<![CDATA[Abstract—Algorithmic trading models for highly volatile assets such as Tesla (TSLA) stock need to adapt to rapidly changing market conditions. Deep Reinforcement Learning (DRL) provides an adaptive approach for this task. Dueling Double Deep Q-Network (D3QN) combines Double DQN and Dueling DQN to reduce overestimation bias and improve state-value representation. However, D3QN does not explicitly prioritize individual state features, which might limit its ability to extract informative signals from noisy financial data. In this study, we propose Gated Attention Dueling Double Deep Q-Network (GA-D3QN). GA-D3QN applies a gated attention mechanism after the flattened state and before the shared feature layer of D3QN to adaptively weight state information during Q-value estimation. D3QN and GA-D3QN were trained and evaluated on TSLA data from 2018 to 2024, with two configurations (OHLCV and OHLCV + Technical Indicators), using 3-fold expanding-window walk-forward evaluation. Performance was evaluated using cumulative return, Sharpe ratio, maximum drawdown, and convergence speed. GA-D3QN improved mean cumulative return and Sharpe ratio while reducing maximum drawdown in both configurations; however, the improvements were not consistent across folds. Under OHLCV, cumulative return improved from -1.713% to 1.750%, Sharpe ratio improved from 0.038 to 0.179 and maximum drawdown decreased from 40.183% to 36.833%. Under OHLCV + Technical Indicators, cumulative return increased from 9.380% to 15.630%, Sharpe ratio increased from 0.207 to 0.531, and maximum drawdown decreased from 35.183% to 33.430%. GA-D3QN converged faster for OHLCV but slower for OHLCV + Technical Indicators. These results suggest that the effect of Gated Attention on convergence depends on the input configuration. Keywords—Deep Reinforcement Learning, Dueling Double Deep Q-Network, Gated Attention Mechanism, Algorithmic Trading, Walk-Forward Evaluation, Technical Indicators.]]>
Copyrights © 2026
