Introduction
In the field of quantitative finance, researchers have long sought the "holy grail" of stock price prediction using increasingly complex neural network architectures. While many published models boast impressive accuracy based on standard regression metrics like Mean Absolute Error (MAE) and Mean Squared Error (MSE), these models often fail to deliver profits in real-world trading environments.
This paper identifies a fundamental "statistical flaw" in current evaluation methodologies: the Absolute Value Constraint. By treating stock prediction as a simple distance-minimization task, researchers inadvertently ignore the most critical factor in trading—the direction of price movement.
The Absolute Value Constraint Explained
Most Neural Network Stock Price Prediction (NNSPP) models are evaluated using metrics that calculate the magnitude of error between predicted and actual prices. However, in securities markets, the sign of the movement is often more important than the magnitude of the error.
Consider the "Color-Blind Pilot" analogy: a pilot who can measure light intensity but cannot distinguish red from green. In trading, a model might predict a price change with high precision (low MAE), but if it predicts a slight rise when the stock actually falls, the resulting trade is a loss. Metrics like MAE ($|y - \hat{y}|$) remove the negative sign by definition, effectively blinding the evaluator to whether the model correctly identified the trend direction.
Methodology
To investigate this discrepancy, the study implemented six neural network architectures: Multilayer Perceptron (MLP), RNN, LSTM, GRU, BiRNN, and BiLSTM. To ensure optimal performance, hyperparameters (such as window length and hidden nodes) were tuned using Orthogonal Array Testing, a statistical method for efficient parameter optimization.
Predictions were tested against a robust dataset of 40 stocks (20 from Chinese markets and 20 from US markets like NASDAQ and S&P 500) spanning 2011 to 2017, totaling over 62,000 trading days.
Key Results: The Correlation Gap
The core finding of the research is that traditional error metrics are poor proxies for trading success. The study calculated the Pearson Correlation Coefficient between standard metrics and the Stock Return Direction (SRD).
As shown in the correlation analysis, the relationship between MAE and SRD is weak (approximately -0.2775), and the correlation for $R^2$ is negligible (0.1400). This statistical independence proves that achieving a "more accurate" model in terms of regression error does not necessarily lead to better directional insight or higher returns.
A Tale of Two Stocks
The practical impact of this constraint is best illustrated by comparing the performance of an LSTM model on two different stocks: the Chinese stock 300027 and the US stock AAPL.
Both stocks achieved remarkably low MAE values (0.0104 and 0.0136, respectively), which would typically signal a highly successful model. However, the financial outcomes were polar opposites. Following the model's predictions for 300027 resulted in a 28.61% loss, while for AAPL, it yielded a 30.51% gain. Traditional metrics treated these results as equally valid, failing to penalize the model for its catastrophic directional failure in the first case.
Conclusion
The research concludes that standard regression metrics are insufficient for evaluating financial prediction models. For a model to be useful in a trading context, it must be "directionally consistent" rather than just "precisely wrong." Researchers and developers should move toward evaluation frameworks that incorporate directional accuracy and return rates to avoid the pitfalls of the Absolute Value Constraint.
Summary updated