Innovations In Time Series Forecasting: New Validation Procedures to Improve Forecasting Accuracy and A Novel Machine Learning Strategy for Model Selection

This dissertation is divided into two parts. The first part introduces the p-Holdout family of validation schemes for minimizing the generalization error rate and improving forecasting accuracy. More specifically, if one wants to compare different forecasting methods, or models, based on their performance, one may choose to use “out-of-sample tests” based on formal hypothesis tests, or “out-of-sample tests” based on data-driven procedures that directly compare the models using an error measure (e.g., MSE, MASE). To distinguish between the two “out-of-sample tests” terminologies seen in the literature, we will use the term “out-of-sample tests” for the former and “out-of-sample validation” for the latter. Both methods rely on some form of data split. We call these data partition methods “validation schemes.” We also provide a history of their use with time-series data, along with their formulas and the formulas for the associated out-of-sample generalization errors. We also attempt to organize the different terminologies used in the statistics, econometrics, and machine learning literature into one set of terms. Moreover, we noticed that the schemes used in a time series context overlook one crucial characteristic of this type of data: its seasonality. We also observed that deseasonalizing is not often done in the machine learning literature. With this in mind, we introduce the p-Holdout family of validation schemes. It has three new procedures that we have developed specifically to consider a series’ periodicity. Our results show that when applied to benchmark data and compared to state-of-the-art schemes, the new procedures are computationally inexpensive, improve the forecast accuracy, and greatly reduce, on average, the forecast error bias, especially when applied to non-stationary time series.In the second part of this dissertation, we introduce a new machine learning strategy to select forecasting models. We call it the GEARS (generalized and rolling sample) strategy. The “generalized” part of the name is because we use generalized linear models combined with partial likelihood inference to estimate the parameters. It has been shown that partial likelihood inference enables very flexible conditions that allow for correct time series analysis using GLMs. With this, it becomes easy for users to estimate multivariate (or univariate) time series models. All they have to do is provide the right-hand side variable, the variables that should enter the left-hand side of the model, and their lags. GLMs also allow for the inclusion of interactions and all sorts of non-linear links. This easy setup is an advantage over more complicated models like state-space and GARCH. And the fact that we can include covariates and interactions is an advantage over ARIMA, Theta-method, and other univariate methods. The “rolling sample” part relates to estimating the parameters over a sample of a fixed size that “moves forward” at different “rounds” of estimation (also known as “folds”). This part resembles the “rolling window” validation scheme, but ours does not start at T = 1. The “best” model is taken from the set with all possible combinations of covariates - and their respective lags - included in the right-hand side of the forecasting model. Its selection is based on the minimization of the average error measure over all folds. Once this is done, the best model’s estimated coefficients are used to get the out- of-sample forecasts. We applied the GEARS method to all the 100,000 time-series used in the 2018’s M-Competition, the M4 Forecasting Competition. We produced one-step-ahead forecasts for each series and compared our results with the submitted approaches and the bench- mark methods. The GEARS strategy yielded the best results - in terms of the smallest overall weighted average of the forecast errors - more often than any of the twenty-five top methods in that competition. We had the best results in 8,750 cases out of the 100,000, while the procedure that won the competition had better results in fewer than 7,300 series. Moreover, the GEARS strategy shows promise when dealing with multivariate time series. Here, we estimated several forecasting models based on a complex formulation that includes covariates with variable and fixed lags, quadratic terms, and interaction terms. The accuracy of the forecasts obtained with GEARS was far superior than the one observed for the predictions from an ARIMA. This result and the fact that our strategy for dealing with multivariate series is far simpler than VAR, State Space, or Cointegration approaches shines a light in the future of our procedure. An R package was written for the GEARS strategy. A prototype web application - using the R package “Shiny” - was also developed to disseminate this method