Astra Neural Engine: Gold M30 Demo Setup and Personal Model Training

Astra Neural Engine lets you train neural models locally in MetaTrader 5 and use frozen models for automated execution. This guide explains the first demo test and the separate personal-training workflow introduced in version 2.8.

Open Astra Neural Engine on the MQL5 Market and download the free demo. The Market demo is tested in the Strategy Tester.

1. Start with the included Gold model

You do not need to train a model before this first test. The supplied reference is XAUUSD on M30.

1. Open the MT5 Strategy Tester and select Astra.
2. Select your broker's XAUUSD symbol, M30, and Every tick based on real ticks. Allow MT5 to synchronize the required history.
3. In Inputs, select Workflow = TRADE - included Gold model and Timeframe = M30.
4. Set ExecutionCheck = false. Otherwise the execution diagnostic runs instead of the neural strategy.
5. Keep the included model's architecture and execution contract settings at their defaults.
6. For the documented historical replay, select 1 January 2025 to 7 October 2026. Run a single test without optimization.
7. Inspect the Journal, individual trades, equity drawdown and broker costs.

This reference period was used for training. Replaying it shows the fitted model's behavior. It is not an independent test or a live track record. Test later, unseen dates separately; different broker data and execution conditions can produce different results.

2. Train a separate personal model

Choose TRAIN - tester dates / chart months. In the Strategy Tester, its start and end dates determine the training interval. TrainingMonths applies to chart training instead. Choose the symbol and set Timeframe for your intended model.

1. Give the experiment a distinct ModelId, for example xau_m30_experiment_01.
2. Choose your training dates and use real ticks. Disable optimization and the tester forward split.
3. Choose chronological validation or full-period fitting. For full fitting, set UseValidation=false, ValidationPct=0 and Stride=1.
4. Start the test. History is replayed first; fitting starts after replay completes. This training pass does not measure broker trading performance.
5. Follow the epoch reports. NEW BEST refers to the selected loss metric, not guaranteed trading profitability.
6. Wait for MODEL READY and DONE. Save your input preset and the bundle path printed in the Journal.

Training initializes new weights and does not reuse the included Gold weights. A failed run does not activate a partial model. Use a new ModelId for each experiment so an earlier completed run cannot be mistaken for the new one.

3. Test your saved personal model

Select TRADE - personal trained model and ExecutionCheck=false. In the tester, keep the same ModelId and training configuration, set UseLatestModel=true, and leave ModelBundle empty. Selection is tied to the symbol, timeframe and account/broker identity.

To inspect trading on the personal model's training dates, enable TesterDiagnosticReplay in the tester. Treat this as an in-sample diagnostic. For a chart, use the explicit saved bundle path. Preserve the complete generated model files and their dependencies, not just the bundle index.

What the network evaluates

Two causal TCN branches read the same timeframe over different temporal horizons. Their combined representation feeds a causal Transformer, and the complete network is trained jointly. LONG and SHORT outputs estimate net R relative to trade risk; they are not win-probability percentages.

If the first test has no trades

• Confirm ExecutionCheck=false and the intended TRADE workflow.
• For the included model, use the Gold M30 reference setup and matching defaults.
• For a personal model, confirm successful training and the exact ModelId or bundle path.
• Check the Journal for missing history, incompatible settings, model-date restrictions, cost filters or risk limits.
• Do not disable risk controls simply to force trades. A model can also legitimately remain out of the market.

For setup assistance, use the Astra product comments. Include the EA version, symbol, timeframe, workflow and relevant Journal message, removing private account details.

Try it before choosing

Visit Astra Neural Engine and test the free Market demo. Start with the included model, then evaluate whether local training fits your own research workflow.

Trading involves risk of capital loss. Historical and training-data results do not guarantee future performance.