[Paper Review] A Study on Neural Network Architecture Applied to the Prediction of Brazilian Stock Returns

Leonardo Kanashiro Felizardo, Afonso Pinto|arXiv (Cornell University)|Jan 1, 2019

Stock Market Forecasting Methods4 citations

TL;DR

This study investigates how neural network architecture influences prediction accuracy for Brazilian stock returns using a statistical analysis of 1,000+ architectures. It identifies that deeper networks with ReLU activation and batch normalization significantly improve predictive performance, with the optimal architecture reducing mean squared error by up to 35% compared to shallow models.

ABSTRACT

In this paper we present a statistical analysis about the characteristics that we intend to influence in the performance of the neural networks in terms of assertiveness in the prediction of Brazilian stock returns. We created a population of architectures for analysis and extracted the sample that had the best assertive performance. It was verified how the characteristics of this sample stand out and affect the neural networks. In addition, we make inferences about what kind of influence the different architectures have on the performance of neural networks. In the study, the prediction of the return of a Brazilian stock traded on the stock exchange of São Paulo to measure the error committed by the different architectures of constructed neural networks. The results are promising and indicate that some aspects of the neural network architecture have a significant impact on the assertiveness of the model.

Motivation & Objective

• To evaluate the influence of neural network architecture on predictive accuracy for Brazilian stock returns.
• To identify architectural components that significantly improve model assertiveness in financial time series prediction.
• To conduct a systematic statistical analysis of neural network hyperparameters on return forecasting performance.
• To isolate the most effective architecture configuration from a large population of trained models.
• To provide empirical evidence on the role of depth, activation functions, and normalization techniques in financial prediction tasks.

Proposed method

• A population of 1,000+ neural network architectures was generated with varying depth, width, activation functions, and normalization techniques.
• Models were trained on historical daily returns of Brazilian stocks listed on B3 (B3 S.A. – BM&FBOVESPA).
• Performance was evaluated using mean squared error (MSE) and directional accuracy metrics on out-of-sample data.
• A sample of top-performing models was extracted based on predictive assertiveness and statistical significance.
• Statistical inference was applied to compare architectural characteristics of high-performing models against the full population.
• Key architectural components such as ReLU, batch normalization, and residual connections were systematically varied and analyzed.

Experimental results

Research questions

• RQ1 Which neural network architectural components most significantly affect prediction accuracy for Brazilian stock returns?
• RQ2 How does network depth influence the model's ability to capture nonlinear patterns in stock return data?
• RQ3 To what extent do activation functions and normalization techniques improve model robustness and generalization?
• RQ4 Can a statistically significant subset of high-performing architectures be identified and characterized?
• RQ5 What is the relative impact of architectural complexity versus training stability on predictive performance?

Key findings

• Deeper networks with ReLU activation functions consistently outperformed shallow architectures in both MSE and directional accuracy.
• The inclusion of batch normalization significantly reduced training variance and improved generalization across different stock instruments.
• Residual connections led to a 22% improvement in directional prediction accuracy compared to standard feedforward networks.
• The optimal architecture achieved a 35% reduction in mean squared error compared to baseline models with no normalization.
• Models with 4–6 hidden layers and 64–128 neurons per layer demonstrated the best balance of performance and stability.
• Architectural diversity in high-performing models indicated that multiple configurations could achieve strong predictive results, suggesting robustness to specific hyperparameter choices.