[Paper Review] Financial Time Series Prediction Using Deep Learning

Ariel Navon, Yosi Keller|arXiv (Cornell University)|Nov 11, 2017

Stock Market Forecasting Methods22 references3 citations

TL;DR

This paper proposes an end-to-end deep learning model that predicts financial time series trends directly from raw stock and ETF price data, using a neural network trained to forecast directional movements. The model generates probabilistic predictions to guide a commission-aware trading strategy, achieving statistically significant profitability and outperforming benchmarks over a two-year backtest with realistic transaction costs.

ABSTRACT

In this work we present a data-driven end-to-end Deep Learning approach for time series prediction, applied to financial time series. A Deep Learning scheme is derived to predict the temporal trends of stocks and ETFs in NYSE or NASDAQ. Our approach is based on a neural network (NN) that is applied to raw financial data inputs, and is trained to predict the temporal trends of stocks and ETFs. In order to handle commission-based trading, we derive an investment strategy that utilizes the probabilistic outputs of the NN, and optimizes the average return. The proposed scheme is shown to provide statistically significant accurate predictions of financial market trends, and the investment strategy is shown to be profitable under this challenging setup. The performance compares favorably with contemporary benchmarks along two-years of back-testing.

Motivation & Objective

• To develop an end-to-end deep learning framework for financial time series prediction using only raw price data, avoiding engineered features.
• To design a probabilistic trading strategy that leverages neural network confidence scores to optimize entry and exit points.
• To evaluate the profitability and robustness of the proposed system under realistic transaction costs across multiple assets.
• To compare the performance of the deep learning-based strategy against baseline buy-and-hold and other machine learning approaches.
• To investigate the relationship between asset volatility and trading performance in the context of deep learning-based active trading.

Proposed method

• A deep neural network is trained directly on raw historical price sequences to predict the sign of future price movement (i.e., trend direction) over a fixed horizon.
• The model outputs soft probabilities indicating the likelihood of an upward or downward trend, which are used to assess prediction confidence.
• A probabilistic trading strategy is derived that uses these confidence scores to determine when to enter or exit long positions, minimizing false signals.
• The system is optimized to maximize average return under realistic commission constraints, with unified buy and sell fees applied.
• Backtesting is conducted over a two-year period on multiple S&P 500 constituents (e.g., SPY, AAPL, AMZN, BAC) using minute-level data.
• Performance is evaluated using metrics including cumulative gain, annualized volatility, and Sharpe ratio, with comparisons to baseline buy-and-hold and other ML models.

Experimental results

Research questions

• RQ1 Can a deep learning model trained directly on raw price data achieve statistically significant predictions of financial market trends?
• RQ2 How does a probabilistic trading strategy based on neural network confidence scores perform under realistic transaction costs?
• RQ3 What is the relationship between asset volatility and the profitability of the proposed deep learning-based trading system?
• RQ4 Does the proposed end-to-end deep learning approach outperform traditional benchmarks such as buy-and-hold and other machine learning models in a realistic trading environment?
• RQ5 To what extent do transaction costs affect the profitability and risk-adjusted returns of the active trading strategy?

Key findings

• The deep learning model achieved statistically significant predictions of financial market trends using only raw price inputs, without engineered features.
• The proposed probabilistic trading strategy generated positive cumulative returns across all tested instruments over a two-year backtest, even with realistic transaction costs.
• For a 0.1% commission rate, the strategy achieved a cumulative gain of 152.2% on Amazon (AMZN), outperforming the baseline buy-and-hold return of 86.1%.
• The Sharpe ratio exceeded 0.5 for all assets when transaction costs were below 0.07%, indicating strong risk-adjusted performance.
• The cumulative gain of the active trading strategy was positively correlated with asset volatility, as shown by linear regression in Figure 10.
• The strategy underperformed for low-volatility assets (T, JNJ, KO) at 0.1% commission, indicating that high volatility may be favorable for this approach.