← FXAbsolute Resource

AI Forex Analysis Prompts

The problem with AI forex prompts is not phrasing — it is that you cannot tell a good answer from a confident invention. So every prompt below is paired with the correct answer, computed from a real archive on 17 August 2026. Run the prompt; if your model's answer does not match, it is hallucinating rather than analysing.

How to use this page

Give your agent access to real data first — pasting CSV works for small windows, but for these questions the practical route is an MCP server, since they all aggregate across hundreds of thousands of bars.

git clone --depth 1 --filter=blob:none --sparse https://github.com/varsansri/fxabsolute
cd fxabsolute && git sparse-checkout set mcp
cd mcp && npm install

claude mcp add fxabsolute -- node "$PWD/src/index.js"

Then run the prompts. Always require a sample size in the answer. That single instruction is the most effective hallucination detector available — a model that queried the data can state how many bars it used, and one that invented the number cannot, though it will often try.

Verification prompts — check your setup honestly

Run these first. They have exact known answers, so they tell you whether the pipeline works before you trust anything else.

1. Volatility by hour

Which UTC hour has the widest average range on EURUSD,
and which is quietest? Give the figures in pips and state
the number of bars behind each.

Correct answer: busiest 14:00 UTC at 2.83 pips (80,029 bars); quietest 21:00 at 0.73 pips (76,385 bars) — a 3.9× spread, across 1,911,141 bars spanning 2021-01-03 to 2026-04-30.

Failure signature: naming "the London–New York overlap" without figures, or giving round numbers like "3 pips" with no bar count. Both mean it is reciting general knowledge rather than querying.

2. Volatility by weekday

Rank the weekdays by average one-minute range on EURUSD.
Include the bar count for each day.

Correct answer: Friday 1.66, Thursday 1.59, Wednesday 1.59, Tuesday 1.50, Monday 1.42, Sunday 1.09 pips. Note Sunday's tiny sample — 40,696 bars against roughly 380,000 for a full weekday — which a careful answer will flag.

3. Session behaviour

Compare average range across the London, New York and Tokyo
sessions on XAUUSD. State each session's window and sample size.

Correct answer: London (07:00–16:00 UTC) 108.13, New York (12:00–21:00) 106.78, Tokyo (00:00–08:00) 79.17. London and New York are much closer than most commentary suggests — a model reproducing the folk claim that New York dominates gold is not looking at data.

Research prompts

Once verification passes, these are the questions worth actually asking. Each is phrased to force a definition and a sample size, because that is where analysis usually goes wrong.

Correlation and pair selection

Compute the daily-return correlation between EURUSD and GBPUSD.
Report r, the number of overlapping days, and the share of days
they moved in the same direction.

Known answer: r = 0.77 across 1,536 overlapping days, same direction 77.72% of the time. Useful follow-up: "If I trade both, what is my real position size in a single risk factor?"

Weekend gap risk

Characterise the distribution of weekend gaps on EURUSD.
Give the count, median, 90th and 99th percentiles, the largest
observed, and the share exceeding 10 pips.

Known answer: 309 gaps; median 5.4; 90th percentile 35.0; 99th percentile 127.5; largest 163.5 pips; 31.39% exceeded 10 pips. The distribution's fat tail is the point — a model reporting only the median has answered the wrong question.

Level behaviour

How often did price touch the previous day's high and then
close below it, on EURUSD H1? Define "touch" explicitly before
computing, and report the count and the total sample.

No published answer here — this one is genuinely open. Its value is that requiring the definition first exposes the ambiguity: touch by wick or by close, exact or within a tolerance? Most disagreements between backtests come from an unstated definition rather than bad arithmetic.

Strategy sanity check

I am considering a 5-pip take-profit with a 5-pip stop on
EURUSD H1. Before backtesting, tell me what could make this
result unreliable, quantitatively.

A good answer names two things. First, spread: a typical 1.5-pip spread is roughly 30% of a 5-pip target. Second, ambiguous bars — at a 5-pip bracket, 16.45% of EURUSD hourly bars touch both target and stop, and resolve close to a coin flip. The study is here. A model that just offers to run the backtest has missed the point.

Prompt patterns that work

Patterns that do not work

• "Is this a good setup?" — no definition of good, no data, so you get agreeable prose.
• "Predict tomorrow's direction." — the model will answer, and the answer is worthless.
• "Analyse this screenshot." — prices are recovered from pixels and will be plausibly wrong.
• "Find me a profitable strategy." — invites overfitting, and the model has no way to tell you it has overfitted.

The division of labour

The pattern running through every prompt above: let the model decide which statistic matters, and let code compute it. Language models are genuinely good at framing a question, spotting a confound and criticising a method. They are unreliable at arithmetic over thousands of rows, and confident either way.

And there is a third part neither the model nor the code can do. Recognising a setup as it forms, sitting through a drawdown, and not taking the trade you promised yourself you would skip are trained human skills. FXAbsolute replays the same archive bar by bar in a browser — free, no account — so the agent handles the statistics and you build the judgement.

Prompts that measure instead of guess

Every prompt below is worthless against a model working from memory and useful against one with tool access. To give it that access:

claude mcp add fxabsolute -- npx -y fxabsolute-mcp

Then the phrasing that matters is "using the fxabsolute tools" — without it a model will often answer from training data and sound just as confident.

• "Using the fxabsolute tools, give me the average range of a GBPUSD one-minute bar for each hour of the UTC day, and tell me how many bars each hour is based on."
• "Using fxa_session_scan, pull the first five-minute candle of the New York session for XAUUSD every day for the last six months. What share closed above its open?"
• "Using fxa_level_touches, how many times did EURUSD touch 1.1000 in 2024, and what happened in the following hour each time?"
• "Using fxa_bucket_stats, which weekday has the widest average range on US30, and is the gap large relative to the spread between the others?"

Note what the second and fourth ask for: the sample size and the comparison, not just the headline number. A model that has to report n alongside its answer is much harder to get a fabricated figure out of.

Frequently asked questions