3 Boyutlu Evrişimsel Sinir Ağları Temelli Algoritmik Alım Satım Sistemi

Abstract

Artificial neural networks are widely used in financial forecasting models. Although the most preferred model is LSTM, some studies based on CNN can also be found. In this study, the developed CNN model applies the convolution operation on three-dimensional data with a different approach. During data preparation, 18 different technical analysis indicators were selected. These indicators were calculated based on 20 different values, corresponding to periods ranging from 5 to 25 for each day. The resulting two-dimensional daily data was augmented with 20 days of past values, forming datasets of size 18 × 20 × 20 for each day. The data was labeled with Buy, Sell, and Hold classes. Based on the model's outputs, trading activities conducted over 750 trading days between 2022 and 2024 on Dow30 stocks and selected exchange-traded funds achieved an average annual return of 18.15% and 20.16%, respectively, outperforming the buy-and-hold strategy. © 2025 Elsevier B.V., All rights reserved.