Markets turn when a crowd changes its mind.
MULLINER watches for those moments, forecasts what follows, and publishes its record whether or not it was right.
The current research has not established a durable edge.
A turn happens when buyers and sellers who want opposite things arrive at the same price. MULLINER looks at everything surrounding that moment and forecasts what comes next. Why that might work →
USD / CAD5 min candles
———Each row follows one model across its recorded history. Hover or tap a cell to see its vote and model ID.
Experiments and their outcomes.
Start with the question, the registered pass criteria and the recorded result. Read the experiment history, including failures →
The workshop below shows the current models and their development evidence.
The next generation.
Models inherit, adapt, and compete. Admission depends on recorded outcomes; new candidates must pass tests before influencing a trade.
How models evolve
A neural network is a program that learns numerical patterns from examples of market conditions and later price moves.
Its DNA is a recipe. Each gene is a setting: how the model is built, how it learns, or which clues it uses.
Parent models supply settings for a new candidate. Some settings change. The workshop trains and tests the candidate before selection.
A generation is one round of model selection. The five-minute timer checks evidence; it does not require a new generation. Each retained model must supply four new, non-overlapping simulated outcomes, or finish its retirement review. Four outcomes set a review budget; they do not prove an edge. More generations do not automatically mean better results.
Select a model below to see its layers, settings, and parents. Open the test results to compare all candidates.
Waiting for evidence status
Team admission and retirement Four outcomes per review
Live models can influence trades. Shadow models make recorded forecasts with zero trading influence. A model retires after a losing review or too little activity across eight fully observed market hours. A missing feed does not count as silence.
| Model | State | Resolved outcomes | Observed hours | Reason |
|---|
Select a neural candidate
Compare model test results Historical tests · simulated results
| Candidate | Type | Side | Selected | Trades | Base / stress · pips | Test score |
|---|
HOMOLOGATION · The next test
Registered before the runCan a varied ML team beat the original pair? September 14–19, UTC. Frozen models, equal starting weights and the same future prices. Missing evidence counts as failure.
Awaiting the experiment monitor.
Costs and the pass criteria
Base costs: 1.5-pip spread, 0.2-pip slippage per side, 0.5 pip per day financing. Stress: 3 / 0.5 / 1. A spread sweep changes only the spread. Scores are uncalibrated; 0.55 is not a win rate.
The varied team must survive both cost sets, beat all five controls and pass the registered activity, coverage and paired day-bootstrap checks. These are one-hour simulations with fixed stops, not the live trailing-stop account.
Forecast ledger
The scorebook: what did price do after each forecast? Results are simulated.Select a result to inspect the original forecast. Pending means more price data is needed. These are calculated outcomes, not account profits.
| Recorded | Model | Sim. entry | Evaluation | Base / stress · pips |
|---|
The ledger starts with forecasts recorded by the running system.
Price level reference
These prices come from prior UTC days and weeks. Chart lines show the latest reference set across the view.
| Period · UTC | Starting | Reference | Price |
|---|