Fuzzy Mathematics and MicroTEDAClus in Time Series Analysis: Applications in Bitcoin Price Prediction

Abstract

Time series forecasting presents significant challenges for traditional models due to inherent noise, non-stationary patterns, and uncertainty. This study proposes a novel hybrid framework integrating fuzzy logic, dynamic clustering, and LSTM networks to address these challenges. While machine learning approaches like Long Short-Term Memory (LSTM) networks have shown promise, they often struggle with noise, uncertainty, and non-stationary patterns inherent in cryptocurrency markets. This study addresses these limitations by proposing a novel hybrid framework that integrates fuzzy logic, MicroTEDAClus dynamic clustering, and LSTM networks. The framework introduces three key innovations: fuzzy logic-based categorization to handle market uncertainty and linguistic ambiguity, MicroTEDAClus clustering to adaptively segment data streams and capture evolving price regimes, and an LSTM architecture optimized for modelling temporal dependencies in clustered subsets. The methodology, applied to historical Bitcoin price data, underwent rigorous evaluation through preprocessing, dynamic clustering, and multi-phase model training. The experimental results demonstrate the framework’s effectiveness, achieving a Train RMSE of 2687.93, a Test RMSE of 4872.49, a Train MAPE of 2.73%, and a Test MAPE of 4.40%. These findings highlight the potential of hybrid models for improved financial decision-making in volatile markets.