Transfer Learning for Market Models

Article MQL5 articles

Summary

The article evaluates transfer learning for a neural model trained to detect price fractals. It compares a model initialized with random weights against models that reuse the encoders of previously trained variational autoencoders, including fully connected and LSTM-based versions. It considers two ways to make the comparison fair: match the architecture after the reused encoder, or match the full architecture while randomizing the weights. A shared testing expert advisor loads saved models and trains them under common conditions.

The reported conclusion is that transfer learning reduces the new model’s training time, but the measurement excludes the time needed to pretrain the autoencoder. Including that work could erase the advantage, and the article does not establish that transfer learning is faster overall. It may be useful when an encoder is reused across tasks, when training a full model is difficult, or when model complexity is increased gradually. The approach also depends on new training data being similar to the donor model’s source data.

Key ideas

The experiment compares random initialization with transfer learning from variational autoencoder encoders.
It tests both architecture matching after the borrowed encoder and matching the complete architecture.
A shared expert advisor is used to train the candidate models under common conditions.
The reported training-time benefit excludes the cost of pretraining the donor autoencoder.
Transfer learning may be most useful when reusing a model block or when full-model training is difficult.

Testing Variational Autoencoder Transfer Learning for Market Models

Summary

Key ideas

Tags