Machine Learning–Based Algorithmic Trading Bots: A Comparative Study of kNN and LVQ Models on Bitcoin Markets

The rapid development of cryptocurrency markets has intensified the use of algorithmic trading systems capable of operating under high volatility and non-stationary market conditions. In this context, machine learning–based trading bots have gained increasing attention due to their potential to adapt dynamically to evolving price patterns. This study investigates the performance of two supervised machine learning classification approaches—k-Nearest Neighbors (kNN) and Learning Vector Quantization (LVQ)—when applied to automated trading in the Bitcoin market. The research addresses the following question: Do adaptive prototype-based learning models outperform instance-based classification models in algorithmic trading applications under identical market conditions? To answer this question, two fully automated trading bots were developed by the authors and implemented on the TradingView platform using Pine Script. Both bots were evaluated under identical experimental settings, trading the BTC/USDT pair on an hourly time frame over the period January 2025 to January 2026, with an initial capital of 1000 USD. Additional market strength and trend filters were applied to enhance signal robustness. The empirical results reveal notable differences between the two learning paradigms. The kNN-based trading bot exhibited a conservative trading behavior, characterized by lower trade frequency and moderate drawdown, but failed to generate positive cumulative returns. In contrast, the LVQ-based trading bot achieved superior profitability and demonstrated a stronger capacity to adapt to changing market regimes, albeit at the cost of significantly higher drawdown and increased risk exposure. The findings suggest that adaptive prototype-based learning models such as LVQ are better suited for algorithmic trading in highly volatile cryptocurrency markets when profitability and adaptability are prioritized. However, the increased risk associated with such models highlights the importance of incorporating robust risk management mechanisms. Overall, this study contributes to the literature by providing a controlled, implementation-level comparison of two machine learning trading strategies using realistic market conditions and fully automated trading systems.