Enhancing Forecast Accuracy: The Impact of Data Transformation in Time Series Models

Abstract

Aim: The aim of the article was formulate suggestions on which preprocessing method is preferable for various forecasting algorithms, including machine learning approaches, particularly for forecasting stock values.

Methodology: Research study on actual stock values prediction on an example of 10 average NYSE enterprises, comparing five scenarios of data preparation.

Results: The results confirm theoretical assumptions and recommendations for the proper design of benchmark studies and real forecasting models.

Implications and recommendations: As stated in the literature of the subject data transformation for models based on stochastic processes, such as ARIMA and GARCH, transforming data to rates of return (a form of differentiation) is a desirable approach. For machine learning models, especially recurrent neural networks, such as the Long Short-Term Memory Network and the Gated Recurrent Unit, the min-max normalisation data transformation should be applied. For exponential Smoothing and Brownian motion methods, the best results were achieved for non-transformed (raw) data. The guidance relevant to benchmark studies and real forecasting models is presented in the final section of the paper. The central thesis was to emphasise, through the example of stock values forecasting, that proper benchmark studies and real-life applications should be designed in a way that ensures proper preprocessing is used for the given model. Using the same preprocessing for different models may sometimes yield misleading results.

Originality/value: The topic of data preparation and transformation, although commonly present in the literature of the subject, is rarely confirmed by research studies on real datasets. To the best of the author's knowledge, this type of analysis has not been conducted on real data to date.