Artificial intelligence-based public recruitment fund position penetration method

Abstract

The invention relates to the technical field of data processing in the financial field, in particular to an artificial intelligence-based public recruitment fund position penetration method, which comprises the following steps of: filling and complementing seasonal report disclosure data through fund half-year report disclosure data, and extracting other-class factor information in real time in combination with fund related information; and converting the analysis model into solving a global optimization problem, and calculating the base position ratio of the fund every day, thereby realizing the daily frequency penetration of the large-class, industry and individual stock position conditions of the fund.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology in the financial field, and in particular to a public fund position penetration method based on artificial intelligence. Background Art

[0002] The analysis, penetration and real-time accurate evaluation of the holdings of public funds are the urgent needs of the majority of financial investment users. The analysis of the underlying holdings of public funds is of great significance to investment research and investment consulting. It is conducive to studying the investment style of funds, penetrating the distribution of major asset classes and even individual stocks, so as to analyze the source of fund income, evaluate the asset management ability of fund managers, and effectively improve the efficiency and effect of public fund research. At present, public fund research faces problems such as delayed disclosure of public fund quarterly reports, missing disclosure, opaque information during the disclosure interval, quarterly report whitewashing, information hiding, etc., resulting in large attribution errors in product research performance, untimely risk control warnings, and limited product research depth. With the continuous development of the financial market and the deepening of financial innovation, the regulatory authorities have increasingly stringent requirements for the supervision of the financial market. As an important force in the market, the holdings of public funds have an important impact on the stability and development of the market. At the same time, the market's transparency requirements are also an important driving force for the penetration of public fund holdings. Investors need to understand the investment portfolio and holdings of public funds in order to make more informed investment decisions. Therefore, penetrating management helps to reveal the fund's true holdings, improve market transparency, and protect investors' legitimate rights and interests. Regulatory authorities can also understand the fund's true investment targets and risk control status to ensure the healthy operation of the market.

[0003] In addition, with the continuous development of technologies such as big data and artificial intelligence, technical support has also been provided for the penetration of public fund holdings. By analyzing and mining massive data through technical means, we can more accurately understand the fund's holdings, conduct in-depth research and analysis and risk monitoring on public funds and fund managers, and at the same time improve the efficiency and accuracy of supervision.

[0004] The quarterly reports of public funds only disclose the top ten heavily-weighted stocks. The data disclosed in the quarterly reports can be supplemented through quarterly report completion to obtain estimated quarterly individual stock holdings data. However, due to the long lag in the disclosure of public fund reports and the limited information disclosed in the quarterly reports, the fund holdings data cannot be known during the non-disclosure period, thus limiting the depth of fund research. In addition, investment researchers are constrained by the complexity of factor model construction and find it difficult to make breakthroughs in portfolio optimization, factor research and factor investing.

[0005] In the prior art, a fund research report entitled "Public Fund Holdings Restoration and Its Practical Application" proposes to achieve penetration of fund holdings through three stages: known holdings, manager holdings, and regression fitting. Although this model method has improved the prediction accuracy of the traditional method of normalizing heavy holdings, it is difficult to ensure that the industry ratio of fund holdings is consistent with the actual holding ratio when screening the alternative pool through regression yield. In addition, the model is only established to complete the fund's quarterly report data, and lacks daily penetration of major asset classes, Shenwan first-level industries and individual stocks. Summary of the invention

[0006] In view of the defects of the prior art, the present invention improves the prior art from the following two aspects: first, by scaling the industry when completing the fund quarterly report data, it more accurately fits the actual holdings of the public fund industry; second, the daily frequency penetration of major asset classes, daily frequency penetration of Shenwan first-level industries, and daily frequency penetration of individual stocks are modeled, which solves the problems currently faced by the research on public fund holdings penetration from multiple dimensions. When performing daily frequency penetration of fund major categories, industries, and individual stock holdings, the semi-annual report disclosure data can be directly applied to fill in and complete the quarterly report disclosure data, and alternative factor information can be extracted in real time through public fund related information. The analysis model is transformed into a solution to the global optimization problem, and the daily fund underlying industry holdings are calculated.

[0007] The solution for completing the quarterly report of the present invention is:

[0008] S1. Establish backup positions: Based on the data published in the fund's most recent semi-annual report, divide the holdings published in the semi-annual report into the top ten major holdings and remaining holdings. Based on the data published in the fund's most recent quarterly report, divide the quarterly report holdings into the top ten major holdings and backup positions. Furthermore, all holdings in the above semi-annual report that are not among the top ten major holdings in the quarterly report are inherited into the above backup positions.

[0009] S2, weight adjustment: based on the rise and fall of each stock from the latest semi-annual report to the latest quarterly report, recalculate the new weight of the reserve holdings after the rise and fall. At the same time, the above weights must be adjusted to be less than the smallest weight among the top ten weights in the quarterly report;

[0010] S3, supplement reserve holdings: determine whether there are new stocks that can be included in the reserve holdings through the restricted stock holdings data disclosed in the quarterly report or the list of the top ten circulating shareholders published by the listed company. If there are new stocks in the reserve holdings, their weight in the reserve holdings will be determined based on the public information in the quarterly report or the number of shares held and the stock price;

[0011] S4, industry alignment: According to the industry distribution data of the China Securities Regulatory Commission (CSRC) disclosed in the fund quarterly report, the weight of the reserve holdings is scaled so that the sum of the weights of the individual stocks belonging to a certain industry is equal to the allocation ratio of the industry disclosed in the fund quarterly report. The weights of individual stocks can be scaled proportionally. If the weight of individual stocks needs to exceed the top ten heavy holdings to fill the industry weight after scaling, other stocks in this industry will be selected to fill the weight;

[0012] S5, splicing verification: splice the top ten holdings in the quarterly report with the backup holdings calculated through the above steps to obtain all the calculated quarterly report holdings, and then verify whether they are consistent with the public information, including whether the top ten holdings and proportions are the same, whether the industry distribution is the same, and whether the total holding positions are the same.

[0013] The daily frequency solution for major asset classes of the present invention integrates product net asset value data, quarterly, semi-annual, annual report disclosure data, major asset class data, etc., to provide key basic information for the application of major asset class penetration to the major asset allocation of public funds. The goal is to predictably distribute the holdings of the funds to be penetrated across six major asset classes: A-shares, Hong Kong stocks, U.S. stocks, cash, gold, and bonds (if there are no data restrictions, major asset classes such as commodities and overseas bonds can be expanded based on the provided data).

[0014] The model design framework is to divide the target funds to be penetrated into passive funds and active funds, among which passive and active funds are further subdivided into index ETF connection funds, other passive funds, etc., and the rule engine is designed to calculate the distribution results of passive major asset classes in combination with the details of regulatory rules and the actual investment behavior of passive funds; for active funds, the fitting errors of product daily yield and major asset daily yield and the errors between the prediction results and the previous quarterly report are comprehensively considered, and the Bayesian method is used to optimize and minimize the two errors to obtain the final major asset allocation prediction results. In the major asset model, the risk exposure is calculated. In actual investment research, the actual risk exposure is what investors need to consider. Therefore, the model is more in line with actual investment logic and application than disclosure.

[0015] The specific operations are:

[0016] S1, estimate model hyperparameters, including net value rough fitting variance parameter, product position change variance parameter, and yield fitting term time weight decay parameter:

[0017] The net value rough fitting variance parameter is obtained by regression calculation of fitting residuals according to the following formula:

[0018] r it =X t B it +α i +ε it , t∈M

[0019]

[0020] where r it is the daily return of target fund i at time t, X t is the industry fitting weight, it is the daily return rate of target fund i industry, α i is the intercept corresponding to fund i, ε it is the fitting residual of the target fund at time t, M is the time window, is the rough fitting variance of the net value of target fund i at time t;

[0021] The product position change variance parameter is calculated by the following formula:

[0022]

[0023] where η i,j,t＝T is the variance parameter of the target fund i’s holdings in industry j at time t, S i is the number of quarterly reports published by target fund i since its establishment, m i,j,s is the industry holding weight of target fund i in industry j in the Sth published quarterly report period;

[0024] The time weight decay parameter of the yield fitting term is calculated by taking the coefficients from the 1st to the Mth term in the formula, that is, {λ,λ(1-λ),λ(1-λ) 2 ,…,λ(1-λ) M-1}, and normalize the above coefficients, and record the normalized coefficients as γ t＝T ={γ T-1 ,γ T-2 ,…,γ T-M}, the calculation formula is:

[0025]

[0026]

[0027] Where λ is the EWMA parameter and HL is the half-life.

[0028] S2, solve the optimal solution of the following function to get the weights of major asset classes:

[0029]

[0030] where w i,t＝T is the industry weight of the target fund at i at time T, w i,t＝T The transpose of index,t is the industry index return at time t.

[0031] In the existing technology, daily industry penetration can be based on the fund's daily net value curve and daily industry index data. These two data can be used to perform regression fitting on the fund's net value to obtain the weight of each industry, that is, the holding ratio. However, the daily holding penetration estimated using only the above two data is not ideal. Research has found that the accuracy can only reach 60%.

[0032] The solution for daily frequency penetration of the industry in the present invention is:

[0033] S1, model data pre-verification: based on the fund calculation date, obtain the quarterly report announcement date that is less than the fund calculation date, and take the reporting period corresponding to the quarterly report announcement date;

[0034] S2, estimate hyperparameters, including the rough fitting variance parameter of net value, the variance parameter of product position change, and the time weight decay parameter of yield fitting item:

[0035] The calculation method of the net value rough fitting variance parameter is the same as above;

[0036] The calculation formula of the variance parameter of product position change is the same as above. In addition, to allow the fund to add industries that have not been historically held, η i,j,t＝T Sort from largest to smallest, keep the largest 5 unchanged, and remember that π is equal to all the remaining η i,j,t＝T The average value of η i,j,t＝T <π, then let η i,j,t＝T =π;

[0037] The calculation method of the time weight decay parameter of the yield fitting term is the same as above.

[0038] S3, set the prior of holding weight: suppose the current model calculation date is T, the latest model penetration result calculation date is T1, and the latest model penetration result is The current model calculates the daily difference T m =T-T1, take the latest quarterly industry weight announcement date as T2, the latest quarterly announcement date as m i,j,s＝T2 The difference between the current model calculation date and the quarterly report release date is T p =T-T2, if the model results exists, and T p ≤T m , then prior=m i,j,s＝T2 ,otherwise If the model results If it does not exist, then prior=m i,j,s＝T2 ;

[0039] S4, solve the equation to get the optimal solution to obtain industry distribution:

[0040]

[0041] st

[0042] ∑ j w i，j，t＝T =1;

[0043] S5, post-processing of calculation results:

[0044] For the single-day model running results of a single fund, if the industry weight is less than 0.02%, it is set to 0. After normalizing the single fund single-day model results and the actual disclosed weight results, the same rate as the previous period weight calculation is calculated, and the results are printed in the log. The calculation process is as follows:

[0045]

[0046] The historical time series of a single fund is processed as follows:

[0047] a. Smoothing for each fund and each industry

[0048] i. For historical data, calculate the EWMA with a half-life of 5 days for each industry;

[0049] ii. For daily new data, the calculation formula is:

[0050]

[0051] in The model result is the result of EWMA calculation of target fund i in industry j on day T-1 and stored in the database. The newly added model results for target fund i in industry j on day T;

[0052] b. For a single fund or a single industry historical time series, the starting date T1 and ending date T2 of the grouping are the reporting dates of two consecutive quarterly reports. The daily frequency penetration results within the interval are combined into one group; from the most recent "quarterly report release" date to the present, a separate group is formed;

[0053] i. Define each group: G i，j，g =[w i，j，t＝T1 ,···,w i，j，t＝T2-1 ] T , the number of N in each group is T2-T1, and each group is processed as follows:

[0054] ii. Let w i，j，t =0.00001, if w i，j，t =0&t∈[T1,...,T2], for each group G i，j，g , if the number of groups is greater than 11 and Do the following:

[0055] Yi，j，t =|(w i，j，t -w i，j，t-1 ) / w i，j，t-1 |

[0056] Y i，j，t =[Y i，j，t＝T1 , ..., Y i，j，t＝T2-1 ]

[0057] Y i，j，median =median(Y i，j，m )

[0058] If: Y i,j,median ≤0.006&i≠bond, take the target fund i, industry j, the closing price of the corresponding index of the Shenwan first-level industry in the period from T1 to T2-1, and calculate the cumulative rate of return based on the closing price. i,j,t

[0059] R i，j,m =[r i，j,t＝T1 ,…,r i,j,t＝T2-1 ]

[0060] Gi ，j,g =G i，jg *((R i,j,m -1)*0.4+1)

[0061] iii.w i,j,t =0, if w i,j,t =0.00001&t∈[T1,…,T2]

[0062] iv. If conditions ii and iii are not met, the original data will be returned directly;

[0063] c. Reset the final weight to 1:

[0064] W i，j，t / (∑ j (W i,j,t )

[0065] The individual stock penetration model of the present invention is mainly used to solve the problems of low data update frequency, insufficient accuracy, low processing efficiency and lack of real-time monitoring and optimization in the existing industry penetration technology of equity-oriented funds. Traditional penetration methods rely on quarterly reports, which leads to information lag and cannot reflect market changes in a timely manner. In addition, these methods are usually unable to effectively process large amounts of data and are easily affected by missing data and outliers during the processing process, thereby reducing the accuracy and reliability of the penetration results. The present invention provides an efficient, accurate and real-time updated daily frequency industry penetration method for equity-oriented funds, which can improve the accuracy and efficiency of fund investment analysis. The specific solutions of the present invention are:

[0066] S1, model data preprocessing, the data input to the model is processed as follows:

[0067] a. Based on the fund calculation date, obtain the available quarterly report date that is less than the fund calculation date, and take the reporting period corresponding to the quarterly report date;

[0068] b. If the fund rate of return and index rate of return is greater than 20%, the previous period will be used to fill the gap;

[0069] c. For Hong Kong stocks, only the time points of A-share trading days are taken, and the daily increase or decrease data of A-share trading days but Hong Kong stocks non-trading days are filled in as 0;

[0070] d. Ensure that there are no missing values ​​in the input model data;

[0071] S2, establish a candidate stock pool:

[0072] a. Funds with less than four quarterly report disclosure periods will be excluded from the fund pool for individual stocks to be penetrated;

[0073] b. Obtain the full disclosure of holdings in the fund's historical semi-annual and annual reports;

[0074] c. Obtain the holdings after the historical quarterly reports are completed;

[0075] d. Use the asset allocation table disclosed in the latest quarterly report and the disclosed weights of equity as the weights of the fund’s equity holdings;

[0076] e. Organize a pool of candidate targets for each fund

[0077] i. Use the latest quarterly report to complete the position results. If the semi-annual report or annual report corresponding to the latest quarterly report has been disclosed, use the latest semi-annual report or annual report to fully disclose the position. For example, assuming that the quarterly report on June 30 is disclosed on July 20 and the semi-annual report is disclosed on August 25, then before July 20, use the first quarterly report to complete the position results. From July 20 to August 25, use the second quarterly report to complete the position results. After August 25, use the semi-annual report.

[0078] ii. Individual stocks whose holding weights account for more than 1.0% of the equity holding weights will be included in the candidate target pool, and the holding ratio of the individual stock will be recorded at the same time (note that the holding ratio here refers to the proportion of net assets, not the proportion of equity);

[0079] iii. If the candidate pool is less than 30 stocks, 30 stocks will be added from the remaining disclosures in descending order of weight, and the holding ratio of the stock will be recorded at the same time;

[0080] iv. If the candidate pool exceeds 50 stocks, only the 50 stocks with the largest weights will be retained;

[0081] S3, first layer modeling:

[0082] a. Configure the calculation time window, i.e. the most recent M trading days, with the default value of M = 44;

[0083] b. Take one of the following two options to calculate the residual sequence ε of the return on equity assets it :

[0084] i. Option 1: Do not subtract A shares and Hong Kong stocks

[0085]

[0086] ii. Option 2: Subtract A shares and Hong Kong stocks

[0087]

[0088] where r t,mkt,A is the return rate of CSI 300 Index, r t,mkt,HK is the Hang Seng Index return, w i,t＝T,A-sahre is the A-share asset weight result of fund i on day t, w i,t＝T,HK-share is the asset weight result of Hong Kong stocks of fund i on day t.

[0089] S4, second layer modeling:

[0090] a. Estimate the rough fitting variance parameters of the net value (same method as above);

[0091] b. Estimate the variance parameters of the changes in the holdings of individual stocks of a given fund;

[0092] i. First disclosure of fund i m i,s＝1 (The disclosure date is t=s). The first disclosure can be a semi-annual report, an annual report, or a quarterly report after the stock holdings are completed. Calculate the average holding weight of all "other funds" of each stock j in the candidate target pool on the disclosure date t=s-1, recorded as m i,j,s＝0 , each stock in the candidate target pool is recorded separately;

[0093] ii. Calculate the variance of the change in holdings of a given fund product on that day (t = T):

[0094]

[0095] iii. To allow the fund to add industries that it has not held in the past, η i,j,t＝T Sort from largest to smallest, keep the 10 largest ones unchanged, and remember that π is equal to all the remaining η i,j,t＝T If π = 0, then let π be equal to the smallest non-zero η i,j,t＝T ; if η i,j,t＝T <π, then let η i,j,t＝T =π.

[0096] c. Calculate the hyperparameters of a given fund stock:

[0097] j∈Fund i candidate stock target pool

[0098] d. Calculate the hyperparameters of the Shenwan industry of a given fund, and complete the hyperparameters of the Shenwan industries that are not covered by the candidate stock target pool. The completion method is the same as the hyperparameter completion method for individual stocks:

[0099] j∈Fund i candidate stock target pool, k∈Shenwan industry

[0100] e. Estimate the time weight decay parameter of the yield fitting term:

[0101] i. Half-life HL days, default HL = 44 days, when HL is positive infinity, time weighting degenerates to equal weight;

[0102] ii. Calculate EWMA parameters:

[0103]

[0104] iii. Calculation:

[0105]

[0106] iv. Take out the coefficients from the 1st to the Mth term in the previous step, that is, {λ,λ(1-λ),λ(1-λ) 2 ,…,λ(1-λ) M-1}, then normalize these M coefficients so that each of them is greater than zero and their sum is 100%, and record the normalized coefficient as γ t＝T ={γ T-1 ,γ T-2 ,…,γ T-M}

[0107] f. Set the prior of position weight

[0108] i. Assume that the current model calculation date is T, the latest model penetration result calculation date is T1, and the latest model penetration result is The current model calculates the daily difference T m =T-T1, take the industry weight report date disclosed in the latest quarterly report as T2, and the latest quarterly report disclosure as m i,j,s＝T2 , the current model calculation date is the difference between the quarterly report release date and T p =T-T2;

[0109] ii. If the model results Exists: If T p ≤T m , Tp ≠0, otherwise

[0110] iii. If the model results Does not exist:

[0111] g. Calculate the proportion of individual stock and industry holdings to the net value of equity assets (note that this is not the proportion of total net assets), and jointly calculate the residual of the return on equity assets in Scheme 1 and Scheme 2, and use them with Scheme 1 and Scheme 2 here:

[0112] Option 1: Daily return of individual stocks minus market, Hong Kong stocks

[0113]

[0114] Option 2: The daily return of individual stocks does not reduce the market

[0115]

[0116] Optimize the following problems and record the optimal solution

[0117] The optimization function is as follows:

[0118]

[0119] st

[0120]

[0121] w i,k,t *(w i,k,t -0.002)≥0

[0122]

[0123] where w i,k,t is the penetration weight result of individual stock j of fund i on day t, and the sum of individual stock weight vectors can be obtained w i,j,t is the penetration weight result of individual stock j of fund i on day t, and the sum of individual stock weight vectors can be obtained

[0124] S5, the third layer modeling:

[0125] a. Given a fund i in a certain industry k, if w in the above results i,k,t＝T >0, the industry needs to be further mapped back to stock holdings;

[0126] b. Obtain historical holdings, obtain the holdings data of fund i in the above candidate stock pool, and sort out the list of individual stocks belonging to Shenwan Industry k;

[0127] c. Obtain research shares:

[0128] i. Get the list of fund managers of fund i;

[0129] ii. For each fund manager at time t = T, use the database to sort out all the stocks that he has researched in the past year;

[0130] iii. Summarize all the researched stocks and compile a list of stocks belonging to Shenwan Industry K;

[0131] d. Combine the historical holdings and research stocks, remove duplicates, and then remove the candidate target pool stocks. The cleaned stock list is recorded as KH and KM stocks respectively;

[0132] e. If KH+KM == 0:

[0133] i. If the new platform stocks meet condition 1 (stocks of the same industry in the platform stocks are sorted from large to small by taking the log of market value, and weights are allocated according to market value ratio, and the largest weight is less than the smallest weight of the top 10 heavily-weighted stocks), then the estimated weight w of fund i in industry k is i,k,t＝T Delete and estimate the weight w of fund i in industry k i,k,t＝T Allocate to stocks in the same industry k as the platform stock;

[0134] ii. If the same industry in the platform stocks cannot meet the above condition 1, then obtain the stocks in the corresponding industry as a supplement, so that the sum of the platform stocks and the corresponding industry stocks can meet the above condition 1, and then estimate the weight w of fund i in industry k i,k,t＝T The sum of the allocation to platform stocks and corresponding industry stocks;

[0135] f. If KH+KM==1, calculate the estimated weight w of industry k in fund i i,k,t＝T Delete and add the stock at the same time, and the estimated weight of the stock is w i,k,t＝T ;

[0136] g. If KH+KM>1, then

[0137] i. Perform constrained linear regression (M days) on the index return of industry k for the above KH+KM stocks. The maximum weight shall not be greater than the minimum weight of the top 10 holdings, that is,

[0138]

[0139] st

[0140] The intercept is 0

[0141]

[0142] ii. The final weight of each stock in KH+KM is w i,k,t＝T ·[w kh ,,w kh ] T

[0143] S6, result processing after model calculation:

[0144] a. For historical data: for each fund and each stock, perform EWMA calculation with half-life HL = 5 days;

[0145] b. For daily new data: For the current fund i, the new "model result" on day T is After calculating the EWMA of fund i on day T-1, the model result stored in the database is

[0146] BRIEF DESCRIPTION OF THE DRAWINGS

[0147] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in describing the embodiments.

[0148] Figure 1 Shown is the overall flow chart of the present invention.

[0149] Figure 2 A flow chart of the quarterly report completion of the present invention is shown.

[0150] Figure 3 A flow chart showing the daily frequency penetration of major asset classes of the present invention.

[0151] Figure 4 A flow chart showing the daily frequency penetration of industries in the present invention.

[0152] Figure 5 A flow chart of daily frequency penetration of individual stocks according to the present invention is shown. DETAILED DESCRIPTION

[0153] The embodiments will now be described more fully with reference to the accompanying drawings. The embodiments can be implemented in various forms and should not be understood as being limited to the embodiments. The purpose of the embodiments is to make the present application more comprehensive and complete and to fully convey the concept of the present invention to those skilled in the art.

[0154] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and steps, nor must they be executed in the order described. For example, some steps can be decomposed, and some steps can be combined or partially combined, so the actual execution order may change according to the actual situation.

[0155] Figure 1 The overall technical solution according to the present invention is shown, wherein the quarterly tasks include completing the quarterly report according to the semi-annual report and obtaining the industry distribution after the quarterly report is completed, and the daily tasks include obtaining the Shenwan industry distribution of the index, Shenwan first-level industry penetration, individual stock penetration, Shenwan second-level industry penetration and major asset penetration.

[0156] like Figure 1 As shown, the present invention first completes the quarterly report, and then performs daily frequency penetration of major asset classes, or performs daily frequency penetration of industries or daily frequency penetration of individual stocks based on the industry distribution obtained after completing the quarterly report and combining it with the industry distribution of the index. When performing daily frequency penetration of individual stocks, the results of daily frequency penetration of major asset classes are also needed.

[0157] When completing the quarterly report, Figure 2 As shown, first, the quarterly report backup holdings are established by position inheritance, and all stocks in the most recent semi-annual report are inherited to the most recent quarterly report. For the remaining holdings (outside the top ten) in the semi-annual report, the stocks that have appeared in the top ten of the quarterly report are removed. In addition, the stocks that belong to the top ten of the semi-annual report but not the top ten of the quarterly report must also be included to constitute the quarterly report backup holdings. When adjusting the weights of the stocks in the backup holdings in S22, the method of proportional reduction can be adopted, or the proportion of the weight exceeding the minimum weight of the top ten in the quarterly report can be directly adjusted to the minimum weight of the top ten. When supplementing the backup holdings according to shareholder information in S23, it is considered that if a very few funds have changed their full names, and the listed companies still disclose shareholders according to the old fund names, if the disclosed quarterly report and the top ten circulating shareholders exist at the same time, the quarterly report disclosure is retained, and only the quarterly report is not spliced ​​and filled, so as to avoid missing this information due to mismatch. When performing industry alignment in S24, if the individual stocks cannot fill the industry weight after scaling, and the individual stock weight needs to break through the top ten to meet the industry weight alignment, other individual stocks in the industry will be selected to fill in. According to the direction of the fund company's collective entry and exit of platform stocks during the same period, new platform stocks can be given priority as fill-in. If there are no platform stocks, select the top one-third of the industry's market value from the entire market as a substitute. After industry alignment, it should also be noted that the total position of the reserve holdings is the total position of the fund minus the total position of the top ten. If there is a difference in the data, the weight of the reserve holdings needs to be re-scaled to ensure that the positions are aligned. In addition, for funds holding Hong Kong stocks, due to the different methods adopted by various fund companies for the current industry classification of Hong Kong stocks, some use GICS classification, and some use Bloomberg classification, individual funds need to be analyzed individually. When performing splicing verification in S25, annotate the source behind each stock, including "quarterly report disclosure", "circulating shareholder disclosure", "restricted sale disclosure", "platform stock filling", "other filling", etc.

[0158] Figure 3 The method for constructing the daily frequency penetration model of major asset classes of the present invention is shown. After the quarterly report of the fund is completed, the major asset classes are penetrated. The model is run daily and all types are penetrated, but it is necessary to exclude funds with less than 1 quarterly report, terminated funds and funds with weekly net value disclosure. All funds are divided into passive and other types. If the fund has a parent fund code, only the parent fund code is retained when running the model. The penetration categories mainly include: A shares, bonds, currencies, gold, Hong Kong stocks, and US stocks.

[0159] like Figure 3 As shown, when penetrating the fund's major asset classes, it includes constructing the objective function and calculating the optimal solution of the function. First, the model hyperparameters are estimated, including the net value rough fitting variance parameter, the fund product position change variance parameter and the yield fitting term time weight decay parameter. In step S31, the fitting residual is first calculated by regression according to the following formula:

[0160] r it =X t Β it +α i +ε it , t∈M

[0161] Then calculate the net value rough fitting variance parameter:

[0162]

[0163] In S32, to estimate the variance parameter of the holdings change of a given fund product, calculate the variance of the holdings change of a given fund product on that day (t=T), take the difference, square and mean of each disclosure and each asset weight, and the calculation formula is:

[0164]

[0165] In S33, the yield fitting time term decay parameter is calculated according to the following formula:

[0166]

[0167] And we need to extract the coefficients from the 1st to the Mth term in the above formula, that is, {λ,λ(1-λ),λ(1-λ) 2 ,…,λ(1-λ) M-1}, and then normalize these M coefficients, where the EWMA parameters Finally, in S34, the weight is calculated according to the following formula:

[0168]

[0169] When conducting daily industry penetration, Figure 4As shown, the construction of the model includes three steps. In S41, if the annual report and semi-annual report are released, the annual report and semi-annual report are used to replace the completed quarterly report. In the fund quarterly report weight, the industry weight of the industry with a share of less than 0.00001 is set. For the fund yield and the date when the index yield is greater than 20%, the net value is filled with the previous period. The other steps for estimating the model hyperparameters are repeated with the daily frequency penetration of major asset classes, so they will not be repeated here. After estimating the model hyperparameters, when setting the prior of the position weight, set the current model calculation date to T, the most recent model penetration result calculation date to T1, and the most recent model penetration result to T2. The current model calculates the daily difference T m =T-T1, take the latest quarterly industry weight announcement date as T2, the latest quarterly announcement date as m i,j,s＝T2 The difference between the current model calculation date and the quarterly report release date is T p =T-T2, if the model results exists, and T p ≤T m , then prior=m i,j,s＝T2 ,otherwise If the model results If it does not exist, then prior=m i,j,s＝T2 ; Then find the optimal solution according to the following formula to get the industry distribution:

[0170]

[0171] st

[0172] ∑ j w i,j,t＝T =1.

[0173] As shown in S43, the model results need to be processed at last. For the single-day model running results of a single fund, if the industry weight is less than 0.02%, it is set to 0. After normalizing the single fund, single-day model results, and the actual disclosure weight results, the same rate as the weight calculation of the previous day is calculated, and the results are printed in the log. The calculation process is as follows:

[0174]

[0175] The historical time series of a single fund is processed as follows:

[0176] a. Smoothing for each fund and each industry

[0177] i. For historical data, calculate the EWMA with a half-life of 5 days for each industry;

[0178] ii. For daily new data, the calculation formula is:

[0179]

[0180] in The model result is the result of EWMA calculation of target fund i in industry j on day T-1 and stored in the database. The newly added model results for target fund i in industry j on day T;

[0181] b. For a single fund or a single industry historical time series, the starting date T1 and ending date T2 of the grouping are the reporting dates of two consecutive quarterly reports. The daily frequency penetration results within the interval are combined into one group; from the most recent "quarterly report release" date to the present, a separate group is formed;

[0182] i. Define each group: G i，j，g =[w i，j，t＝T1 ,…,w i，j，t＝T2-1 ] T , the number of N in each group is T2-T1, and each group is processed as follows:

[0183] ii. Let w i，j，t =0.00001, if w i，j，t =0&t∈[T1,...,T2], for each group G i，j，g , if the number of groups is greater than 11 and Do the following:

[0184] Y i，j，t =|(w i，j，t -w i，j，t-1 ) / w i，j，t-1 |

[0185] Y i，j，t =[Y i，j，t＝T1 , …, Y i，j，t＝T2-1 ]

[0186] Y i，j，median =median(Y i，j，m )

[0187] If: Y i，j，median ≤0.006&i≠bond, take the target fund i, industry j, the closing price of the corresponding index of the Shenwan first-level industry in the period from T1 to T2-1, and calculate the cumulative rate of return based on the closing price. i，j，t

[0188] R i，j，m =[r i，j，t＝T1 ,…,r i，j，t＝T2-1 ]

[0189] G i，j，g =G i，j，g *((R i，j，m-1)*0.4+1)

[0190] iii.w i，j，t =0, if w i，j，t =0.00001&t∈[T1,...,T2]

[0191] iv. If conditions ii and iii are not met, the original data will be returned directly;

[0192] c. Reset the final weight to 1:

[0193] w i，j，t / (∑ j (w i，j，t ).

[0194] When performing daily penetration of individual stocks, Figure 5 As shown, the construction of the model includes 5 steps.

[0195] When preprocessing the model data in S51, the following operations are performed:

[0196] a. Based on the fund calculation date, obtain the available quarterly report date that is less than the fund calculation date, and take the reporting period corresponding to the quarterly report date;

[0197] b. If the fund rate of return and index rate of return is greater than 20%, the previous period will be used to fill the gap;

[0198] c. For Hong Kong stocks, only the time points of A-share trading days are taken, and the daily increase or decrease data of A-share trading days but Hong Kong stocks non-trading days are filled in as 0;

[0199] d. Ensure that there are no missing values ​​in the input model data.

[0200] When establishing a candidate target pool in S52, the specific operations are as follows: a. Eliminate funds with less than four quarterly report historical disclosures from the fund pool for individual stocks to be penetrated; b. Obtain the full holdings disclosed in the fund's historical semi-annual and annual reports; c. Obtain the holdings after the historical quarterly holdings are completed; d. Use the asset allocation table disclosed in the latest quarterly report and the equity disclosure weight as the equity holding weight of the fund, where the asset allocation table disclosed in the latest quarterly report is the Juyuan table MF_AssetAllocationNew, and the word The segment is RINOfStock, e. Arrange the candidate target pool for each fund, i. Use the latest quarterly report to complete the holding results. If the semi-annual report or annual report corresponding to the latest quarterly report has been disclosed, then use the latest semi-annual report or annual report to fully disclose the holdings. For example, assuming that the quarterly report on June 30 is disclosed on July 20 and the semi-annual report is disclosed on August 25, then before July 20, use the first quarterly report to complete, from July 20 to August 25, use the second quarterly report to complete, and after August 25, use the semi-annual report. ii. Individual stocks whose holding weight accounts for more than 1.0% of the equity holding weight will be included in the candidate target pool, and the holding ratio of the individual stock will be recorded (note that the holding ratio here refers to the proportion of net assets, not the proportion of equity); iii. If the candidate target pool is less than 30 stocks, 30 stocks will be added from the remaining disclosures in descending order of weight, and the holding ratio of the individual stock will be recorded; iv. If the candidate target pool exceeds 50 stocks, only the 50 stocks with the largest weight will be retained.

[0201] When the first layer modeling is performed in the next S53, the specific operations are as follows:

[0202] a. Configure the calculation time window, i.e. the most recent M trading days, with the default value of M = 44;

[0203] b. Take one of the following two options to calculate the residual sequence ε of the return on equity assets it :

[0204] i. Option 1: Do not subtract A shares and Hong Kong stocks

[0205]

[0206] ii. Option 2: Subtract A shares and Hong Kong stocks

[0207]

[0208] where r t,mkt,A is the return rate of CSI 300 Index, r t,mkt,HK is the Hang Seng Index return, w i,t＝T,A-sahre is the A-share asset weight result of fund i on day t, w i,t＝T,HK-share is the asset weight result of Hong Kong stocks of fund i on day t.

[0209] Next, the second layer modeling is performed as shown in S54, and the specific operations are as follows:

[0210] a. Estimate the rough fitting variance parameters of the net value (same method as above);

[0211] b. Estimate the variance parameters of the changes in the holdings of individual stocks of a given fund;

[0212] i. First disclosure of fund i m i,s＝1 (The disclosure date is t=s). The first disclosure can be a semi-annual report, an annual report, or a quarterly report after the stock holdings are completed. Calculate the average holding weight of all "other funds" of each stock j in the candidate target pool on the disclosure date t=s-1, recorded as m i,j,s＝0 , each stock in the candidate target pool is recorded separately;

[0213] ii. Calculate the variance of the change in holdings of a given fund product on that day (t = T):

[0214]

[0215] iii. To allow the fund to add industries that it has not held in the past, η i,j,t＝T Sort from largest to smallest, keep the 10 largest ones unchanged, and remember that π is equal to all the remaining η i,j,t＝T If π = 0, then let π be equal to the smallest non-zero η i,j,t＝T ; if η i,j,t＝T <π, then let η i,j,t＝T =π;

[0216] c. Calculate the hyperparameters of a given fund stock:

[0217] j∈Fund i candidate stocks pool;

[0218] d. Calculate the hyperparameters of the Shenwan industry of a given fund, and complete the hyperparameters of the Shenwan industries that are not covered by the candidate stock target pool. The completion method is the same as the hyperparameter completion method for individual stocks:

[0219] j∈Fund i candidate stock target pool, k∈Shenwan industry;

[0220] e. Estimate the time weight decay parameter of the yield fitting term:

[0221] i. Half-life HL days, default HL = 44 days, when HL is positive infinity, time weighting degenerates to equal weight;

[0222] ii. Calculate EWMA parameters:

[0223]

[0224] iii. Calculation:

[0225]

[0226] iv. Take out the coefficients from the 1st to the Mth term in the previous step, that is, {λ,λ(1-λ),λ(1-λ) 2 ,…,λ(1-λ) M-1}, then normalize these M coefficients so that each of them is greater than zero and their sum is 100%, and record the normalized coefficient as γ t＝T ={γ T-1 ,γ T-2 ,…,γ T-M};

[0227] f. Set the prior of position weight

[0228] i. Assume that the current model calculation date is T, the latest model penetration result calculation date is T1, and the latest model penetration result is The current model calculates the daily difference T m =T-T1, take the industry weight report date disclosed in the latest quarterly report as T2, and the latest quarterly report disclosure as m i,j,s＝T2 , the current model calculation date is the difference between the quarterly report release date and T p =T-T2;

[0229] ii. If the model results Exists: If T p ≤T m , T p ≠0, otherwise

[0230] iii. If the model results Does not exist:

[0231] g. Calculate the proportion of individual stock and industry holdings to the net value of equity assets (note that this is not the proportion of total net assets), and jointly calculate the residual of the return on equity assets in Scheme 1 and Scheme 2, and use them with Scheme 1 and Scheme 2 here:

[0232] Option 1: Daily return of individual stocks minus market, Hong Kong stocks

[0233]

[0234] Option 2: The daily return of individual stocks does not reduce the market

[0235]

[0236] Optimize the following problems and record the optimal solution

[0237] The optimization function is as follows:

[0238]

[0239] st

[0240]

[0241] W i,k,t *(W i,k,t –0.002)≥0

[0242]

[0243] where w i,k,t is the penetration weight result of individual stock j of fund i on day t, and the sum of individual stock weight vectors can be obtained w i,j,t is the penetration weight result of individual stock j of fund i on day t, and the sum of individual stock weight vectors can be obtained

[0244] The last step is to perform the third layer modeling as shown in S55. The specific steps are:

[0245] a. Given a fund i in a certain industry k, if w in the above results i,k,t＝T >0, the industry needs to be further mapped back to stock holdings;

[0246] b. Obtain historical holdings, obtain the holdings data of fund i in the above candidate stock pool, and sort out the list of individual stocks belonging to Shenwan Industry k;

[0247] c. Obtain research shares:

[0248] i. Get the list of fund managers of fund i;

[0249] ii. For each fund manager at time t = T, use the Juyuan database, investor relations activity research details LC_InvestorDetail and LC_InvestorRa tables to sort out all the stocks that he has researched in the past year, and use fund companies to distinguish duplicate names;

[0250] iii. Summarize all the researched stocks and compile a list of stocks belonging to Shenwan Industry K;

[0251] d. Combine the historical holdings and research stocks, remove duplicates, and then remove the candidate target pool stocks. The cleaned stock list is recorded as KH and KM stocks respectively;

[0252] e. If KH+KM == 0:

[0253] i. If the new platform stocks meet condition 1 (stocks of the same industry in the platform stocks are sorted from large to small by taking the log of market value, and weights are allocated according to market value ratio, and the largest weight is less than the smallest weight of the top 10 heavily-weighted stocks), then the estimated weight w of fund i in industry k is i,k,t＝T Delete and estimate the weight w of fund i in industry k i,k,t＝T Allocate to stocks in the same industry k as the platform stock;

[0254] ii. If the same industry in the platform stocks cannot meet the above condition 1, then obtain the stocks in the corresponding industry as a supplement, so that the sum of the platform stocks and the corresponding industry stocks can meet the above condition 1, and then estimate the weight w of fund i in industry k i,k,t＝T The sum of the allocation to platform stocks and corresponding industry stocks;

[0255] f. If KH+KM==1, calculate the estimated weight w of industry k in fund i i,k,t＝T Delete and add the stock at the same time, and the estimated weight of the stock is w i,k,t＝T ;

[0256] g. If KH+KM>1, then

[0257] i. Perform constrained linear regression (M days) on the index return of industry k for the above KH+KM stocks. The maximum weight shall not be greater than the minimum weight of the top 10 holdings, that is,

[0258]

[0259] st

[0260] The intercept is 0

[0261]

[0262] ii. The final weight of each stock in KH+KM is w i,k,t＝T ·[w kh ,w kh ] T ;

[0263] After the model is calculated, the results need to be processed:

[0264] a. For historical data: for each fund and each stock, perform EWMA calculation with half-life HL = 5 days;

[0265] b. For daily new data: For the current fund i, the new "model result" on day T is After calculating the EWMA of fund i on day T-1, the model result stored in the database is

[0266]

[0267] The innovative points of the present invention are:

[0268] Solve the problem of delayed quarterly disclosure of public funds, unknown operations and opaque information during the disclosure interval, and solve the problem of public fund managers whitewashing quarterly holdings to hide their holdings. The higher the accuracy of the model in predicting fund holding changes, the more accurate the understanding of the daily dynamics of the fund can be, and the problem of lagging information on the style / industry drift of public funds can be solved, which can improve the risk control ability of product selection. Based on the real-time predicted fund holdings information, it can more quickly identify risks such as fund style / industry drift, excessive shareholding concentration, and excessive industry concentration, and issue early warnings to investment research personnel to help public investment advisory business teams improve their portfolio risk control capabilities. Based on real-time expected fund holdings information and performance attribution models, it can better monitor the changes in risk exposure of fund portfolio styles / industries, improve market research capabilities, and increase the confidence of investment decisions. Based on real-time predicted fund holdings data, it can derive statistics on the dynamic changes in the holding styles / industries / individual stocks of public funds in the entire market, monitor the investment behavior of public institutions, and assist investment research personnel in judging market trends.

[0269] Obviously, the embodiments described above are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work, as well as changes made in specific implementation based on the ideas of the present invention, should all fall within the scope of protection of the present invention.

Claims

1. A public fund holding penetration method based on artificial intelligence, characterized in that: include: The quarterly report is supplemented based on the semi-annual report of the target fund, and further, the daily frequency penetration of the target fund is carried out.

2. The method for penetrating public fund holdings based on artificial intelligence according to claim 1, characterized in that: The steps of completing the quarterly report include: S1, holding inheritance: the remaining holdings except the top ten holdings in the most recent semi-annual report are inherited to the reserve holdings except the top ten holdings in the most recent quarterly report; S2, adjust weights: recalculate the weights of each stock in the above-mentioned reserve holdings based on the rise and fall of each stock from the above-mentioned semi-annual report to the quarterly report; S3, supplement reserve holdings: determine whether there are new shares entering the reserve holdings based on the restricted shares holdings data disclosed in the above quarterly reports or the top ten circulating shareholders data published by the listed company, and determine their weights; S4, industry alignment: scale the weights of each stock in the above reserve holdings according to the industry distribution data in the above quarterly report; S5, splicing verification: splice the backup positions calculated above with the top ten major positions published in the quarterly report, and verify based on the data published in the above quarterly report.

3. The method for penetrating public fund holdings based on artificial intelligence according to claim 2, characterized in that: The daily frequency penetration is the daily frequency penetration of major asset classes, and the steps include: S1, Fund classification: Target funds are divided into passive funds and active funds; S2, estimating hyperparameters, wherein the hyperparameters include the rough fitting variance parameter of the net value, the variance parameter of the fund product holding change, and the time weight decay parameter of the yield fitting item; S3, find the optimal solution of the function.

4. The method for penetrating public fund holdings based on artificial intelligence according to claim 2, characterized in that: The daily frequency penetration is the industry daily frequency penetration, and its steps include: S1: Estimate hyperparameters, including the net value rough fitting variance parameter, the fund product holding change variance parameter and the yield fitting term time weight decay parameter; S2: Set the prior of position weight; S3: Find the optimal solution of the function; S4: Result processing after model calculation.

5. The method for penetrating public fund holdings based on artificial intelligence according to claim 2, characterized in that: The daily frequency penetration is the daily frequency penetration of individual stocks, and its steps include: S1, model data preprocessing; S2, establish a candidate stock pool: include stocks whose individual stock holding weight accounts for 1% higher than the equity holding weight into the candidate target pool; S3: Calculate the residual sequence of equity asset returns; S4: Estimate hyperparameters, including net value rough fitting variance parameters, individual stock position change variance parameters, individual stock hyperparameters, Shenwan industry hyperparameters and yield fitting item time weight decay parameters; S5: Set the prior of position weight; S6: Calculate the proportion of individual stock and industry holdings to the net value of equity assets; S7: Arrange the list of individual stocks in the Shenwan industry in the candidate target pool; S8: Obtain the stocks that each fund manager of the target fund has researched; S9: After taking the union of the stocks in S7 and S8, the duplicate stocks and the stocks in the candidate stock pool are removed.

6. According to the artificial intelligence-based cemetery fund holding penetration method according to claims 3 to 5, it is characterized in that: The calculation method of the net value rough fitting variance parameter is: S1, calculate the fitting residual: r it =X t B it +a i +e it ,t∈M S2, calculation and solution: where r it is the daily return of target fund i at time t, X t is the industry fitting weight, B it is the daily return rate of target fund i industry, α i is the intercept corresponding to fund i, ε it is the fitting residual of the target fund at time t, M is the time window, is the rough fitting variance of the net value of target fund i at time t; The calculation method of the position change variance parameter is: where η i，j，t＝T is the variance parameter of the target fund i’s holdings in industry j at time t, S i is the number of quarterly reports published by target fund i since its establishment, m i，j，s is the industry holding weight of target fund i in industry j in the Sth published quarterly report period. The calculation method of the time weight decay parameter of the yield fitting term is: Where λ is the EWMA parameter, HL is the half-life, and the coefficients from the 1st to the Mth item in the formula are taken out during calculation, that is, {λ, λ(1-2), λ(1-λ) 2 ,…,λ(1-λ) M-1 }, and normalize the above coefficients, and record the normalized coefficients as γ t＝T ={γ T-1 , γ T-2 …, γ T-M }.

7. The method for penetrating public fund holdings based on artificial intelligence according to claim 3, characterized in that: The expression of the function is: where w i，t＝T is the industry weight of the target fund at i at time T, w i，t＝T The transpose of index，t is the industry index return at time t.

8. The method for penetrating public fund holdings based on artificial intelligence according to claim 4, characterized in that: The prior method for setting the position weight is: S1: Assume that the current model calculation date is T, the latest model penetration result calculation date is T1, and the latest model penetration result is The current model calculates the daily difference T m =T-T1; S2: Take the latest quarterly industry weight announcement date as T2, and the latest quarterly announcement date as m i，j，s＝T2 The difference between the current model calculation date and the quarterly report release date is T p =T-T2; S3: If the model results exists, and T p ≤T m , then prior=m i，j，s＝T2 ,otherwise If the model results If it does not exist, then prior=m i，j，s＝T2 .

9. The method for penetrating public fund holdings based on artificial intelligence according to claim 4, characterized in that: The expression of the function is: st ∑ j w i，j，t＝T ＝1。 10. The method for penetrating public fund holdings based on artificial intelligence according to claim 5, characterized in that: There are two methods for calculating the yield residual sequence in S3: Method 1, without subtracting A shares and Hong Kong stocks: Method 2: Subtract A shares and Hong Kong stocks: where r t，mkt，A is the return rate of CSI 300 Index, r t，mkt，HK is the Hang Seng Index return, w i，t＝T，A-sahre is the A-share asset weight result of fund i on day t, w i，t＝T，HK-share is the asset weight result of Hong Kong stocks of fund i on day t.

Images