In the field of financial forecasting, the complexity–accuracy paradigm—the assumption that more complex models yield superior performance—is frequently challenged by market noise and non-stationarity. This study tests this paradigm by evaluating advanced LSTM variants: the core Long Short-Term Memory (LSTM) unit (sLSTM), the matrix LSTM unit (mLSTM), and the extended LSTM architecture (xLSTM), which integrates these units into stacked residual blocks. We systematically benchmark these variants against standard LSTMs and the advanced benchmark model, TimesNet. Extensive experiments span six diverse financial datasets (comprising mature U.S. equities, a macro index, and high-volatility Chinese A-shares) and four historical window lengths. Results demonstrate that the core sLSTM and mLSTM units consistently deliver superior forecasting performance. Crucially, the targeted architectural innovations of sLSTM and mLSTM not only outperform the standard LSTM and TimesNet benchmarks individually but also surpass the more structurally complex xLSTM module configuration. This advantage remains robust across different asset types, indicators, and window lengths, with particularly outstanding performance at the 10-day length window. This study thus provides strong counterevidence to the “complexity–accuracy” paradigm in this field, proposing a data-driven innovation direction for practical trading systems: prioritizing efficient, high-performance core model innovations over generalized architectural complexity.
Huang et al. (Sun,) studied this question.