Adaptive mixture of experts framework for multi horizon stock market forecasting using heterogeneous deep learning models

Accurate stock price forecasting remains a fundamental challenge in computational finance due to the inherent non stationarity, volatility clustering, and complex temporal dependencies characterizing financial time series. This paper presents an Adaptive Mixture of Experts (AMoE) framework that dynamically integrates four specialized deep learning architectures i.e. Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) through a learned gating mechanism for multi horizon stock price prediction. The proposed framework is evaluated on five major technology stocks (Apple Inc., Microsoft Corporation, International Business Machines, Infosys Limited, and Meta Platforms Inc.) over the period from 01 July 2024 to 30 June 2025, encompassing 252 trading days across three forecasting horizons: one day ahead ( $$h_1$$ ), five day ahead ( $$h_5$$ ), and ten day ahead ( $$h_{10}$$ ). Experimental results demonstrate that while individual experts exhibit varying strengths across different stocks and forecasting horizons with GRU achieving superior performance for short horizon predictions on high liquidity stocks and CNN demonstrating advantages in capturing local pattern structures the AMoE framework consistently produces more stable and reliable forecasts by adaptively leveraging these complementary characteristics. For the $$h_1$$ horizon, the AMoE achieves mean Root Mean Square Error (RMSE) values of 2.847 for AAPL, 3.124 for MSFT, 1.892 for IBM, 0.423 for INFY, and 4.567 for META, with directional accuracy ranging from 54.3% to 57.9% (mean: 56.6%) across stocks. Statistical robustness analysis reveals that the AMoE framework reduces forecasting variance by 12.4% to 23.8% compared to individual experts while maintaining competitive accuracy. The primary contributions of this work lie in demonstrating adaptive expert collaboration mechanisms, quantifying horizon dependent forecasting behavior, and establishing a robust framework for ensemble based financial prediction that prioritizes consistency over marginal accuracy improvements.Empirical evaluations across five equities over 1-day, 5-day, and 10-day forecast horizons demonstrate that AMoE achieves statistically significant gains over standalone neural experts and statistical baselines like Diebold Mariano $$p < 0.05$$ at $$h_1$$ and $$h_5$$ ), reducing out of sample RMSE by 5.8% to 30.1%. Furthermore, analysis across multiple random initializations reveals a 21.1% to 51.6% reduction in optimization standard deviation compared to standalone deep networks, demonstrating superior training stability. These findings highlight the effectiveness of adaptive expert collaboration in exploiting complementary temporal patterns while maintaining robust forecasting performance across heterogeneous stocks and prediction horizons.