Randomized Weight Search for Neural Network Trading Models in MetaTrader 5
Summary
The article builds on earlier neural-network trading experiments and describes testing small feedforward networks inside MetaTrader 5. It compares three input approaches, including candle-size information, price-shape measurements, and moving-average slope angles, using networks with different layer configurations. The proposed goal is to develop an Expert Advisor that trains and trades without external software.
To address the Strategy Tester’s limited ability to explore many weight and bias combinations, the author proposes optimizing a single pass counter while the EA generates randomized parameter sets internally. A CSV file records combinations already tried so that the process can avoid repeats. The article reports experiments across six EA variants and says the results warrant further optimization, but the supplied text does not provide enough complete performance evidence to judge robustness.
The author cautions that deeper training and forward testing are still needed and that the experiments require substantial computing resources. Random search and in-sample optimization alone do not demonstrate predictive value or durable profitability.
Key ideas
- The experiments compare neural networks using different price-derived inputs, including candle measures and moving-average slopes.
- A small network is presented as a starting point before testing a deeper architecture.
- The optimization method varies one tester input while generating weight and bias combinations within the EA.
- A CSV file is used to retain tried parameter sets and avoid repeating them.
- The reported experiments require further training and forward testing before their trading value can be assessed.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.