Neural Network Classification for Technical-Indicator Stock Selection

Summary

This tutorial introduces a basic neuron model, describing weighted inputs, a threshold, and an activation function. It explains why a smooth sigmoid function is used in practice in place of an idealized step function, then applies a neural network to supervised stock classification. The proposed features are six measures drawn from valuation, market capitalization, and technical analysis: total market value, OBV, price-to-earnings, Bollinger Bands, KDJ, and RSI. Each sample is labeled according to whether the stock rises by more than a stated threshold over the following 20 trading days. Feature values are standardized by subtracting their mean and dividing by their standard deviation.

The stated trading rule buys when the model predicts a positive class and no position is held, and sells an existing position on a negative prediction. The document gives an initial capital amount and a daily backtest period, but provides no performance statistics or interpretation of the referenced equity curve. It also omits details needed to assess the model, including the network architecture, training procedure, validation design, transaction costs, and safeguards against look-ahead bias. The example therefore outlines a strategy concept rather than demonstrating robust predictive performance.

Key ideas

• The tutorial frames stock selection as a supervised classification problem using a neural network.
• Inputs combine valuation, market-capitalization, and technical-indicator features.
• Labels depend on whether a stock exceeds a specified forward return threshold over 20 trading days.
• Features are standardized using their sample means and standard deviations.
• The document states a daily buy-and-sell rule but reports no numerical performance evaluation.

Tags

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