Price Embedding with a Backpropagation-Trained Restricted Boltzmann Machine

Summary

This article proposes transforming financial price changes into a lower-dimensional representation with a Restricted Boltzmann Machine (RBM), then passing that representation to a multilayer perceptron (MLP) that predicts a subsequent price change. It presents an unconventional RBM training approach: backpropagation minimizes reconstruction error between the input and its reconstruction, instead of relying on Gibbs sampling and contrastive divergence. The hidden layer is treated as a learned embedding that may capture useful structure in price inputs.

The author suggests comparing the RBM-to-MLP model with an MLP trained directly on prior price changes to assess whether embedding helps. The article also mentions possible extensions to other time series inputs, such as indicators or candle patterns. The supplied material motivates the method and sketches its implementation, but gives no comparative trading results or evidence that the representation adds predictive value. Its analogy to word embeddings is a conceptual framing, not proof that the network captures attention or improves forecasts.

Key ideas

• The proposed pipeline uses an RBM hidden representation as input to an MLP price-change predictor.
• The RBM is trained by backpropagation to reconstruct its input, using that input as its own target.
• This approach replaces the traditional Gibbs sampling and contrastive divergence training process described in the article.
• A raw-input MLP provides a suggested benchmark for evaluating whether price embedding adds value.
• The article offers a modeling proposal but no comparative results establishing trading efficacy.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.