Method & limits

How the signals are produced, and what this service deliberately is not.

The cadence

The market is reviewed three times every trading day, at fixed times. Nothing is published in between. The schedule exists to stop the service from reacting to every twitch in the tape, and to make the output comparable from one day to the next.

The watchlist

We cover a deliberately small, liquid universe. Narrow coverage means each instrument gets real attention, and tight spreads mean the simulated results stay closer to what is actually achievable.

• FX majors — EUR/USD, GBP/USD, USD/JPY, AUD/USD, USD/CHF
• Indices — US 500, US Tech 100, Germany 40, Europe 50
• Commodities — gold, silver, WTI crude
• Crypto — BTC/USD, ETH/USD

What each review does

1. Context. Establish the higher-timeframe trend, the current volatility regime, and what the macro calendar holds in the next 24 hours.
2. Levels. Mark the structural levels that matter: prior highs and lows, the previous day's range, round numbers, the levels where the market has already reacted.
3. Setups. Look for a location where price meeting a level would give a specific, falsifiable reason to be long or short.
4. Checklist. Test each candidate against the rules below.
5. Publish. Anything that passes is published in full to every registered user at the same moment. Anything that does not is discarded, and the review is still written up as a market note.

The checklist

An idea is only published if it clears every one of these:

• A defined invalidation level. There must be a specific price at which the idea is simply wrong. No invalidation, no trade.
• Reward-to-risk of at least 1.5 to the first target. Measured from entry to stop against entry to target 1.
• The stop sits behind structure, not at an arbitrary distance chosen to make the ratio look good.
• No first-tier data release inside the horizon of an intraday idea on the affected instrument.
• Not a duplicate. No second open idea expressing the same directional bet on a correlated instrument.
• Liquid hours only. No entries in thin sessions where the spread distorts the result.

Most reviews produce nothing. A day with no signal is a normal outcome, not a failure — it is published as a market note so you can see what was considered.

Risk, expressed as a percentage

Every idea risks a percentage of a notional account between entry and stop — typically 1%, never more than 2%. We never publish a position size in lots, contracts or currency, for two reasons: we do not know your account, and a percentage keeps the record readable at any capital. That is why the track record lets you view the same history against $100, $1,000 or any other figure — the percentages do not change, only the currency amounts do.

Results are recorded in R, multiples of the risk taken. An idea that made twice what it risked is +2R; one that hit its stop is −1R. R is what makes a 30-pip idea and a 300-point idea comparable.

How an idea is closed

Once published, an idea is followed until one of these happens, and the outcome is then permanently recorded:

• Target reached — closed at target 1 or target 2.
• Stop reached — closed at the stop, recorded as −1R.
• Expiry — the horizon runs out with neither level touched; closed at the market price, for whatever it gives.
• Cancellation — the entry was never reached and the setup is no longer valid; recorded at 0R and still shown, so the count of published ideas always reconciles.

Records are append-only. A published idea is never edited, re-priced or removed, and each one carries a visible event log.

How the method learns

A signal service that never revisits its own output is just a random idea generator with a website. Every closed idea therefore goes through a post-mortem before the next review is allowed to use the slot, and the conclusions are published on the what we learned page.

Outcome and process are graded separately

This is the part most track records get wrong. A single trade tells you almost nothing about whether the decision behind it was good: an idea can be built correctly, sized correctly, and still lose, because the outcome of any one trade is dominated by chance. Another can win after a real error — entering late, ignoring a data release — and reward exactly the behaviour that should be corrected.

So each review records two independent things: the result (in R, unchangeable) and a verdict on the process — sound, flawed, or inconclusive. Losses with a sound process are left alone. Wins with a flawed process are flagged just as loudly as losses. Learning from profit and loss alone is how a method teaches itself the noise.

From a mistake to a rule

When a review finds a real process error it is filed against a fixed vocabulary — stop too tight, entry too late, data release overlooked, hidden duplicate, stale data, and so on. A closed vocabulary is what makes the errors countable: the same mistake made three times is visible as a pattern instead of three unrelated anecdotes.

A recurring pattern becomes a rule, written as a single actionable line, and every subsequent review receives the full list of active rules before it is allowed to publish anything. That is the mechanism by which the service actually changes: not by intending to do better, but by carrying a written constraint forward. A rule keeps a counter of how many times its error has occurred, so a mistake seen five times outweighs one seen once.

Learning from the numbers, not just the narrative

Alongside the written reviews, results are broken down by instrument, asset class, direction, review slot, horizon, conviction score and exit type. If conviction-5 ideas quietly underperform conviction-3 ideas, the numbers say so even when every individual write-up sounded convincing.

The honest limits of this

• Small samples lie. Below roughly twenty closed ideas, almost any pattern you can see is noise. The learning page labels the sample size rather than hiding it, and rules are meant to rest on at least a handful of occurrences.
• Adapting can make things worse. A method that rewrites itself after every loss ends up fitted to the recent past instead of the market. Rules are added deliberately and can be archived when the evidence stops supporting them.
• Learning is not a trend. Nothing here guarantees that results improve over time. The process gets more disciplined; the market does not become more predictable. Anyone promising that an AI "gets better every week" at making money is selling something.

Where the AI sits in this

An AI model performs the review and drafts the rationale, working from public market data and the rules above. The checklist is applied mechanically; the model is not free to publish an idea that fails it. What the model does bring is the reading of context — and that reading can be wrong. AI systems misinterpret data, work from stale prices, miss what has not yet shown up in the tape, and can be confidently wrong. The checklist reduces that risk; it does not remove it.

Every idea shows the model that produced it. That is a transparency choice, and it also means you can judge the record against the tool that generated it.

What this method deliberately is not

• Not a backtest. The record starts from the day the service went live and grows forward. There is no historical simulation and no curve fitting, because there is no history to fit to.
• Not a strategy you can outsource. The ideas are a starting point for your own analysis, not instructions.
• Not personalised. Everyone sees the same idea, whatever their account size, experience or situation. See the Recommendation Disclosure.
• Not a promise. The published record is hypothetical, and past results say nothing reliable about future ones.