Data Schemas for Forecasting (with Examples in R)

摘要

Evaluation of forecasting performance using real-world data is inevitably connected with the question of how to store actuals and forecasts in a convenient way. The issue gets complicated when it comes to working with rolling-origin out-of-sample forecasts calculated for many time series. This setup can be met in both research tasks (such as forecasting competitions or when some new method is proposed) and in practical settings. When designing data schemas for forecasting it is important to provide access to the information needed for exploratory time series analysis and accuracy evaluation. We found that existing approaches to store forecasting data often cannot be applied efficiently as they are either not flexible enough or they require too much resources to implement and maintain the data storage. Here we propose a flexible yet simple way of keeping forecasting data allowing the storage and exchange of actuals, forecasts, and other relevant information. We also present an R package that helps perform exploratory data analysis and accuracy evaluation based on the data schemas proposed.