|
|
 |
| |
|
|
|
An end-to-end Combined Forecasting Architecture: Forecasting Stock Price Data |
|
|
|
PP: 59-75 |
|
|
|
doi:10.18576/jsap/140105
|
|
|
|
Author(s) |
|
|
|
Katleho Makatjane,
Claris Shoko,
Caston Sigauke,
|
|
|
|
Abstract |
|
|
| In this paper we evaluated five models within a Bayesian framework in quantifying aleatoric uncertainty of financial time series data and, in particular, stock prices. Statistical-based predictions with deep learning algorithms improve the performance of stock price forecasting models not by choosing the model structure expected to predict the best but by developing a model whose results are a combination of models with different shapes. Using the minimum combined score to conglomerate the TBATS and the α-RNN, we found that the combined model mimics the forecasting error compared to individual deep learning algorithms, with coverages of 98.7% and 88.76%, for five and ten-day steps ahead, respectively. By leveraging TBATS for capturing complex seasonality and α-RNN for modeling memory decay and long-term dependencies, the estimated model demonstrates robust performance in short and medium-term predictions. our findings further highlight the ability of the model to address volatility clustering, trend detection, and seasonality more effectively than traditional methods such as ARIMA, GARCH, and standalone RNN-based models methods. The model significantly lowers error metrics, improves forecast accuracy and better handles financial uncertainties. |
|
|
|
|
|
|
|
