Skip to content

How KAI works

What KAI looks at, and what it never does

KAI is Northmark’s project to have an AI (a large language model) make FX trading decisions and to test its learning and reasoning in public. Every closed trade since tracking began on November 1, 2026, from the one account we record, is published, losses included.

It is not a trading instruction or recommendation, and it does not promise future results.

Current model
Claude Haiku 5.5 (since October 13, 2026)
Process records start
August 27, 2026
Tracking since
November 1, 2026
Last checked
November 3, 2026

KAI in one trade

Narrated (same words as the captions). It starts without sound.

A short film of one trade: how KAI sets its stop-loss before anything else, and how it records the result.

How a decision is made

Fixed rules are checked by code before and after the AI decides. If any check fails, KAI does not trade.

  1. 01

    Find candidates

    Every 30 minutes, KAI picks from the 28 pairs of the 8 major currencies the ones whose trading cost is low compared with how much price moves.

  2. 02

    Check when not to trade

    Weekends, the first hours after the weekly open and the time around major releases are skipped here.

  3. 03

    Read the market

    Code that applies the same rules to rises and falls summarizes charts on four timeframes. No AI is used here.

  4. 04

    The AI decides

    The AI writes the direction, stop-loss, take-profit, time limit and invalidation price. It can also skip.

  5. 05

    Challenge

    An AI in a separate role argues against the decision, and its view is recorded.

  6. 06

    2 of 3 agree

    The AI is asked 3 times about the same moment. Unless at least 2 answers agree on direction, KAI does not trade.

  7. 07

    Final check

    Cost and price drift are measured again right before the order. If they are out of range, KAI skips.

  8. 08

    Place the order

    The order goes in with a stop-loss and a take-profit. Open trades are watched on every 5-minute check.

What KAI looks at

  • Charts on four timeframes: 15-minute, 1-hour, 4-hour and daily
  • Trading cost (the spread) compared with the average range of price moves (ATR)
  • The schedule of economic releases and central bank announcements
  • Its open trades and how many it has opened that day

KAI is not judged on backtests. The AI may already know past prices, so it is measured only on trades recorded going forward.

What KAI never does

  • No new trades at weekends, in the first 120 minutes after the weekly open, from 30 minutes before to 60 minutes after 17:00 New York time, or after noon New York time on Fridays.
  • No trades around US jobs data, US inflation, FOMC and other central bank decisions, or flash inflation, jobs and GDP figures for major economies. No trade may stay open across such a release. If the schedule can’t be read, KAI does not trade.
  • No trade when the cost is high compared with price movement (spread ÷ 1-hour ATR above a line measured for each account, roughly 0.10–0.20).
  • At most 4 new trades a day, 3 in 6 hours and 1 per check. At most 2 trades with the same currency, and 1 trade selling yen at a time.
  • The same pair is not judged again until 240 minutes have passed.
  • It never adds to a losing trade.
  • It does not close trades in parts or move the take-profit once the trade is open.

Stop-loss rules

A high win rate can still lose money. What decides whether pips are left over is the size of each loss, so KAI fixes that first.

  • Every trade has a stop-loss, placed together with the order.
  • The AI sets the stop-loss at 0.8–3.0× the average range of price moves (ATR). Code enforces that range, and the AI cannot override it.
  • Trade size is set so that a stop-out costs about 0.33% of the account. Stops across open trades are capped at 2%, and losses in a day at 3%.
  • A stop-loss is never moved further away after the order. It can be moved closer; such moves are included in the notifications, though a notice may not always arrive.
  • Once open profit equals one stop-loss distance, the stop moves to the entry price.

A stop-loss limits losses in normal conditions, but it is not a promise. In fast markets or when prices gap at the weekly open, a trade can close at a worse price than planned.

Entries and exits

  • The take-profit is at most 3.0× the average range of price moves.
  • A trade is held for at most 96 hours.
  • Before ordering, the AI writes the price that would show its idea was wrong. If price reaches it, the trade is closed.
  • When a trade moves into profit, the stop follows at 2.25× the average range.

How losing trades closed since November 1, 2026

  • Stop-loss: 7
  • Invalidation price: 1

See the full record

When KAI stops

  • After 4 losses in a row within 72 hours, new trades stop for 12 hours.
  • If the day’s losses reach 3% of the account, no new trades that day.
  • At most 12 open trades, with stops totaling no more than 2% of the account.
  • The operator can stop new decisions at any time. While stopped, posts on X stop too.
  • If the AI can’t be reached, KAI does not open new trades.

KAI is run to test an AI, so we do not guarantee that it runs or that updates arrive. We may change or stop it without notice, and we do not announce outages or maintenance. See KAI’s status

Learning and reasoning, honestly

KAI is not an AI that gets better by learning from its trades. What it does today is record and test.

  • Right after a trade closes, the AI scores its own decision, and the score is recorded.
  • 24 hours after the close, the decision is checked against what price did next and tagged, for example “exited too early”.
  • Candidate lessons from these reviews are collected and tested against results.
  • A hypothesis is not trusted until chance can be ruled out, and hypotheses under study are never fed straight into decisions.

The key point: none of this learning is fed back into live decisions yet. Our tests of the record showed that giving the AI its memory pulled its direction. It will stay out until it is shown to help.

Known weaknesses

  • There are still too few closed trades to judge the results.
  • So far the record shows no evidence that the AI picks market situations well.
  • It leans heavily to selling. Before tracking began (27 Aug–31 Oct 2026) there were 254 sells and 38 buys. We are still looking into why.
  • Pips are measured from the decision price, so they leave out the difference to the fill price. A pip is worth a different amount in each pair, so total pips don’t equal money.
  • The record is not independently audited. Instead, we publish every closed trade and every correction.

Methods we tested and retired

Before KAI, we tested a number of methods that traded on fixed rules alone, and retired them. Their records, losses included, are kept.

See the methods we retired

Failures and changes

When the system changed and what changed. The reasons and sources are kept in our records.

  1. Record

    Results tracking started.

  2. Rule

    The market read is now done by code that applies the same rules to rises and falls. No AI is used for it.

  3. Safeguard

    The losing-streak pause now counts only losses within the last 72 hours.

  4. Rule

    Candidates were narrowed to pairs of the 8 major currencies.

  5. Rule

    A decision with 2 of 3 in agreement is now ordered at once instead of waiting.

  6. Rule

    The AI is now asked three times about the same moment, and KAI trades only if at least two answers agree.

  7. Rule

    Candidate selection was rebuilt: only pairs whose cost is low compared with price movement, a gap before the same pair is judged again, and caps on the number of new trades.

  8. Model

    The deciding AI was switched to Claude Haiku 5.5.

  9. Record

    We started counting process records such as skips and corrections.

What we don’t publish

Account balance, trade size, money amounts, keys, the AI’s full instructions and server names are not published.

See the record, or join

The full record is open to everyone, no account needed. To get KAI’s decisions and closed trades on Discord, follow the steps to join.

FX trading carries a high risk of loss. Make your own decisions about whether to trade, and get independent advice if you need it. Past performance is not a reliable indicator of future results. We record and publish an AI’s decisions for research. This is not investment advice.