Machine Learning Methods and a Wavelet-Autoencoder-GRU Stock Forecasting Pipeline

Summary

This review surveys machine learning approaches to stock forecasting, including support vector machines, random forests, nearest-neighbor and Bayesian classifiers, neural networks, and recurrent models. It also describes research examples that combine methods, such as wavelet denoising and stacked autoencoders before LSTM prediction, or singular spectrum analysis paired with an SVM. The cited studies report differing model comparisons, including cases where random forests or neural networks performed better under particular settings.

The proposed research pipeline collects OHLC history, preprocesses and denoises it with wavelets, extracts features through stacked autoencoders, and forecasts with a GRU. It names RMSE, MAPE, mean bias error, Theil’s U, and correlation as evaluation measures. The document summarizes prior work and outlines a proposed method, but supplies no results for that new pipeline. It also notes that text and macroeconomic data could complement price and volume inputs; forecast accuracy alone does not establish trading profitability.

Key ideas

• The review covers several classical machine learning and neural network approaches to stock prediction.
• Prior studies summarized in the document report model comparisons whose findings vary by dataset and method.
• The proposed pipeline applies wavelet denoising, stacked autoencoder feature extraction, and GRU forecasting to OHLC data.
• The proposed method lists multiple forecast error and correlation measures but provides no results for its own pipeline.
• News, social media, and macroeconomic variables are presented as potential additions to market data.

Tags

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