We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Finding the Best Trading Strategy.
- Authors
Yuxian Wang; Yue Jiang; Yiling Chen
- Abstract
Market traders seek to maximize returns through buying and selling. Financial markets are complex and volatile, and it is not easy to do so by relying only on experience to determine trading strategies. Quantitative trading methods can help. To this end, we construct an LM-BP Neural Network Model and a Recurrent Decision Model. The two models together constitute a Quantitative Trading Decision Model for a portfolio of cash, gold, and bitcoin. For the LM-BP Neural Network Model, we use the Levenberg-Marquardt (LM) algorithm based on numerical optimization to improve a traditional back propagation (BP) neural network. We use historical price data for 7 days to make long-term and short-term price forecasts. Subject to trading rules, the long-term forecast for gold is 5 days; that for bitcoin is 7 days. Both shortterm forecasts are for the next day. We test the model with historical price data. The coefficient of determination (R²) between prediction and historical price exceeds .99, and the predictions work well. The core idea of the Recurrent Decision Model is to execute a "buy low, sell high" strategy based on future price trend. It consists of four states: "Stand by," buy, hold, and sell. No decisions are made for the first 10 days, so as to accumulate historical price data. Buy if there is an uptrend in the short- or long-term prediction to a certain threshold. Thresholds for purchase are related to transaction costs and expected returns. The short-term (long-term) thresholds are 2% (3%) for gold and 3.5% (4.5%) for bitcoin. We use the Sharpe ratio to measure the riskiness of a portfolio and to determine the purchase share of each product in the portfolio. Sell if both short-term and long-term forecast prices decline. After model solving, model checking, and sensitivity analysis, we find that our Quantitative Trading Decision Model after a specified five-year trading period would realize an asset value of $270,836 from an initial capital of $1,000. We provide evidence that it makes the best decisions by showing that the model's parameter values are optimal. Also, the investment return is also higher than simple long-term trading, short-term trading, and the strategies of some high-performance investment companies. We find that both our trading strategy and its results are susceptible to changes in the bitcoin commission payment ratio change but less sensitive to that for gold. Finally, we write a memo to traders describing our models, strategies, and results'.
- Subjects
RECURRENT neural networks; PRICES; BACK propagation; TRANSACTION costs; SHARPE ratio
- Publication
UMAP Journal, 2022, Vol 43, Issue 4, p433
- ISSN
0197-3622
- Publication type
Article