Azimuth: proprietary neural decision-making for Gold

Azimuth is built on the exact same proprietary foundation as Meridian, but introduces a different internal market representation and a more advanced architecture.

Instead of relying directly on conventional timeframe candles, Azimuth uses an internal custom bar structure built directly from tick data. The bars are generated from market activity rather than simply from the passage of time.

Because of this, tick data is essential when testing Azimuth. The system should always be tested using tick, regardless of which classical timeframe is selected in the Strategy Tester.

The result is a model that does not depend exclusively on the way a broker aggregates the market into conventional M1, M5, M15 or H1 candles.

The neural architecture

The model is based on the same proprietary neural structure used as the foundation of Meridian, where neurons themselves have been redefined.

A conventional neural network fundamentally relies on neurons aggregating their inputs before passing the resulting information through an activation function. This architecture is extremely powerful and works well for many classes of problems.

Financial markets, however, are particularly difficult to model.

The relationship between two market inputs can change depending on context, volatility, recent structure and the behaviour that preceded the current signal. The same input can therefore have a completely different meaning under different market conditions.

Azimuth and Meridian are designed around a different internal topology, specifically built to process these non-linear relationships.

Rather than treating every input as an independent signal that simply contributes to a final score, the structure is designed to analyse how the information interacts and develops inside the network.

This allows the model to work with more complex relationships between its inputs and to react to the structure of a signal rather than simply adding signals together.

The architecture is proprietary and its internal topology is not exposed.

Designed for non-stationary markets

Financial markets are not static systems.

Volatility changes. Market structure changes. The relationship between different variables changes. A behaviour that was profitable in one period can disappear completely in another.

This is one of the main reasons why a model can look exceptional on historical data and then behave very differently once exposed to unseen data.

Azimuth was designed with this problem in mind.

The objective is not to find a configuration that perfectly explains one historical period. The objective is to find a model that can continue to behave correctly when the data changes.

Walk-forward training

The training process is one of the most important parts of Azimuth.

The model is not trained on one large historical dataset and accepted simply because the final equity curve looks good.

Instead, Azimuth is trained through a walk-forward matrix with multiple independent out-of-sample periods.

The process begins with:

15% Learning → 5% completely unseen OOS data

If the model passes the first OOS period, the process moves forward and repeats:

15% Learning → 5% OOS
15% Learning → 5% OOS
15% Learning → 5% OOS
15% Learning → 5% OOS
15% Learning → 5% OOS

This creates five separate out-of-sample validation periods.

The OOS data is not used to train the model.

More importantly, if any OOS period shows a degradation beyond the required criteria, the model is discarded and the entire learning process starts again.

It does not get patched.
It does not get manually adjusted.
The training process starts over.

Only models capable of surviving the complete sequence of unseen periods are retained.

This approach puts the emphasis on generalisation and out-of-sample behaviour, rather than simply fitting historical data.

How Azimuth trades

Azimuth uses direct market execution rather than the pending-order approach used by Meridian.

When the model identifies a valid opportunity, the system executes directly at market.

The model continuously evaluates the internal market representation and determines whether the current structure provides a valid trading opportunity.

Once a position is opened:

• The system does not average down.
• There is no grid.
• There is no martingale.
• Positions have a defined maximum lifetime.
• Risk parameters are determined according to the model's current market assessment.

The objective is to allow the model to make the decision itself rather than surrounding the network with a large collection of fixed trading rules.

High trade frequency

The combination of its internal market representation, direct market execution and architecture results in a high trade frequency.