Get the latest tech news
Time-series forecasting through recurrent topology
Chomiak and Hu introduce a versatile time-series data prediction algorithm using recurring local topological patterning. This algorithm reduces computational complexity and cost, and its feasibility is demonstrated across various dynamic systems including macroeconomic, wearable sensor, and dynamic population systems.
In this section we introduced FReT, an algorithm based on decoding recurrent patterns in a series’ local topology that may offer an effective approach to forecast time-evolving dependencies between data observations in a time-series. To provide additional evidence that FReT can decode important information regarding unseen future events, SETAR, NNET, and D-NNET model data are shown for comparison. It has been shown to be capable of outperforming neural network-based models in revealing digital biomarkers in time-series data 40, and may therefore offer a computational tool to decode topological events that may reflect a system’s upcoming dynamic changes.
Or read this on Hacker News