Financial Time Series Prediction Using Deep Learning
Financial Time Series Prediction Using Deep Learning
This paper presents a deep learning framework for predicting short-term trends in financial time series and a probabilistic trading strategy that leverages the model's confidence estimates. The neural network uses raw one-minute closing prices from the past 60 minutes as input to classify whether the price will rise or fall over a horizon T (typically 20-30 minutes). The trading strategy opens positions only when the predicted probability exceeds an adaptive threshold, and closes when the model's direction changes, incorporating a safety-switch to avoid consecutive losses. Backtested on SPY ETF and nine S&P 500 stocks over two years (2014-2016), the approach achieves cumulative gains of 7-152% depending on the asset and commission rate, outperforming both SVM-based active trading and buy-and-hold. The results demonstrate that deep learning can extract predictive signals from raw price data, and that selective trading based on model confidence can overcome transaction costs.
Highlights
- 1End-to-end deep learning approach for financial time series prediction using raw price data without engineered features.
- 2Probabilistic trading strategy that uses soft outputs from the neural network to select high-confidence trades.
- 3Outperforms SVM-based active trading and buy-and-hold strategy over two-year backtesting with realistic transaction costs.
- 4Achieves statistically significant trend prediction accuracy (53%) and profitable trading across multiple S&P 500 stocks and ETFs.
Methods
- MDeep neural network with 5 fully-connected layers (500, 200, 40, 20, 2 units) and ReLU activations, trained with dropout and early stopping.
- MProbabilistic trading strategy using adaptive soft-thresholding and safety-switch mechanisms to filter trades based on prediction confidence.
- MCross-validation to select prediction horizon T and adaptive threshold window D for each asset.
Results
- RSPY ETF cumulative gain of 61.1% over two years with 0.1% transaction costs, compared to 10.63% buy-and-hold.
- ROutperforms SVM-based strategy across all commission rates (0.02% to 0.2%).
- RActive trading profitability correlates with asset volatility; higher volatility assets (e.g., AMZN, GOOGL) yield higher returns.
- RSharpe ratios above 0.5 for commissions ≤0.07% across all tested assets.
Analyze Paper
Generate insights from "Financial Time Series Prediction Using Deep Learning".