0050 Quantitative Trading Experiment: When Machine Learning Strategies Meet Simple Buy-and-Hold

This open-source project records the failure journey of a quantitative trading experiment. By comparing machine learning strategies with the simple buy-and-hold strategy of the 0050 ETF, it reveals the limitations of complex models in live trading. The experimental results show that the complex strategy not only failed to generate excess returns but also incurred losses, leading the author to eventually return to the simple approach of holding 0050.

The project stems from a classic question in quantitative investing: Can complex algorithms outperform simple market indices? The author initially believed that machine learning models could build market-beating strategies, but the experimental results were the opposite. 0050 is a representative ETF in Taiwan, tracking the Taiwan 50 Index. It has low management fees and good liquidity, serving as a simple way to capture overall market returns—similar to the S&P 500 in the U.S. stock market.

Machine Learning Strategy: Trained on historical data to dynamically adjust positions and rotate assets, using technical indicators and fundamental data as features. It outperformed the benchmark in some stages during backtesting.

Simple Buy-and-Hold Strategy: Buy 0050 at the start of the experiment and hold it long-term without any operations—no complex models or frequent position adjustments.

In backtesting, the machine learning strategy showed excess returns under specific parameters, but live trading performance was the opposite: the ML strategy incurred losses while the 0050 hold strategy maintained positive returns. Reasons for the discrepancy include overfitting, look-ahead bias, transaction costs, and changes in market structure.

Key Indicators: The simple hold strategy had a higher Sharpe ratio (better risk-adjusted return), while the ML strategy had a negative Alpha (underperformed the benchmark).

1. Complexity does not equal better: Market noise and non-stationarity in financial markets make ML predictions difficult; it is wiser to acknowledge the unknown and choose simple strategies.
2. Cost control is important: Fees and slippage from frequent trading erode returns; the passive strategy of 0050 has a clear cost advantage.
3. Behavioral finance perspective: Overconfidence and illusion of control lead to overestimating one’s ability to beat the market.

Implications: Quantitative traders need to emphasize out-of-sample testing and avoid over-optimization; ordinary investors can choose index funds (e.g., 0050); investment education should face failure squarely.

Project Value: The open-source tool provides functions such as transaction record visualization, strategy comparison, and risk assessment. Transparently recording failure experiences has educational significance.

Limitations: The results do not generalize to all quantitative strategies and are based only on Taiwan market data.