Dataset characteristics effect on time series forecasting : comparison of statistical and deep learning models

Sammanfattning: Time series are points of data measured throughout time in equally spaced periods. They present characteristics such as level, noise, trend, seasonality, and outliers. Time series forecasting is the attempt to predict single or multiple future values. It holds significant relevance in numerous fields,including, but not limited to, healthcare, finance, and weather forecasting. It has recently gained more attention due to the COVID-19 pandemic, which highlighted the importance of predicting and managing crises. Two distinct methods of forecasting utilise either statistical or deep learning models, and the debate about the best model is still inconclusive. This thesis aimed to explicate the benefits and drawbacks of each approach pertaining to singlestep and multi-step forecasting. The study applied four models, two of each method, on datasets of varying characteristics and measured their prediction accuracy and computing time. The prediction accuracy of each model was measured using commonly used evaluation metrics, including Root MeanSquare Error. Subsequently, the results were compared with the features of the datasets to identify possible interconnecting relations between the factors. The findings concluded that the deep learning models generally produced a more accurate prediction but required more processing power and computing time. Contrastingly, the statistical models' predictions were less accurate butmarginally faster. Furthermore, the forecast accuracy's most impactful characteristics were the dataset's trend and linearity. The code and datasets were published at: https://github.com/Adam20Taylor/BScThesis