An overview of time series point and interval forecasting based on similarity of trajectories, with an experimental study on traffic flow forecasting

TL;DR

This paper reviews trajectory similarity-based time series forecasting methods, introduces new variations, and compares their performance with standard models like ARIMA and Random Forest using traffic flow data.

Contribution

It presents a general framework for trajectory similarity-based forecasting, proposes new approaches for selecting and assembling similar trajectories, and extends the method to interval forecasting.

Findings

Trajectory similarity methods perform competitively with traditional models.

New approaches improve forecast accuracy and interval prediction.

Traffic flow forecasting benefits from trajectory-based methods.

Abstract

The purpose of this paper is to give an overview of the time series forecasting problem based on similarity of trajectories. Various methodologies are introduced and studied, and detailed discussions on hyperparameter optimization, outlier handling and distance measures are provided. The suggested new approaches involve variations in both the selection of similar trajectories and assembling the candidate forecasts. After forming a general framework, an experimental study is conducted to compare the methods that use similar trajectories along with some other standard models (such as ARIMA and Random Forest) from the literature. Lastly, the forecasting setting is extended to interval forecasts, and the prediction intervals resulting from the similar trajectories approach are compared with the existing models from the literature, such as historical simulation and quantile regression. Throughout the paper, the experimentations and comparisons are conducted via the time series of traffic flow from the California PEMS dataset.