LSTM in Algorithmic Investment Strategies on BTC and S&P500 Index

Area of Science:

• Quantitative Finance
• Computational Finance
• Machine Learning in Finance

Background:

• Algorithmic trading strategies require accurate market forecasting.
• Traditional models often struggle with the volatility of assets like Bitcoin and the S&P 500 index.
• Long Short-Term Memory (LSTM) networks offer potential for time-series financial forecasting.

Purpose of the Study:

• To forecast the value of Bitcoin (BTC) and the S&P 500 index using LSTM networks.
• To introduce an innovative loss function enhancing LSTM forecasting for algorithmic investment.
• To develop and evaluate algorithmic investment strategies based on LSTM-generated signals.

Main Methods:

• Utilized LSTM networks with daily, 1-hour, and 15-minute data from 2013-2020.
• Developed and applied a novel loss function to improve forecasting accuracy.
• Implemented a rolling window approach for in-sample optimization and out-of-sample testing.
• Created ensemble models by combining signals from various data frequencies.
• Focused on meticulous data preprocessing and hyperparameter selection to prevent overfitting.

Main Results:

• LSTM models, enhanced by the new loss function, generated actionable buy and sell signals.
• Ensemble models combining different data frequencies showed promising performance in out-of-sample testing.
• The study successfully created equity lines demonstrating the profitability of the algorithmic strategies.
• Sensitivity analysis provided insights into key parameter and hyperparameter impacts.

Conclusions:

• The proposed LSTM-based approach with an innovative loss function is effective for forecasting BTC and S&P 500.
• The developed algorithmic investment strategies show potential for generating returns in financial markets.
• Ensemble modeling and careful data handling are crucial for robust financial forecasting with LSTMs.