Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Boosting Bitcoin Minute Trend Prediction Using the Separation Index

Zeinab Shahsafdari, Ahmad Kalhor

TL;DR

Bitcoin minute-trend forecasting is challenging due to extreme volatility. The authors propose Separation Index (SI), a normalized distance-based separability metric, to rank and selectively retain observations that maximize class separation before training: $SI(Data) = \sum_{i=1}^{m} \delta(l_i, l_{i^*})$, with $i^* = \arg_{q\neq i} \min ||x_i - x_q||^2$. Leveraging minute-level data from Freqtrade on KuCoin (2018–2022) and Bayes-optimized technical indicators plus lag features, they train a BiLSTM for direction and a CNN for magnitude, combined via a voting classifier. The method yields strong test performance, with a maximum of $83.70\%$ accuracy (average $80.23\%$) and shows that SI-guided data selection can closely align with or exceed results obtained from larger, less carefully curated feature sets. The approach demonstrates a practical, data-efficient path toward high-frequency Bitcoin forecasting without relying on fundamental or social-sentiment data, offering meaningful implications for HFT decision-support systems.

Abstract

Predicting the trend of Bitcoin, a highly volatile cryptocurrency, remains a challenging task. Accurate forecasting holds immense potential for investors and market participants dealing with High Frequency Trading systems. The purpose of this study is to demonstrate the significance of using a systematic approach toward selecting informative observations for enhancing Bitcoin minute trend prediction. While a multitude of data collection methods exist, a crucial barrier remains: efficiently selecting the most informative data for building powerful prediction models. This study tackles this challenge head-on by introducing the Separation Index, a groundbreaking tool for fast and effective data (feature) subset selection. The Separation Index operates by measuring the improvement in class separability (i.e. upward vs. downward trends) with each added feature set. This innovative metric guides the creation of a highly informative dataset, maximizing the model's ability to differentiate between price movements. Our research demonstrates the effectiveness of this approach, achieving unprecedented accuracy in minute-scale Bitcoin trend prediction, surpassing the performance of previous studies. This significant advancement paves the way for a new era of data-driven decision-making in the dynamic world of cryptocurrency markets.

Boosting Bitcoin Minute Trend Prediction Using the Separation Index

TL;DR

Bitcoin minute-trend forecasting is challenging due to extreme volatility. The authors propose Separation Index (SI), a normalized distance-based separability metric, to rank and selectively retain observations that maximize class separation before training: , with . Leveraging minute-level data from Freqtrade on KuCoin (2018–2022) and Bayes-optimized technical indicators plus lag features, they train a BiLSTM for direction and a CNN for magnitude, combined via a voting classifier. The method yields strong test performance, with a maximum of accuracy (average ) and shows that SI-guided data selection can closely align with or exceed results obtained from larger, less carefully curated feature sets. The approach demonstrates a practical, data-efficient path toward high-frequency Bitcoin forecasting without relying on fundamental or social-sentiment data, offering meaningful implications for HFT decision-support systems.

Abstract

Predicting the trend of Bitcoin, a highly volatile cryptocurrency, remains a challenging task. Accurate forecasting holds immense potential for investors and market participants dealing with High Frequency Trading systems. The purpose of this study is to demonstrate the significance of using a systematic approach toward selecting informative observations for enhancing Bitcoin minute trend prediction. While a multitude of data collection methods exist, a crucial barrier remains: efficiently selecting the most informative data for building powerful prediction models. This study tackles this challenge head-on by introducing the Separation Index, a groundbreaking tool for fast and effective data (feature) subset selection. The Separation Index operates by measuring the improvement in class separability (i.e. upward vs. downward trends) with each added feature set. This innovative metric guides the creation of a highly informative dataset, maximizing the model's ability to differentiate between price movements. Our research demonstrates the effectiveness of this approach, achieving unprecedented accuracy in minute-scale Bitcoin trend prediction, surpassing the performance of previous studies. This significant advancement paves the way for a new era of data-driven decision-making in the dynamic world of cryptocurrency markets.

Paper Structure

This paper contains 13 sections, 4 equations, 4 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. Using Separation Index for Observation Selection
  4. Bitcoin Minute Data Collection and Preprocessing
  5. Data Collection
  6. Data Preprocessing
  7. Experiments, Results and Discussion
  8. Performing SI on Collected Data
  9. Selected Models for Bitcoin Trend Prediction
  10. BiLSTM – next minute’s close price direction
  11. CNN – next minute’s close price change magnitude
  12. Results and Comparison with Related Works
  13. Conclusion

Figures (4)

  • Figure 1: Expected increasing trend of selected subsets obtained by the algorithm
  • Figure 2: SI values obtained on different subsets of observations – Observation $\#$1: Bitcoin 1-minute historical data, Observation $\#$2: Technical indicators with hyper-optimized parameters, Observation $\#$3: Lagged data on all features and Observation $\#$4: Other cryptocurrencies such as Ethereum, Tether and Litecoin
  • Figure 3: Minute trend prediction flowchart
  • Figure 4: Voting classifier functionality