About · how we research
Where every number comes from, and who or what wrote it
The sources behind the terminal, how a market gets its score, the jobs we give to AI and the ones we keep as plain arithmetic.
Where the data comes from
- Bank research
- Notes from leading banks, added through the trading day as they arrive. Each one is broken down into its direction, conviction, argument and levels. The reports themselves stay inside the terminal for Pro members.
- Positioning
- The CFTC Commitments of Traders report each Friday, for positions as of the Tuesday before, and aggregated retail broker positioning through the day.
- Rate expectations
- Exchange-traded short-rate futures, read into the chance of each move at each meeting. Where a central bank has no such market, the panel says so and shows no probability.
- Central banks
- Each of the 8 banks' own decisions, statements and meeting calendar.
- Economic data
- Official releases from statistics offices and central banks, each set against its consensus forecast.
- Prices
- A licensed market data feed.
Where we use AI
- Reading bank research
- A language model reads each report and records the bank's direction, conviction, argument and levels in a fixed format. Each call is shown with the bank it came from.
- Writing the explanations
- The plain-English notes beside the panels, the daily and weekly reports and the morning brief are written by a model from the terminal's own figures, with the figures passed in. It is told never to attach a number the data doesn't hold.
- Sentinel
- Our AI analyst answers questions by reading the board first: scores, positioning, pricing, research and the calendar. It lists the sources it read under each answer.
Where we don't
- The scores
- Arithmetic on published weights. The bank research part counts the calls read from each report, and a model never sets a weight or a score.
- Rate probabilities
- The step the market has priced between meetings, divided by the size of a policy step.
- Positioning and seasonality
- Percentiles, ranges and averages calculated from the published history.
- Missing data
- When an input has no data, it drops out of the score and the others are reweighted. The panel shows what is missing.
How we check our work
- Tested before it counts
- A new input goes into a score only after we have tested it against history. The WTI inventory read added on 17 November 2026, for example, was tested from 1996 to 2023 before it carried any weight.
- Every weight published
- The weights, the contrarian formula and the bias bands are on the scoring model and how it works pages, so you can rebuild any score yourself.
- Every fix dated
- When we find something wrong, the fix goes in the changelog with the date and what it changed, including the figures before and after.
- Assumptions marked
- Where a reading rests on an assumption, as New Zealand's priced path does, the panel marks it.
Questions
Does WatchTower use AI to make its scores?
No. The scores are arithmetic on published weights. AI reads bank research into a fixed format that the score then counts, writes the explanations and reports, and runs Sentinel.
Can I read the bank research on the free pages?
No. The research and each bank's calls are for Pro members inside the terminal. The public pages show how the research is used.
How do you know a new input helps?
We test it against history before it carries any weight, and the changelog records when it went in and what the test found.
Is any of this financial advice?
No. WatchTower describes data and shows the reasoning behind a read. It never gives an entry, a stop or a position size.