LSTM Neural Network-Based Stock Price Prediction System: From Data Acquisition to Model Deployment

The open-source project stock-market-prediction introduced in this article is an LSTM neural network-based stock price prediction system developed by Sagar Gupta, built using the PyTorch framework. It supports price prediction for multiple tech stocks including Apple (AAPL), Microsoft (MSFT), Amazon (AMZN), and Tesla (TSLA), covering the complete workflow of data acquisition, preprocessing, model training, evaluation, and visualization. It aims to provide a reproducible machine learning engineering example and is suitable as an introductory learning resource for financial machine learning.

Stock price prediction is a challenging problem in the financial field due to the influence of multiple factors such as macroeconomics, company fundamentals, and investor sentiment. With the development of deep learning technology, the application of LSTM in time series prediction provides the possibility to capture potential patterns in stock prices. The core objective of this project is not only to implement the prediction function but also to provide a complete and reproducible machine learning engineering workflow, with full-link documentation from data to deployment, to help learners master financial machine learning practices.

LSTM Network Structure

The input layer includes 6 features (open price, high price, low price, close price, trading volume, average price). The LSTM hidden layer has 100 units, and the output layer has 1 unit to predict the future close price. The sequence length is 5 days (using data from the first 5 days to predict the 6th day's price).

Optimizer and Loss Function

The Adam optimizer (learning rate 0.001) is used, and the loss function is Mean Squared Error (MSE).

Data Flow

• Acquisition: Daily data from 2020 to 2023 is obtained via the Yahoo Finance API;
• Feature Engineering: Calculate average price, normalize with MinMaxScaler (independent scaling per stock), generate 5-day sequence inputs and 6th-day close price outputs using sliding windows;
• Splitting: The latest 30 days are used as the test set, and the rest as the training set (to avoid look-ahead bias).

Training Configuration

150 epochs of training, batch processing, reset LSTM hidden state for each sequence, Adam optimizer adjusts parameters automatically.

Performance Evaluation

Mean Squared Error (MSE) is used as the main metric to quantify the deviation between predictions and actual values.

Visualization

Generate comparison charts of actual vs. predicted prices, 20-day/50-day moving averages, and feature importance analysis to help understand model performance.

The project supports multi-dimensional customization:

• Stock Symbols: Modify the list to analyze any stock listed on Yahoo Finance;
• Time Range: Adjust the start and end dates of data to study different periods;
• Model Parameters: Adjust LSTM hidden layer size, sequence length, learning rate;
• Training Epochs: Flexibly adjust the number of iterations. Expansion directions include trying GRU/Transformer architectures, integrating indicators like RSI/MACD, and developing real-time prediction APIs.

The project has the following limitations:

• Market Efficiency: The Efficient Market Hypothesis suggests that historical prices are difficult to predict the future;
• Overfitting Risk: Neural networks tend to memorize noise rather than real patterns;
• External Shocks: Cannot predict the impact of black swan events. Disclaimer: The prediction results are for educational research only and do not constitute investment advice. For actual decisions, please consult a professional advisor.

Tech Stack

PyTorch (deep learning framework), yfinance (data acquisition), pandas (data processing), scikit-learn (preprocessing/evaluation), matplotlib (visualization), numpy (numerical computation).

Community Contributions

The project is open-source under the MIT license. Issues or Pull Requests are welcome. Future directions include integrating more indicators, implementing other architectures, portfolio optimization, and real-time API development.

stock-market-prediction is a well-structured and well-documented implementation of LSTM-based stock price prediction, providing a valuable reference example for beginners in time series prediction and financial machine learning. However, it should be emphasized that stock price prediction is extremely difficult, and no model can guarantee stable profits. The core value of the project lies in education and research, and readers should respect the complexity of the financial market.