Everyone Can Become Next Buffett: Portfolio Trading Strategies Based on Online Learning and Improved Mean-Variance Model

Gold and Bitcoin are two of the most popular assets in the market, characterized by frequent price fluctuations that complicate accurate predictions and effective investment strategies. In this study, we leverage online learning techniques, combining XGBoost with the FTRL algorithm, to construct a time series model that forecasts daily prices based on five years of trading data (September 11, 2016, to September 10, 2021). Using these forecasted prices, we calculate the daily rate of return for each asset. To optimize the investment ratio, we apply an improved Mean-Variance model that addresses three key deficiencies of the traditional approach, making it more applicable to real-world scenarios. Additionally, we develop two distinct trading strategies tailored for gold trading days and non-gold trading days, respectively. An initial investment of $1,000 grows to $59,854.87 by September 10, 2021. Further analysis utilizing the complete historical data set enables us to maximize returns, achieving a final value of $564,542.95 by the same date. We also construct two comparative models, one focusing solely on gold trading and the other exclusively on Bitcoin. The respective re- turns for these models are $1,086.26 and $8,766.19 on September 10, 2021. The observed similarity in return trends between our model and the full historical data sug- gests that our approach, based on past daily prices, offers a robust investment strategy. To assess the model's sensitivity to commission fee rates, we introduce perturbations and analyze the resultant daily returns and trading behaviors. Finally, we compile our findings into a memorandum to assist the trader in for- mulating more effective portfolio trading strategies.