TL;DR
This paper explores the investment relationship between gold and Bitcoin, proposing a machine learning model to predict price fluctuations and optimize portfolio strategies. The research aims to help traders make informed decisions about buying, holding, or selling these assets, finding that Bitcoin and gold exhibit a strong substitution effect and can be valuable hedging tools.
What Problem Does This Paper Address?
In recent years, Bitcoin has emerged as a significant digital asset, often compared to "digital gold" due to its potential as an inflation hedge and investment vehicle. This has sparked considerable interest in understanding its relationship with traditional safe-haven assets like gold. However, the high volatility and unique characteristics of both assets present a challenge for investors seeking to construct optimal portfolios. This paper addresses the need for a data-driven approach to guide investment decisions in gold and Bitcoin, specifically by developing a model that can predict daily price movements and suggest optimal trading strategies.
Methods
The researchers employed two primary machine learning techniques:
- Linear Regression Prediction: This method was used to model the relationship between input variables (like past price fluctuations and price differences) and output variables (future price changes) for both gold and Bitcoin. The model aims to find a function that best fits the historical data to predict future trends.
- K-Nearest Neighbors (KNN) Algorithm: This algorithm was used to predict feature similarity, aiming to place new data points (predicted prices) in the same category as existing data. It classifies a data point based on the majority class of its 'K' nearest neighbors in the feature space.
The study also involved:
- Data Preprocessing: This included cleaning the data by removing irrelevant entries (e.g., blank lines in gold trading data) and ensuring consistency by aligning Bitcoin data with gold trading days.
- Data Segmentation: The dataset was split into a training set (90%) and a test set (10%) to build and evaluate the predictive models.
- Goodness-of-Fit Test (Chi-square): This statistical test was used to evaluate the accuracy and reliability of the prediction models for both gold and Bitcoin.
- Sensitivity Analysis: The model's robustness was tested by varying the initial investment amount to observe its impact on the final return.
Datasets
The study utilized daily pricing data for gold and Bitcoin from September 10, 2016, to September 10, 2021. The data sources were:
- Gold: London Gold and Silver Market Association (LBMA-GOLD.csv)
- Bitcoin: NASDAQ (BCHAIN-MKPRU.csv)
The paper does not introduce new benchmark datasets.
Key Insights and Ideas
- Bitcoin as a "New Gold": The paper validates the growing perception of Bitcoin as a potential alternative or complement to gold as an investment asset, particularly for hedging against inflation.
- Substitution Effect: A key finding is the strong substitution effect between Bitcoin and gold price volatility. This suggests that as one asset becomes more volatile, the other might become a more attractive hedging option, implying a dynamic relationship that investors can leverage.
- Combined Hedging Power: The research indicates that gold, when combined with traditional currencies, can serve as a suitable hedging strategy. This highlights the enduring role of gold in portfolio diversification.
- Machine Learning for Financial Forecasting: The study demonstrates the practical application of machine learning techniques (Linear Regression and KNN) in predicting financial asset prices and informing investment strategies, offering a data-driven approach beyond traditional qualitative analysis.
Key Results
- Linear Regression Prediction:
- Goodness of fit for gold: 89.44%
- Goodness of fit for Bitcoin: 98.43%
- KNN Algorithm (Test Set Prediction):
- Goodness of fit for gold: 97.25%
- Goodness of fit for Bitcoin: 95.06%
- Model Comparison: The paper notes that while linear regression had higher accuracy for Bitcoin (3.37% higher), KNN was more accurate for gold (7.81% higher). It suggests that linear regression is better for data with obvious linear relationships, while KNN might be preferable otherwise.
- Sensitivity Analysis: The initial investment amount showed a positive correlation with the final return, indicating the model's robustness and reliability.
Limitations and Future Work
- Data Consistency Issues: The paper acknowledges that processing data to maintain consistency (e.g., deleting Bitcoin data on gold shutdown days) might introduce errors into the prediction model.
- Model Simplicity: The authors note that simpler models can affect accuracy and that they adopted a combined model and sensitivity analysis to mitigate this.
- Dimensionality Disaster: The KNN algorithm is susceptible to the "curse of dimensionality," which can become more pronounced with increasing data dimensions.
- Future Work: The authors suggest that the developed model could be applied to other financial domains and emerging industries with potential modifications to the model architecture. They also highlight the need to improve the model to better avoid financial risks in the context of globalizing financial markets.
Summary updated