Although the attention-LSTM model without the decoding process trained more quickly than the model with the decoding process, the prediction accuracy of the model without the decoding process was not as good as that of the model with the decoding process . Wang et al. proposed a random recursive network called the CLVSA model. Researchers claim that their technique may be used to predict likely variations in raw financial transaction data. Based on deep LSTM and attention mechanisms, Zheng and Xu have developed a financial data forecasting strategy that is based on deep LSTM and attention mechanisms . This study used daily data as well as time-sharing data to investigate the influence of capital flow variations on stock trend changes, with the finding that the self-attention model is enhanced as a consequence. As a result of the experiments, the suggested technique was able to improve the accura

04 percent and gain 6.562 percent in the two-month backtest experiment, demonstrating that the model has a certain level of efficacy and practicability in the prediction of stock price trends. LSTM has encountered a number of obstacles in the area of financial time series prediction. Improving the LSTM model is essential for producing more accurate results in the highly dynamic field of investment forecasting. According to the article, this model is referred to as the LSTM-P model.

3. Bitcoin and Gold Prices Prediction 3.1. Data The data employed in our model for the empirical application consist only of historical price series for Bitcoin and gold, which were collected between September 2016 and September 2021. Gold daily prices (in US dollars per troy ounce) are sourced from the London Bullion Market Association, while Bitcoin daily prices (in US dollars per Bitcoin) are sourced from the Nasdaq Stock Market. Take a look at Figures 1 and 2. Figure 1 Gold daily price. Figure 2 BTC daily price. Note: Bitcoin can be traded every day, so the data are continuous. However, gold has a difference between trading days and nontrading days and the data are not continuous. We first do a smoothed interpolation of the historical price of gold to facilitate the forecasting model. Moreover, we consider the nontrading day scenario in the trading strategy.

3.2. Flow of Our Work Considering the background information and restricted conditions, the article consists of three parts. The first one is data preprocessing part. We use interpolation fitting and wavelet transform noise reduction for Bitcoin and gold historical price data, in order to get higher accuracy in the later time series prediction. Then, we use a modified LSTM-Plus (LSTM-P) neural network for training and prediction. LSTM-P is characterized by keeping only one control gate in the original LSTM model and adding cellular connections to the candidate hidden states and control gate to improve prediction accuracy. At last, we make a sensitivity analysis. In order to avoid complicated descriptions and intuitively reflect our work process, the specific flowchart of the full article can be referred to in Figure 3. Figure 3