The Establishment Of Effective Portfolio Based On Wavelet Transformation Method And Machine Learning Algorithm

Stock market, as an important part of financial market, has characteristics of low entry barriers and higher liquidity. So, it becomes one of the main investment channels of Chinese people. In this market, institutional investors have lots of advantages, such as with high quality of employees, large amount of funds, and so on. However, in order to attract investors, institutional investors must establish a set of perfect portfolio method. Generally speaking, the advantages of portfolio is in the premise of reducing system risk distribution, obtaining excess returns. Therefore, portfolios built by institutional investors must meet two characteristics: fisrt, robustness:portfolio value cannot be compared with the market index deviated greatly, otherwise difficult to control risk; second, profitable:portfolio can not only simply replicate market index, but also obtain excess returns. This can reflect the value of the portfolio.In the current academic circles, researchers have increasingly favored the use of machine learning methods to process, analyze and forecast the historical data of financial market. However, because the traditional machine learning method can only analysis and predict one sequence in one time, so the efficiency of the model is not high. On the other hand, because the noise content of financial data is large, if use directly, it cannot get the satisfied effect. Although scholars have been using the wavelet analysis method of financial series of decomposition and denoising, but in their use of wavelet, most of them get experience from engineering field, while ignoring the characteristics of financial sequence and denoising demand for predictors.Therefore, the main problem this paper hope to solve is that, in order to build a robust and profitable portfolio, a more effective machine learning prediction model, in the premise of accuracy, should be established.On this basis, according to the characteristics of series of stock--contains a lot of noise, but itself has some volatility, this paper hopes to find a more applicable wavelet denoising method to filter financial sequence, so that it could keep fluctuations containing information as much as possible, and remove the original sequence noise at the same time.In order to solve the problems mentioned above, this paper attempts to apply the matrix decomposition method for recommendation system to analyzing and predicting the price/earnings ratio - date matrix. Would it able to improve the model efficiency by transforming the target from vectors to matrix. At the same time, in the model's input data, this paper tries to use some more moderate wavelet denoising method to process the raw data. In the specific empirical analysis, First of all, this paper uses the matrix decomposition method to forecast the original series stock logarithmic return rate, and chooses the stock portfolio construction according to the forecasting rate of each holding period. In the second, because the value of investment portfolio may not be better than the index, so this paper tries to denoise the raw data by wavelet method. In the final, by forecasting the logarithmic return generated by denoising sequences, this paper builds up portfolios, which have significantly improved values and effects comparing by the former. This indicates that the matrix decomposition prediction method is efficient method and wavelet denoising technology can effectively improve the prediction accuracy.Through empirical analysis, this paper determines the global threshold wavelet denoising method for the optimal denoising method, the shareholding period T=20 and the number of stocks N=20 is the optimal parameters of the investment portfolio. What's more, the prediction model, the denoising model and portfolio has the following characteristics:model are in good condition as a whole, the portfolio's value is relatively stable, the portfolio returns in most of the time is better than the market index, and incurred losses will not occur with larger offset. Secondly, because the advantages of the forecast model is to seize the trend of the sequence, so when the market index rose or fell, the portfolio effect is better, the portfolio's value is usually higher than the index value. Thirdly, as the global threshold wavelet denoising method is more moderate, so when the index is severely fluctuant, the method would filter the noise well and retain the effective wave. On the contrary, when the index is relatively stable, wavelet denoising would filter more effective wave so that the effect of improving portfolio's value would be unsatisfactory. Fourthly, the investment portfolio strategy have slow reaction and tracking on the trend of the index transformation and reverse. This may make the investment portfolio missed part of the gains from the index, but also may avoid some of the risk of decline in the bubble burst.The innovations of this paper are as follows. Firstly, this paper has processed the stock return rate-time matrix by multidimensional matrix decomposition technique, and has predicted the stock return rate of each holding period. This has effectively enhanced the efficiency of machine learning models in the stock market prediction. Secondly, different from other literature on the original sequence of wavelet decomposition will be high-frequency and low-frequency sequence of the same time prediction approach, this paper considers a reasonable reduction of the wavelet coefficients in order to achieve the effect of denoising smoothing. Finally, this paper validates the established portfolio with effectiveness, stability, yield, excellent properties, not only return rate are greater than the index in the most of the time, but also the portfolio's value deviate little from the index when a loss occurs. In addition, because of the reasons for the wavelet denoising processing, the investment portfolio's value is higher than the index, especially in the period of index is rising or falling.