Meet the team behind each decision.

One price. Two possible directions.

Forex means trading currencies. MULLINER studies USD/CAD: how many Canadian dollars one US dollar buys. A price of 1.35 means US$1 buys C$1.35.

A long position gains when that price rises. A short position gains when it falls. Either can lose when price moves the other way. Flat means no open position.

The aim is to find moves large enough to cover trading costs and leave a profit. An edge is a repeatable advantage after those costs. The current research has not established a durable edge.

Why turns?

At any moment there are three groups in the market: people betting the price rises, people betting it falls, and people waiting to commit.

When the price climbs away from a level and then comes back, all three groups act at once, for completely different reasons. The people who bought there want to buy more at the same price. The people who bet against it are now losing, and want out at the level where they'd break even. The people waiting see momentum and want in.

Three groups, three motives, one price. They buy at the same moment and the price bounces. Sellers produce the mirror image above.

This isn't a theory about chart patterns. Studies of actual currency orders find them clustered at exactly these prices — take-profits at round numbers, protective stops just beyond. The orders are really there. A confirmed turn is what it looks like when that cluster gets used up.

MULLINER's question is whether the conditions around one of those moments — which levels it formed against, how price approached, how volatile the session was, what every other measurable clue was doing — say anything useful about what happens next.

These motives are an explanatory hypothesis, not something the candles identify. Price data alone cannot recover the orders or prove that a particular turn exhausted an order cluster.

Order-flow evidence: Carol L. Osler, Currency Orders and Exchange-Rate Dynamics (2001), examines stop-loss and take-profit orders at a foreign-exchange bank and their clustering around round numbers. Stop-Loss Orders and Price Cascades (2002) distinguishes trend-amplifying stops from take-profits that can oppose a move. This evidence motivates the test; it does not establish MULLINER’s edge.

A machine with a legible record.

The vehicular names express the author's approach to markets: a machine with carefully fitted parts, a legible instrument panel and a record of how it was built. The luxury is in the workmanship — clear interfaces, reproducible models, controlled changes and details that remain accessible without crowding the driver.

The team, from clue to result.

COMPRESSOR · Find the turns

Price changes constantly. COMPRESSOR picks out turns after price moves back far enough to confirm them. A smaller scale reveals smaller turns, like adding detail to a map. A confirmed turn describes what happened; it does not guarantee the next move.

Models · Make the forecasts

A model is a program that estimates what might happen next. The active team learns from examples with neural networks, a linear classifier and a tree ensemble. Archived rules retain their earlier descriptions and votes. Each model can vote long, short, or flat.

CONSENSUS · Choose one target

The models form an ensemble: a team. CONSENSUS combines their votes into a direction. In experimental mode, that target stays in force until an opposite decision. Neutral votes keep the last target. Each model has a weight, meaning a share of influence. Measured results can change that share. New models first run in shadow: they make forecasts with zero trading influence. They must pass historical and later-outcome tests before joining the live vote. Losing or persistently inactive models retire with zero influence. If all supporting models retire, the old target is cleared.

OANDA · Carry out the trades

OANDA is the broker: the service that executes orders and holds the trading account. When execution is enabled, MULLINER uses its target to manage the position. The broker panel shows what the account actually holds.

The ledger · Keep score

A forecast is recorded first. Later prices show how a simulated trade would have performed after estimated costs. Pending means the system is waiting for enough later prices. Broker trades have separate records.

The workshop · Test the next team

A neural model’s DNA is its recipe. Genes are individual settings. New candidates inherit settings from parent models, with occasional changes. The workshop trains and tests the candidates, then selects a team. One selection round is a generation. The timer checks every five minutes, but replacement waits for four new resolved outcomes per model or a completed retirement review. An outcome is resolved when the later prices needed to score it are available. Overlapping trades do not count twice. The workshop shows the count and each model’s state.

Neuro-evolution is the name for combining neural models with this evolutionary search over their recipes. The models are not conscious. A new generation is a new test, not a promise of improvement.

Try it in one minute.

1. Read Model target. This is the position the models want.
2. Read Broker exposure. This is what the account holds. Open Position details for the amount, entry price, and current gain or loss.
3. Open Model record and select a model. Its votes stand out on the chart. The inspector explains its idea.
4. Open Research and select a candidate. See its settings and which earlier models supplied them.
5. Open Forecast ledger and select a result. Return to the forecast that came before it.

Green means long or rising. Red means short or falling. Colour alone does not mean profit or loss.

Execution and consensus policy

Live mode can place orders on the real account. Practice mode uses the broker’s practice account. Signal mode records decisions without placing orders. A live broker connection alone does not enable trading.

Experimental policy gives admitted models positive prior influence; shadow models still have zero influence. This is not the stronger evidence-gated qualified policy. Four resolved outcomes trigger a review; they do not establish an edge. Qualified policy additionally requires sufficient development and prospective evidence before a model receives positive influence.

Translate a term.

The learned team and its next test

The current team uses learned models: Intraday and Micro neural networks, a linear classifier and a tree ensemble. Each has a long and short specialist. A score of 0.55 is a screening score, not a 55% win rate. New models first make forecasts with no influence over trades; resolved results determine admission. Archived models retain their name, approach and recorded votes. HOMOLOGATION is a separate frozen experiment with a public result, including failure; it does not change the live team.