An optimized convolutional neural network with a novel spherical triangular fuzzy pooling layer for an algorithmic trading model

Abstract

Pooling layers of Convolutional Neural Networks (CNNs) are applied to reduce the dimensionality of input features while overcoming the overfitting issue. Typically, average pooling assigns the same weight to each activation in the pooling region. However, each region in an input image may not be equally crucial in financial data. To overcome this drawback, we propose a novel pooling layer that incorporates spherical fuzzy logic to account for the inherent uncertainty of financial data in a novel algorithmic trading model. To that end, we propose two minor models for prioritizing activations based on technical indicators of extreme points and time positions, respectively. Based on these two models, we propose an Indicator-Time-based Spherical Triangular Fuzzy pooling (ITSFP) model to capture critical market signals with greater accuracy and adapt more effectively to market conditions. To optimize the algorithmic trading model, a Genetic Algorithm (GA) has been designed to fine-tune the architecture of the proposed CNN, improving its adaptability and performance. The results demonstrate that the optimized ITSFP outperformed the others, achieving an accuracy of 80.08 %, while the optimized Fuzzy Pooling (FP), Max Pooling (MP), and Average Pooling (AP) achieved accuracies of 69.84 %, 63.60 %, and 59.77 %, respectively. The algorithmic trading based on the ITSFP model achieved the highest compound return (2.5866), Sharpe ratio (0.3632), and Sortino ratio (0.5242), reflecting its superior risk-adjusted returns and recorded the lowest Maximum Drawdown (2.8632), indicating superior resilience during market downturns. The Wilcoxon signed-rank test has been applied to show significant outperformance of the proposed ITSFP against others.