Wasserstein VAE Inference for MA and Stochastic Trading Patterns
Summary
The article presents an inference-learning workflow for patterns built from moving averages and the stochastic oscillator. It frames inference as estimating latent structure from observed data, then describes a Wasserstein variational autoencoder that encodes feature, state, action, and reward data into a latent representation. A maximum-mean-discrepancy objective is used to align encoded distributions with a prior; a separate linear regression model is used to estimate missing values in the combined dataset. The models are trained in Python and integrated into a MetaTrader workflow.
The reported evaluation uses data from 2023 through 2024, with a held-out period covering the final six months of 2024. The author says only patterns 1 and 5 appear to benefit in this short evaluation, while pattern 2 does not. The article supplies no performance figures in the provided text and cautions, in effect, that the evidence comes from a limited train/test window; broader robustness and transfer across markets remain unestablished.
Key ideas
- Inference is presented as a way to estimate hidden structure from observed trading data.
- A Wasserstein VAE compresses feature, state, action, and reward information into latent encodings.
- Maximum mean discrepancy is used to compare encoded distributions with a prior.
- Linear regression is applied to infer missing entries in the combined dataset.
- In the reported short evaluation, only patterns 1 and 5 were described as benefiting from inference.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.