LiqFormer: A Hybrid Liquid-Transformer Approach for Multi-Horizon Stock Market Forecasting
Year: 2026
Type: conference-paper
Abstract: Predicting financial markets is difficult due to their nonlinearity, volatility, and non-stationarity. Deep learning models such as RNNs, LSTMs and Transformers typically capture either short-term temporal dependencies or global contextual linkages, but not both. In this work, a hybrid Liquid-Neural Network and Transformer architecture proposed LiqFormer. It combines the adaptive, continuous time dynamics of liquid neurons with the self-attention mechanism of a Transformer. The Boruta feature selection method prioritizes relevant features to enhance interpretability and mitigation. Four major stock indices: the S&P 500, Dow Jones Industrial Average, NSE India, and Shanghai Composite are used to test the approach for 5, 10 and 15 years. Both computational metrics (MSE, R2) and financial profitability metrics (Sharpe, Total Return) are used to assess performance. LiqFormer outperforms baseline LNN, LSTM, RNN and Transformer models, with a Sharpe ratio of 2.0, R2= 0.99, MSE ≤ 0.01. High prediction accuracy and risk-adjusted profitability are attained by the hybrid approach LiqFormer.
Authors: Anisha Singh, Alok Kumar, Balu L Parne
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7203597516
