EXPERIMENTAL ANALYSIS OF THE N-BEATS MODEL FOR MULTI-HORIZON BITCOIN PRICE FORECASTING

Abstract

Cryptocurrency price forecasting is a critical task in financial analysis due to the highly volatile and non-linear nature of the market. Bitcoin, as the cryptocurrency with the largest market capitalization, poses a significant challenge to traditional forecasting models. This study evaluates the performance of the N-BEATS deep learning model, highlighting its novelty in leveraging an independent residual stacking architecture to automatically decompose time series for forecasting Bitcoin prices across multiple time horizons, and compares it with baseline models including CNN, LSTM, GRU, Bi-LSTM, and a hybrid model. Data was collected from Yahoo Finance and processed using a 30-day sliding window technique for forecasting horizons of 1, 3, 5, and 7 days. Model performance was evaluated using MSE, RMSE, MAE, and R² metrics. Quantitative empirical results indicate that at the 1-day horizon, CNN is highly competitive, achieving an MAE of 1957.65 (compared to 1981.51 for N-BEATS). However, as the time horizon increases, N-BEATS demonstrates superior performance and robustness. Specifically, at the longest forecasting horizon (7 days), N-BEATS achieves the lowest error rates (RMSE = 7815.738, MAE = 6184.976) and the highest goodness-of-fit (R² = 0.919365). This demonstrates that N-BEATS achieves higher accuracy across most time horizons and exhibits lower error degradation as the forecasting window is extended.