out-of-band predictions: a target level outside the permitted range around the current price.
• Contract violations: missing structured fields (headline, meta description, key takeaways, predictions) or malformed prediction records.
• Blacklisted language: a maintained list of hedging and filler phrases that signal the model is padding rather than analysing.

Where a draft fails repeatedly, it is not published.

5. Stale and contradictory data

The most dangerous failure mode for an automated desk is a data feed that keeps returning a confidently wrong number. We guard against it explicitly. Before an instrument can be used in an article, its quote is dropped if:

• the quote is four or more days old;
• a fresher row in our own price history diverges from the quote by more than 1%;
• the intraday range is implausibly dead, which usually means a frozen feed rather than a quiet market.

If every instrument in scope fails those checks, the desk publishes thematic coverage with no hard price claims rather than reporting stale prices as current. This check has caught real provider outages, including a quote endpoint that stayed frozen for days while the underlying time series kept moving.

6. How predictions are recorded and verified

This is the part of our methodology we would most like to be judged on.

1. Recorded at publication. Every prediction is extracted from the article and written to our database with its instrument, direction, target level, deadline and the price at the time of writing. It is timestamped and immutable.
2. Verified on a schedule. A daily job checks predictions whose deadline has passed against market data and marks each one confirmed or missed.
3. Published either way. Outcomes roll up into the accuracy figures on our Track Record page, each symbol hub, and each reporter profile. Misses are shown alongside hits, with the same prominence.
4. Recapped weekly. A weekly article reviews the calls that resolved in the previous seven days, including the ones that did not work.

Accuracy percentages are computed only from verified predictions. Pending calls are excluded, so the denominator cannot be inflated by predictions we have not yet been graded on. Because our reporters publish continuously, these figures move — they are a running record, not a marketing number.

7. Charts

The static chart on an article is generated from our own stored daily closes for the instrument, covering the trailing 90 days, with the article's prediction level drawn on it. It is rendered at publication time so what you see is the picture as it stood when the call was made. Interactive charts elsewhere on the site are TradingView embeds and reflect their data, not ours.

8. How brokers are assessed

Our broker directory is deliberately conservative in what it claims. For each firm we publish the legal entity, the trading brand, the regulator and jurisdiction, the official website, and a contact route. We publish a description of the firm's business where we can substantiate it.

What we do not publish is equally deliberate: no star ratings, no "best broker" rankings, no spread or fee tables, and no scores. We have not conducted standardised first-hand account testing across these firms, and inventing a rating to fill a template would be exactly the kind of unsupported claim the rest of this page exists to prevent. If and when we run genuine comparative testing, we will publish the test protocol before publishing the results.

Listing order is editorial, not paid. See our Editorial & AI Policy for the full independence and conflict-of-interest disclosure.

9. Known limitations

10. Changes to this methodology

When we change how the desk works in a way that affects what you read — new data providers, changed prediction rules, changed verification logic — we update this page and revise the date at the top. If you think something described here is not what the site actually does, please tell us.