A novel approach to short time series prediction has emerged, leveraging connectors to improve predictive accuracy and maintain the integrity of periodic characteristics. This innovative methodology, outlined by researchers led by Gao et al., integrates the use of Empirical Mode Decomposition (EMD) with two types of connectors—linear interpolation (LIP) and linear interpolation with random vibration (LRV)—to address the common pitfalls traditionally encountered during the prediction of short datasets.
Time series prediction plays a pivotal role across various sectors, from stock price forecasting to climate modeling. The challenge presented by short series, which often lack sufficient data to enable reliable model training, has long stymied accuracy. Traditional methods tend to rely on concatenation to combine these short series for analysis; yet, this approach creates significant disruptions at the junction points of concatenated datasets, thereby distorting their inherent periodic characteristics. Indeed, as noted by the authors, "Direct concatenation of the short series will disrupt the periodicity and regularity of the nearby data."
To combat these issues, the researchers developed their connector method, which allows for more seamless transitions between series by introducing connectors between them. The LIP connector functions via linear interpolation to smooth over the junctions, successfully reducing fluctuations at the points where datasets previously met; the LRV connector extends this by incorporating random vibrations to maintain the natural variability and regularity of the original series even more effectively.
Data normalization precedes the application of these connectors, ensuring each short series is uniformly evaluated before concatenation. Following this preparation, the combined series undergoes EMD decomposition, which separates the data sequence based on its varying periodic features. By utilizing this structured framework, the resulting sub-sequences align more closely with the characteristics of the original short series.
Experimental results have demonstrated the efficacy of this connector-based model, particularly when applied to real-world datasets such as the Monthly Retail Sales of the USA and several stock price series. The findings indicate improved predictive accuracy when using connectors compared to direct concatenation methods. Specifically, the authors concluded, "The linear and vibrating connector suits those series with obvious periodic characteristics, whereas the simple linear interpolation connector is more appropriate where these characteristics are absent." This flexibility allows various models to be adapted based on the nature of the time series at hand, enhancing the overall predictive capability.
The premise of addressing multi-series datasets through connective methods opens the door to subsequent applications across various scientific and financial domains, where capturing complex patterns remains of utmost importance. By allowing for adaptive connectors like LIP and LRV, the connector method not only fortifies the integrity of data but also reflects its underlying periodicity, offering significant improvements over traditional methodologies.
While promising, this study hints at the importance of future explorations to refine the connector methods utilized and explore their optimal efficacy across diverse datasets. The proposed framework stands as not only a solution to current limitations within short time series prediction but also as an invitation for more nuanced research practices to develop more comprehensive models.