.png){kind=icon}
.png){kind=icon}
Return
Enhancing LSTM performance for financial time series forecasting: the differencing approach - Diff-LSTM
.png){kind=icon}
OA
AI
L
E
H
N
M
DOI：10.1007/s00521-026-12369-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Long Short-Term Memory (LSTM) networks are increasingly preferred over conventional econometric models for forecasting non-linear time series such as stock market data. This study proposes the “Diff-LSTM” approach, designed to enhance the forecasting capabilities of standard LSTM networks for non-stationary time series such as stock market data. Unlike conventional applications, the Diff-LSTM methodology trains the network on first-differenced – hence, stationary- series from which forecasts for the original series (in levels) are reconstructed by updating with the observed previous value of the level series. In LSTM networks, due to the limitations of activation functions such as sigmoid and tanh, a downward bias can be observed in forecasts of non-stationary series that have a unit root. Therefore, the proposed Diff-LSTM approach mitigates this downward bias, enabling more reliable forecasts. Diff-LSTM models are evaluated for forecasting accuracy across five major U.S. stock market indices using three cross-validation schemes and four performance measures. The current study demonstrates the superiority of the Diff-LSTM model, which outperformed standard LSTM models in 56 of 60 evaluated instances. This suggests that using Diff-LSTM models improves the forecasting performance considerably.
Keywords:
Cross-validation
Long Short-Term Memory
Machine Learning
Journal
IF:
4.5
Papers:
729
Citations:
3.2W
Organization
Cited Papers