AN INVESTIGATION OF THE IMPACT OF SERIES LENGTH ON FORECAST BEHAVIOUR USING ARTIFICIAL NEURAL NETWORKS

Abstract

This study investigates the impact of series length on forecast behavior using Artificial Neural Networks (ANNs) across three distinct datasets: rainfall, Consumer Price Index (CPI), and Bitcoin prices. By examining different forecasting horizons (3, 6, and 12 lead times), the analysis highlights how ANN performance varies with dataset characteristics and time horizons. Results indicate that ANN models achieved relatively stable and low Mean Squared Errors (MSE) for rainfall forecasts, demonstrating suitability for natural and moderately volatile series in scenario 1. In contrast, CPI forecasts produced higher MSE values, reflecting the model’s sensitivity to structural shifts and gradual economic trends. Bitcoin forecasts showed the greatest variability, with extreme MSE values in some scenarios, emphasizing the challenge of applying ANNs to highly volatile financial series. Notably, ensemble or combined forecasts consistently outperformed individual scenarios in most cases, suggesting that averaging approaches improve robustness and predictive accuracy. The findings underscore that while ANNs are powerful forecasting tools, their effectiveness depends strongly on data properties and forecast horizons. Stable and cyclical datasets benefit most, while volatile financial data require advanced architectures or hybrid models to capture nonlinear patterns effectively. The study recommends further exploration of deep learning techniques such as LSTM and hybrid ensemble frameworks to enhance predictive accuracy, particularly in high-volatility environments. However, the research demonstrates that forecast reliability improves with careful consideration of series length, dataset volatility, and the use of ensemble strategies.