Multi-horizon ternary time series forecasting
Time series forecasting techniques have been widely applied in domains such as weather forecasting, electric power demand forecasting, earthquake forecasting, and financial market forecasting. Because of the fact that these time series are affected by a multitude of interrelating macroscopic and mic...
| Main Author: | |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | English English English |
| Published: |
2013
|
| Subjects: | |
| Online Access: | http://irep.iium.edu.my/34350/ http://irep.iium.edu.my/34350/ http://irep.iium.edu.my/34350/8/zawzawTOC.pdf http://irep.iium.edu.my/34350/9/zawzawCoverMulti_Horizon.pdf http://irep.iium.edu.my/34350/14/43048.pdf |
| Summary: | Time series forecasting techniques have been widely applied in domains such as weather forecasting, electric power demand forecasting, earthquake forecasting, and financial market forecasting. Because of the fact that these time series are affected by a multitude of interrelating macroscopic and microscopic variables, the underlying models that generate these time series are nonlinear and extremely complex. Therefore, it is computationally infeasible to develop full-scale models with the present computing technology. Therefore, researchers have resorted to smaller-scale models that require frequent recalibration. Despite advances in forecasting technology over the past few decades, there have not been algorithms that can consistently produce accurate forecasts with statistical significance. This is mainly because state-of-the-art forecasting algorithms essentially perform single-horizon forecasts and produce continuous numbers as outputs. This paper proposes a novel multi-horizon ternary forecasting algorithm that forecasts whether a time series is heading for an uptrend or downtrend, or going sideways. The proposed system utilizes a cascade of support vector machines, each of which is trained to forecast a specific horizon. Individual forecasts of these support vector machines are combined to form an extrapolated time series. A higher level forecasting system then forward-runs the extrapolated time series and then forecasts the future trend of the input time series in accordance with some volatility measure. Experiments have been carried out on some datasets. Over these datasets, this system achieves accuracy rates well above the baseline accuracy rate, implying statistical significance. The experimental results demonstrate the efficacy of our framework. |
|---|