Abstract The field of time series forecasting has been evolving rapidly with advances in techniques for modelling and forecasting. However, choosing the right technique for a given series is at the heart of forecasting research. This process is challenging because certain forecasting techniques will perform best on some series while different alternative will perform best on other series. Certainly, each forecasting technique has its own territory and dominance. Discovering the conditions under which a forecasting technique will function well and under which not is useful in identifying model territory and dominance.
Abstract A Tool to Detect Potential Data Leaks in Forecasting Competitions Forecasting competitions are of increasing importance as a mean to learn best practices and gain knowledge. Data leakage is one of the most common issues that can often be found in competitions. Data leaks can happen when the training data contains information about the test data. There are a variety of different ways that data leaks can occur with time series data.
Abstract This work presents two feature-based forecasting algorithms for large-scale time series forecasting. The algorithms involve computing a range of features of the time series which are then used to select the forecasting model. The forecasting model selection process is carried out using a pre-trained classifier. In our first algorithm we use a random forest algorithm to train the classifier. We call this framework FFORMS (Feature-based FORecast Model Selection). The second algorithm use efficient Bayesian multivariate surface regression approach to estimate forecast error for each method, and then using the minimum predicted error to select a forecasting model.
Abstract Peeking inside FFORMS: Feature-based FORecast Model Selection Thiyanga S. Talagala$^1$, Rob J. Hyndman$^1$, George Athanasopoulos$^1$ $^1$Department of Econometrics and Business Statistics, Monash University, Australia Features of time series are useful in identifying suitable forecast models. Talagala, Hyndman & Athanasopoulos (2018) proposed a classification framework, called FFORMS (Feature-based FORecast Model Selection), which selects forecast models based on features calculated from the time series. The FFORMS framework builds a mapping that relates the features of a time series to the “best” forecast model using the random forest algorithm.
Abstract This work presents three feature-based algorithms for large-scale time series forecasting. The algorithms are developed based on meta-learning approach. In our first algorithm we use a random forest algorithm to identify the best forecasting model. We call this framework FFORMS (Feature-based FORecast Model Selection). In the second algorithm, FFORMA (Feature-based FORecast Model Averaging), we use gradient boosting to obtain the weights for forecast combinations. The third algorithm use efficient Bayesian multivariate surface regression approach to estimate forecast error for each method, and then using the minimum predicted error to select a forecasting model or to choose individual models for forecast combinations.
Abstract The seer package provides a novel framework for forecast model selection using time series features. We call this framework FFORMS (Feature-based FORecast Model Selection). The underlying approach involves computing a vector of features from the time series which are then used to select the forecasting model. The model selection process is carried out using a classification algorithm – we use the time series features as inputs, and the best forecasting algorithm as the output.
Abstract A crucial task in time series forecasting is the identification of the most suitable forecasting method. We present a general framework for forecast model selection using meta-learning. A Random Forest is used to predict the best forecasting method using only time series features. The proposed framework has been evaluated using time series from the M1 and M3 competitions, and is shown to yield accurate forecasts comparable to several benchmarks and other commonly used automated approaches of time series forecasting.
Abstract Many applications require a large number of time series to be forecast. Providing better forecasts for these time series is important in decision and policy making. However, large scale time series data present numerous challenges in modelling and implementation due to the high dimensionality. It is unlikely that a single method will consistently provides better forecasts across all time series. On the other hand, selecting individual forecast models when the number of series is very large can be extremely challenging.