Model training method, financial product index value determination method and device, electronic equipment and storage medium

Abstract

The invention discloses a model training method, a financial product index value determination method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a plurality of pieces of sample data, and enabling the sample data to comprise a first type of target samples, a second type of target samples and a third type of target samples, the first type of target samples are feature data samples obtained based on monthly frequency, the second type of target samples are feature data samples obtained based on daily frequency, and the third type of target samples are feature data samples obtained based on rising or non-rising daily values; and training a preset model according to each piece of sample data to obtain a prediction model, wherein the prediction model is used for predicting a financial time sequence according to the monthly degree and predicting rising or non-rising of financial indexes according to the daily degree. According to the invention, the improved model training method is provided, and multi-angle and multi-dimensional prediction analysis is provided for the user.

Description

Technical Field

[0001] This application relates to the technical field of model training, and in particular, to a model training method, a method for determining index values of financial products, an apparatus, an electronic device, and a storage medium. Background Art

[0002] Model training is a process of optimizing model parameters by using a large amount of data so that it can learn the internal patterns and features of the data, and thus accurately predict or classify new data. Model training is widely applied to various artificial intelligence and machine learning tasks.

[0003] In the investment research products in the financial field, there is a lack of providing users with multi-angle and multi-dimensional predictive analysis. Summary of the Invention

[0004] Embodiments of this application provide a model training method, a method for determining index values of financial products, an apparatus, an electronic device, and a storage medium to provide a new model training method and use the new model obtained by model training for determining the index values of financial products.

[0005] Embodiments of this application adopt the following technical solutions:

[0006] In a first aspect, embodiments of this application provide a model training method, where the training method includes:

[0007] Obtain a plurality of sample data, where the sample data includes first-class target samples, second-class target samples, and third-class target samples. The first-class target samples are feature data samples based on monthly frequency, the second-class target samples are feature data samples based on daily frequency, and the third-class target samples are feature data samples based on daily values of rising or non-rising;

[0008] Train a preset model according to each sample data to obtain a prediction model, where the prediction model is used to predict financial time series monthly and predict the rising or non-rising of financial indicators daily.

[0009] In some embodiments, the obtaining a plurality of sample data, where the sample data includes first-class target samples, second-class target samples, and third-class target samples, includes:

[0010] Based on the VAR model and financial features, obtain monthly frequency target samples, where the monthly frequency includes at least one of the following: 2-year financial interest rate, 3-year financial interest rate, 10-year financial interest rate;

[0011] Based on the selection of monthly frequency target features and the search popularity and relevance of key information in the industry, a daily frequency target sample is obtained. The daily frequency includes the monthly averages of the three parameters β0, β1, and β2 in the NS model.

[0012] Taking the increase of more than 0 on the current day compared to the previous day as positive samples and less than or equal to 0 as negative samples, a daily value target sample is obtained. The daily value includes the 10-year financial interest rate.

[0013] In some embodiments, the method further includes: setting the following model objectives:

[0014] The first objective is to predict the numerical values of the next 2-year, 3-year, and 10-year financial interest rates on a monthly basis based on a regression task, and select the 2-year, 3-year, and 10-year financial interest rates as the analysis targets.

[0015] The second objective is to predict the morphological characteristics of the curve composed of different maturities and corresponding interest rates at the same time point of the financial interest rate curve on a monthly basis based on a regression task, and select the translation factor, slope factor, and curvature factor of the financial interest rate curve as the analysis targets.

[0016] The third objective is to predict the increase or non-increase of the 10-year financial interest rate on the next day on a daily basis based on a classification task, select the 10-year financial interest rate with a set daily frequency as the target, and the target output is whether it increases or not.

[0017] In some embodiments, the training of the preset model according to each sample data to obtain a prediction model includes:

[0018] Combining the NS model, VAR model, LSTM machine learning model, and Transformer large language model to predict the first objective, the second objective, and the third objective.

[0019] Among them,

[0020] For the first objective, the features selected by the VAR model are used as inputs to train the LSTM model, and the model structures of LSTM and Transformer are used for predictive analysis.

[0021] For the second objective, the NS model is used to construct the level, slope, and curvature factors, and the model structures of LSTM and Transformer are used for predictive analysis.

[0022] For the third objective, the model structures of LSTM and Transformer are used for predictive analysis.

[0023] In some embodiments, the LSTM and Transformer model structures include:

[0024] Two layers of LSTM, two layers of encoder, and two layers of decoder;

[0025] The bottom layer uses two layers of unidirectional LSTM models. The results of the LSTM models are input into the Transformer network, and MSE LOSS is added to the last layer, and the activation function used is LeakyReLU.

[0026] In some embodiments, the method further includes:

[0027] Perform translation and scaling optimization on the LSTM and Transformer models. By using different learning rates for different network layers of the LSTM model and the Transformer, and setting the bias and weight_decay during model training to 0 or a very small value.

[0028] In a second aspect, an embodiment of the present application further provides a method for determining an index value of a financial product. Wherein, the determination method includes:

[0029] Obtain the source data of the financial product and the user's asset allocation and planning parameters;

[0030] Input the source data of the financial product and the user's asset allocation and planning parameters into a prediction model to obtain index values of different frequencies corresponding to the financial product. The prediction model is trained according to the model training method described in the first aspect.

[0031] In a third aspect, an embodiment of the present application further provides a model training device. Wherein, the training device includes:

[0032] An acquisition module for acquiring a plurality of sample data. The sample data includes a first type of target sample, a second type of target sample, and a third type of target sample. The first type of target sample is a feature data sample obtained based on a monthly frequency, the second type of target sample is a feature data sample obtained based on a daily frequency, and the third type of target sample is a feature data sample obtained based on the daily values of rising or non-rising;

[0033] A training module for training a preset model according to each sample data to obtain a prediction model. The prediction model is used to predict the financial time series on a monthly basis and the rising or non-rising of financial indicators on a daily basis.

[0034] In a fourth aspect, an embodiment of the present application further provides an electronic device, including: a processor; and a memory arranged to store computer-executable instructions. The executable instructions, when executed, cause the processor to execute the above method.

[0035] In a fifth aspect, an embodiment of the present application further provides a computer-readable storage medium storing one or more programs, which, when executed by an electronic device including a plurality of application programs, cause the electronic device to execute the above method.

[0036] The above at least one technical solution adopted in the embodiments of the present application can achieve the following beneficial effects: Obtain a plurality of sample data, where the sample data includes first-class target samples, second-class target samples, and third-class target samples. The first-class target samples are feature data samples obtained based on monthly frequency, the second-class target samples are feature data samples obtained based on daily frequency, and the third-class target samples are feature data samples obtained based on daily values of increase or non-increase. For the monthly frequency prediction target, use the data of monthly frequency and daily frequency mean, and for the daily frequency prediction task, use daily frequency data. Train the model to learn multi-faceted and mixed-frequency data to better fit the real data distribution. Further, train a preset model according to each sample data to obtain a prediction model, which is used to predict the financial time series monthly and predict the increase or non-increase of financial indicators daily. The prediction model obtained through training can be used to predict the financial time series monthly and predict the increase or non-increase of financial indicators daily, so as to provide users with multi-angle and multi-dimensional prediction analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0038] Figure 1 It is a schematic flow chart of the model training method in the embodiments of the present application;

[0039] Figure 2 It is a schematic flow chart of the method for determining the index value of a financial product in the embodiments of the present application;

[0040] Figure 3 It is a schematic structural diagram of the model training device in the embodiments of the present application;

[0041] Figure 4 It is a schematic structural diagram of an electronic device in the embodiments of the present application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0043] The models involved in the embodiments of this application are as follows:

[0044] NS model, the financial interest rate curve is an upward parabola, with the horizontal axis being the maturity and the vertical axis being the interest rate value. As an investment institution, most of the asset maturity allocations are not exactly at the critical maturity point, so the accurate grasp of the changes in the shape of the interest rate curve is the key to guiding investment behavior. In order to effectively characterize the financial interest rate curve, the NS model (Nelson-Siegel model) is usually used to fit the observed data to construct the interest rate curve. The model gives the model itself a strong economic significance by constructing state factors such as level (β0), slope (β1) and curvature (β2). The NS model is as follows, and the values ​​of β0, β1 and β2 can be obtained by fitting the historical data of financial interest rates.

[0045] VAR model is used to predict financial interest rates at a monthly frequency. It has the characteristics of few variables and few data samples, so it is suitable to use the vector auto-regressive model (VAR model). VAR model is a non-structural equation system model proposed by Sims in 1980. The model adopts the form of multiple equations. In each equation of the model, the endogenous variables are regressed on the lagged terms of all the endogenous independent variables of the model, and then the dynamic relationship of all the endogenous variables is estimated. It is often used to predict interconnected time series systems and analyze the dynamic impact of random disturbances on variable systems.

[0046] The VAR model is a simultaneous form of the autoregressive model, so it is an autoregressive model. Assuming that there is a relationship between two variables, if two autoregressive models are established separately, the relationship between the two variables cannot be captured. If the simultaneous form is adopted, the relationship between the two variables can be established. The structure of the VAR model is related to two parameters, one is the number of combined variables N, and the other is the maximum lag order k.

[0047] The LSTM model, Long Short-Term Memory Network (LSTM), is a special RNN model that can learn long-term dependencies. Compared with RNN, it can effectively capture the semantic associations between long sequences and alleviate the gradient disappearance or explosion phenomenon.

[0048] The core structures of LSTM and RNN are similar, and mainly the following two improvements have been made. First, the mechanism of "gates" is added. Each iteration can control the "input gate", "forget gate" and "output gate" to determine which information should be retained and which information can be forgotten, effectively alleviating the problem of gradient disappearance or explosion that may occur in the long sequence problem. Second, the "cell" memory unit is added, which is used to record the information passing through the hidden layer and can "remember" key information for a long time to avoid the problem of long-term dependence. The "input gate" and "forget gate" adjust the weights through the sigmoid module, determining that a certain proportion of the information can be retained in the "cell state", and the remaining information is forgotten. By adjusting the LSTM model, it can be used to predict the above three targets.

[0049] Transformer model. In 2017, Google proposed the Transformer model in the paper "Attention is All you need". It uses the Self-Attention structure to replace the RNN network structure commonly used in NLP tasks. Compared with the RNN network structure, its greatest advantage is that it can perform parallel computing. The biggest difference between Transformer and LSTM is that LSTM is serial. One word needs to be learned before learning the next word, while the training of Transformer is parallel, that is, all words are trained simultaneously, greatly accelerating the computing efficiency. Transformer is essentially an Encoder-Decoder architecture, that is, an encoding component and a decoding component.

[0050] The Transformer model contains multiple Encoders and Decoders. Each layer consists of multiple attention mechanism modules and feed-forward neural network modules. The Encoder is used to encode the input sequence into a high-dimensional feature vector representation, and the Decoder is used to decode the vector representation into the target sequence. In the Transformer model, techniques such as residual connection and layer normalization are also used to accelerate the model convergence and improve the model performance.

[0051] The core of the Transformer model is the self-attention mechanism (Self-Attention Mechanism), whose role is to assign a weight to each position in each input sequence, and then use these weighted position vectors as the output. The calculation process of the self-attention mechanism includes three steps: (1) Calculate the attention weights: Calculate the attention weights between each position and other positions, that is, the importance of each position to other positions. (2) Calculate the weighted sum: Multiply each position vector by the attention weights and then add them together to obtain the weighted sum vector. (3) Linear transformation: Perform a linear transformation on the weighted sum vector to obtain the final output vector.

[0052] The Transformer model can be constructed by continuously stacking multiple self-attention layers and feed-forward neural network layers. For the training of the Transformer model, it is usually pre-trained in an unsupervised manner and then fine-tuned in a supervised manner. During the pre-training process, methods such as autoencoders or masked language models are usually used for training, and the goal is to learn the representation of the input sequence. During the fine-tuning process, it is usually trained in a supervised manner. For example, in the machine translation task, parallel corpora are used for training, and the goal is to learn the mapping relationship that maps the input sequence to the target sequence.

[0053] The Encoder makes predictions in an autoregressive manner. At each time step, the decoder inputs a word and then outputs a word, using the predicted word as the input for the next time step to predict the word until the prediction ends.

[0054] The specific steps are as follows: 1) Position embedding. 2) Self-attention self-attention mechanism. 3) Layer Normalization and residual connection. 4) Apply activation functions to the two-layer linear mapping of the Feed forward.

[0055] Advantages of the Transformer model: 1) Better parallel performance. The Transformer model can compute simultaneously at all positions, thus making full use of the advantages of GPU parallel computing and accelerating the model training and inference processes. 2) Ability to handle long sequences. Traditional recurrent neural network models are prone to problems such as vanishing gradients and exploding gradients when dealing with long sequences, while the Transformer model uses the self-attention mechanism and can consider the information at all positions simultaneously, thus better handling long sequences. 3) Better performance. The Transformer model has achieved many important research results in the field of natural language processing. For example, it has achieved good results in tasks such as machine translation, text generation, and language models.

[0056] Existing financial interest rate prediction technologies use traditional methods and expert rules. These methods are basically based on the assumption that the change law of financial interest rates is linear and do not consider non-linear relationships, etc. However, in fact, the change law of interest rates is intricate and includes both linear and non-linear relationships; traditional methods have poor memory effects for long-term sequences and it is difficult to accurately focus on segmental memory; deep learning methods can well learn non-linear laws and have good long-term and short-term memory capabilities.

[0057] Existing deep learning models for sequence prediction have problems such as bidirectional forgetting of long-term memory and difficulty in focusing on important memories, and are also not suitable for time series prediction scenarios. For example, although LSTM is an improved algorithm of RNN, gradient explosion may occur when the input sequence reaches a certain level, and it cannot focus on important memories and is difficult to capture sudden and abnormal situations; for example, Transformer has excellent learning effect on NLP semantics, and its Attention mechanism pays more attention to the permutation invariance of semantic granularity, but the position information of time series is very important, so its learning of the continuity of time series is poor and its ability to extract the initial representation of input features is weak.

[0058] Due to the relatively stable change trend of the financial interest rate curve and the MSE method selected for model optimization of LOSS, through industry research, it is found that directly using the interest rate value of the previous period to replace the current period's interest rate value can also achieve a very small MSE value, even smaller than the results of existing literature. However, this has a serious problem of trend translation illusion. Similarly, some models may have an illusion of approaching the mean, and the MSE will also be very small; the problem of moving translation frequently appears in financial interest rate curve prediction models, and the model does not really learn the laws and patterns of time series.

[0059] Most existing financial interest rate prediction methods use monthly frequency data and do not use daily frequency data; the data range used is also small and does not fully cover all aspects of economic characteristics. The feature range used in existing financial predictions is small and does not include all aspects of economic characteristics, resulting in intricate and mutually influential model feature relationships, and not selecting appropriate feature combinations, so the model does not learn the main change laws.

[0060] Existing research and investment products do not predict financial interest rates from multiple angles and dimensions, resulting in users not obtaining more effective and accurate investment advice. Due to the limitations of actual prediction effects, the industry has not constructed a daily frequency financial interest rate prediction model, and the training objectives of model design are relatively single.

[0061] The following will describe in detail the technical solutions provided by each embodiment of the present application with reference to the accompanying drawings.

[0062] An embodiment of the present application provides a model training method. As Figure 1 shown, a schematic flow diagram of the model training method in the embodiment of the present application is provided. The method at least includes the following steps S110 to step S120:

[0063] Step S110: Obtain multiple sample data, where the sample data includes first-class target samples, second-class target samples, and third-class target samples. The first-class target samples are feature data samples obtained based on monthly frequency, the second-class target samples are feature data samples obtained based on daily frequency, and the third-class target samples are feature data samples obtained based on the daily values of rising or non-rising.

[0064] The multiple sample data are specifically classified into first-class target samples, second-class target samples, and third-class target samples.

[0065] The first-class target samples are feature data samples obtained based on monthly frequency and are used as data to be input into the model for training. For the monthly frequency target, 38 important features are extracted according to the impulse response of financial macroeconomic features and the optimization process of the pre-VAR model, mainly including features such as fundamentals, policy aspects, sentiment aspects, and capital aspects. At the same time, the training set and test set are set according to the total number of different months.

[0066] The second-class target samples are feature data samples obtained based on daily frequency and are used as data to be input into the model for training. For the daily frequency target, on the basis of the feature selection of the monthly frequency target, in order to effectively expand the feature range, according to the search heat and relevance of key information in the industry, the features are increased to 221, including variable features such as sentiment factors. Some indicators show non-stationary phenomena. Based on the classical methods for time series problems, differential transformations are performed on features such as R007, M2, SR, CPI, PPI, and the trading volume of inter-bank pledged repurchase. At the same time, the training set and test set are set according to the total number of different dates.

[0067] The third-class target samples are feature data samples obtained based on the daily values of rising or non-rising. In the sample, when the daily increase of the 10-year financial interest rate compared with the previous day exceeds 0, it is regarded as a positive sample, and when it is less than or equal to 0, it is regarded as a negative sample. In this way, 1629 positive examples and 3363 negative examples are constructed, and the ratio of positive and negative samples is 1:2.08.

[0068] It can be understood that the specific indicators of the above first-class target samples, second-class target samples, and third-class target samples are only examples and are not used to limit the protection scope in the embodiments of the present application.

[0069] Step S120: Train a preset model according to each of the sample data to obtain a prediction model, where the prediction model is used to predict the financial time series monthly and predict the rising or non-rising of financial indicators daily.

[0070] One of the objectives of the prediction model includes regression tasks. For example, predicting the numerical values of the next 2-year, 3-year, and 10-year financial interest rates on a monthly basis, involving a total of 3 models. Another objective of the prediction model includes regression tasks. For example, predicting the morphological characteristics of the financial interest rate curve (a curve composed of different maturities and corresponding interest rates at the same point in time) on a monthly basis, involving a total of 3 models. The third objective of the prediction model includes classification tasks. For example, predicting whether the 10-year financial interest rate will rise or not on the next day on a daily basis, involving a total of 1 model.

[0071] When training the preset model based on each of the sample data, the preset models selected include, but are not limited to, NS (classical econometric model), VAR (classical econometric model), LSTM (machine learning model), Transformer (large language model), etc. And the above models are effectively combined to make full use of their respective advantages to improve the performance and efficiency of the model, and finally achieve the prediction of the target.

[0072] Through the above model training method, a prediction model is obtained by training the preset model according to each of the sample data, using a method that combines a traditional model (VAR + NS) and a deep learning model. At the same time, the prediction model adopts an improved Transformer algorithm to adapt to financial time series data and uses an LSTM + Transformer model structure.

[0073] Through the above model training method, a daily frequency prediction task is constructed, and multiple prediction objectives are designed to fully display the financial interest rate trend, providing users with multi-angle and multi-dimensional prediction analysis in investment research products and enhancing investment returns.

[0074] Through the above model training method, a model combination of "classical econometric model + deep learning model + large language model" is constructed, making full use of the characteristics and advantages of different models, and making full use of high-frequency and low-frequency data sets from different aspects to ensure that the model fully learns financial laws and patterns. It seems that the method of the present invention can more accurately learn the complex structure of financial interest rates and finally achieve good prediction results.

[0075] In one embodiment of the present application, the obtaining of multiple sample data, where the sample data includes first-class target samples, second-class target samples, and third-class target samples, includes: obtaining monthly-frequency target samples based on a VAR model and financial features, where the monthly frequency includes at least one of the following: 2-year financial interest rate, 3-year financial interest rate, 10-year financial interest rate; obtaining daily-frequency target samples based on the selection of monthly-frequency target features and the search popularity and relevance of the industry for key information, where the daily frequency includes the monthly average of the three parameters β0, β1, and β2 in the NS model; taking the samples with a daily increase exceeding 0 compared to the previous day as positive samples and those less than or equal to 0 as negative samples to obtain daily-value target samples, where the daily value includes the 10-year financial interest rate.

[0076] The first-class target samples are feature data samples obtained based on the monthly frequency and serve as the data to be input into the model for training. For the monthly-frequency target, 38 important features are extracted according to the impulse response of financial macroeconomic features and the optimization process of the pre-VAR model, mainly including features such as fundamentals, policies, sentiment, and liquidity. At the same time, the training set and test set are set according to the total number of different months.

[0077] The second-class target samples are feature data samples obtained based on the daily frequency and serve as the data to be input into the model for training. For the daily-frequency target, on the basis of the selection of monthly-frequency target features, in order to effectively expand the feature range, according to the search popularity and relevance of the industry for key information, the features are increased to 221, including variable features such as sentiment factors. Some indicators show non-stationary phenomena, and based on the classical methods for time series problems, differential transformations are performed on features such as R007, M2, SR, CPI, PPI, and the trading volume of inter-bank pledged repurchase. At the same time, the training set and test set are set according to the total number of different dates.

[0078] The third-class target samples are feature data samples obtained based on the daily values of increase or non-increase. In the samples, the samples with a daily increase in the 10-year financial interest rate exceeding 0 compared to the previous day are regarded as positive samples, and those less than or equal to 0 are regarded as negative samples. In this way, 1629 positive examples and 3363 negative examples are constructed, and the ratio of positive and negative samples is 1:2.08.

[0079] Through the above method, three prediction targets are designed, and based on high-frequency data, a daily-frequency financial interest rate prediction model is designed.

[0080] In an embodiment of the present application, the method further includes: setting the following model objectives: a first objective, based on a regression task, predicting the next 2-year, 3-year, and 10-year financial interest rate values monthly, and selecting the 2-year, 3-year, and 10-year financial interest rates as analysis targets; a second objective, based on a regression task, predicting the morphological characteristics of the curve composed of different maturities and corresponding interest rates on the financial interest rate curve at the same time point, and selecting the translation factor, slope factor, and curvature factor of the financial interest rate curve as analysis targets; a third objective, based on a classification task, predicting the increase or non-increase of the 10-year financial interest rate on the next day daily, selecting the 10-year financial interest rate to set a daily frequency target, and the target output is whether it increases or not.

[0081] Model objective design. According to the actual application needs of investment institutions, the research on financial interest rates in the industry usually divides into two directions. One is the research and analysis of key term (single point) interest rates, and the other is the analysis of financial interest rate curves.

[0082] First, select the 2-year, 3-year, and 10-year financial interest rates as analysis targets among many key terms.

[0083] Secondly, the financial interest rate curve reflects the financial interest rate levels of each maturity. Effectively judging the up / down translation, high / low slope, and large / small curvature of this curve can guide the active change of investment strategies, obtain excess returns by adjusting portfolio duration, convexity, leverage, and the choice of trading time points, and reduce net value drawdown. Therefore, select the translation factor, slope factor, and curvature factor of the financial interest rate curve as analysis targets.

[0084] Finally, since this analysis mainly uses macroeconomic factors and most macro information is disclosed at a monthly frequency, the above objectives are all set monthly, and the target output is specific values. In addition, in order to further explore high-frequency trading, capture trading signals, and reduce market risks, on the basis of the above objectives, a daily frequency target is also set for the 10-year financial interest rate, and the target output is whether it increases or not.

[0085] In summary, organize the objectives as follows:

[0086] Objective 1, regression task, predicting the next 2-year, 3-year, and 10-year financial interest rate values monthly. A total of 3 models are involved.

[0087] Objective 2, regression task, predicting the morphological characteristics of the financial interest rate curve (the curve composed of different maturities and corresponding interest rates at the same time point) monthly. A total of 3 models are involved.

[0088] Objective 3, classification task, predicting the increase or non-increase of the 10-year financial interest rate on the next day daily. A total of 1 model is involved.

[0089] The three targets involve a total of 7 models. Subsequently, 7 models will be launched, and the monthly financial interest rate forecasts for 2-year, 3-year, and 10-year periods will be displayed on the investment research analysis platform. The daily frequency financial interest rate forecast for the 10-year period will be displayed. The forecast information will be displayed by setting a reference suggestion module in the in-house business system to assist the business department in asset allocation and planning. Internally, the bank's investment research capabilities can be output through channels such as the internal platform to serve customers in multiple dimensions and enhance customer stickiness.

[0090] Through the above method, monthly and daily frequency average data is used for the monthly frequency prediction target, and daily frequency data is used for the daily frequency prediction task, allowing the model to fully learn multi-faceted and mixed-frequency data to better fit the real data distribution.

[0091] In an embodiment of the present application, training the preset model according to each sample data to obtain a prediction model includes: combining the NS model, VAR model, LSTM machine learning model, and Transformer large language model to predict the first target, the second target, and the third target. Among them, for the first target, the features selected by the VAR model are used as inputs to train the LSTM model, and the model structures of LSTM and Transformer are used for prediction analysis; for the second target, the NS model is used to construct the level, slope, and curvature factors, and the model structures of LSTM and Transformer are used for prediction analysis; for the third target, the model structures of LSTM and Transformer are used for prediction analysis.

[0092] The problem of financial interest rate prediction is essentially a time series prediction problem. In this modeling, the most suitable models for this modeling goal are first selected: NS (classical econometric model), VAR (classical econometric model), LSTM (machine learning model), Transformer (large language model), etc. These models are effectively combined to make full use of their respective advantages to improve the performance and efficiency of the model, and finally achieve the prediction of the above three targets.

[0093] For Target 1 and Target 2, 1) leverage the strong interpretability of the VAR model to achieve the division of feature importance; 2) use the important features selected by the VAR model as inputs to train the LSTM model; 3) customize a Transformer model architecture for the particularity of time series tasks, use the output of the LSTM model as the input to the Encoder end of the Transformer model, and through training, rely on the strong logical reasoning and long-term memory capabilities of the large model to form a more accurate prediction ability, and finally use it for the prediction of Target 1 and Target 2. At the same time, considering the particularity of Target 3 (classification task), the performance of classical econometric models is not satisfactory, so mainly LSTM + Transformer is used for prediction analysis.

[0094] Through the above method, aiming at the advantages and disadvantages of each model, construct a multi-model combination algorithm LSTM + Transformer. The advantages of the models can make up for the disadvantages of the models. Customize and improve the data input method of the Transformer to make it adapt to the scenario of time series data. The combined model achieves the purpose of model enhancement and improves the model effect.

[0095] Adopt a two-stage method. Use a traditional model (VAR) to extract important features, and then use the selected features to train a deep learning time series model to fully utilize the characteristics of deep learning to improve the model learning ability.

[0096] In an embodiment of the present application, the LSTM and Transformer model structures include: two layers of LSTM, two layers of Encoder, and two layers of Decoder; the bottom layer uses two layers of unidirectional LSTM models. The results of the LSTM models are input into the Transformer network. Add MSE LOSS in the last layer, and the activation function used is LeakyReLU.

[0097] The LSTM + Transformer model used by the machine learning model group. The overall model structure of LSTM + Transformer is: 2Layer-LSTM → 2Layer Encoder → 2Layer Decoder. The bottom layer is two layers of unidirectional LSTM models. The output of the LSTM is all H vectors. The results of the LSTM are input into the Transformer network. The activation function used is LeakyReLU. Add MSE LOSS in the last layer.

[0098] In one embodiment of the present application, the method further includes: performing translational scaling optimization on the LSTM and Transformer models, by using different learning rates for different network layers of the LSTM model and the Transformer respectively, and setting the bias and weight_decay during the model training process to 0 or an extremely small value.

[0099] Model structure optimization: The financial interest rate data is essentially time series data. Transformer is a model technology for the NLP field, and the design of the Decoder stage is for autoregressive learning of language models. How to make the Decoder adapt to the time series prediction scenario is a hot topic in current academic and industrial research. By referring to the top conference papers in recent years, a suitable Transformer structure modified for time series data input can be found. Through comparison, it is found that the effect is better than SOTA in the financial interest rate prediction scenario. After multiple rounds of parameter optimization, such as the model's lr, dropout, batch_size, weight_decay, etc., taking the task of predicting the monthly interest rate of the 10-year financial period as an example, the comparison of the final optimized effect is as follows:

[0100] The effect of using the LSTM model alone is as follows (MSE = 2.99e-07, Ratio = 0.6087);

[0101] The effect of using LSTM + Transformer in combination is as follows (MSE = 2.04e-07, Ratio = 0.6522).

[0102] It can be seen that the MSE and the prediction accuracy of the upward and downward trends of LSTM + Transformer are significantly better than those of using the LSTM model alone. Both models learn well in the initial linear decline trend. LSTM + Transformer learns more fully, and the curve is more fitting. In the middle section, both models have some translations. In the latter half of the curve, both models have some deviations, but the LSTM + Transformer model will slightly improve such problems. Overall, the Transformer structure in LSTM + Transformer has a positive impact, so the LSTM + Transformer structure is effective.

[0103] In the initial optimization process, it was found that the predicted results of the model would shift upward by a certain intercept compared to the true values as a whole, and it was also found that the predicted results of the model tended towards the mean of the true values. The above results all indicate that the model has not been fully learned, and the model has not learned the laws and relationships of interest rate data at all, which is equivalent to a rule. However, the MSE may reach a lower level than using other methods. Methods such as adjusting the learning rate and feature normalization were tried, but the effect did not improve. It was explored and found that the model defaulted to using bias and layer Norm, and during the process, it might normalize the data, resulting in gradient disappearance, etc. Referring to the source code implementation of MOSS at Fudan University, different learning rates were used for different network layers of the LSTM model and Transformer, and the bias and weight_decay during the model training process were set to 0 or a very small value. Obviously, after parameter optimization, the problems of translation and scaling were greatly alleviated.

[0104] Through the above methods, by modifying parameters such as the bias weight (bias) and weight_decay, such problems were alleviated to a certain extent, the effect of the model was improved, and the SOTA effect in this scenario was achieved.

[0105] Model verification stage:

[0106] Daily frequency model construction and effect. The third target is a classification model, that is, to predict the significant increase and non-significant increase of the 10-year financial interest rate on a daily basis. Through multiple rounds of tuning of the model parameters, the best model results (seq_len = 20, lr = 0.02, transf_lr = 0.000001 etc.) were obtained, and auc_test: 0.7349. For example, when prob > 0.562, the corresponding precision is 60.4%, the recall rate is 9.7%, and the probability of predicting a positive example among all prediction samples is 4.9%. For example, if there are 30 days of increase in two months, the model will predict that there are 5 days of increase, among which 3 days are predicted correctly, and the probability of predicting an increase is 60% when predicting an increase. The virtual portfolio is based on the CSI 10-year financial active bond index (931017.CSI), and active position and leverage adjustments are implemented according to the operation strategy. During the period from April 2021 to March 2023, the annualized return level of the simulated portfolio reached 4.50% (for reference only), achieving an excess return of 73BP relative to the benchmark. At the same time, on the eve of the bond market crash in November 2022, additional positions and leverage were avoided, and the net value drawdown was moderately slowed down.

[0107] Model effect evaluation:

[0108] Targets one and two are regression models, and the evaluation methods include MSE (mean squared error) and up-down accuracy. The formulas are as follows:

[0109]

[0110] The RATIO represents the uplink and downlink accuracy rates of the prediction results. Among them, n is the number of samples, is the model fitting value, and y is the original data. For the third objective, conventional accuracy rate, precision rate, and recall rate evaluation methods are adopted. By comparing the MSE index and the Ratio index, it is considered that the LSTM+Transformer model has achieved good prediction of the monthly objective.

[0111] Such as Figure 2 shown, in the embodiment of the present application, a method for determining the index value of a financial product is further provided, wherein the determination method includes:

[0112] Step S210, obtain the source data of the financial product and the user's asset allocation and planning parameters.

[0113] Step S220, input the source data of the financial product and the user's asset allocation and planning parameters into the prediction model to obtain the index values of different frequencies corresponding to the financial product. The prediction model is trained according to the model training method described above.

[0114] By the above method, the source data of the financial product obtained includes but is not limited to: the characteristic data samples obtained based on the monthly frequency, as the data to be input into the model for training, the monthly frequency target, and 38 important characteristics are extracted according to the impulse response of the financial macroeconomic characteristics and the optimization process of the pre-VAR model, mainly including characteristics such as fundamental, policy, sentiment, and capital aspects. At the same time, the training set and the test set are set according to the total number of different months. The characteristic data samples obtained based on the daily frequency, as the data to be input into the model for training, the daily frequency target, on the basis of the selection of the monthly frequency target characteristics, in order to effectively expand the characteristic range, according to the search popularity and correlation of key information in the industry, the characteristics are increased to 221, including variable characteristics such as sentiment factors. The characteristic data samples obtained based on the daily values of rising or non-rising. In the sample, if the increase rate of the 10-year financial interest rate on the current day compared with the previous day exceeds 0, it is regarded as a positive sample, and if it is less than or equal to 0, it is regarded as a negative sample. In this way, 1629 positive examples and 3363 negative examples are constructed, and the ratio of positive and negative samples is 1:2.08.

[0115] The user's asset allocation and planning parameters include but are not limited to: the selection of the user's asset allocation, and the funds are invested in various assets according to different proportions.

[0116] Input the source data of the financial product and the user's asset allocation and planning parameters into the prediction model to obtain the index values of different frequencies corresponding to the financial product. The user's asset planning parameters refer to a series of key parameters and indicators that need to be considered in the process of the user's asset planning, and these parameters help to determine and optimize the management and allocation of the user's assets.

[0117] The embodiment of the present application also provides a model training device 300, as Figure 3 shown, which provides a structural schematic diagram of the model training device in the embodiment of the present application. The model training device 300 at least includes: an acquisition module 310 and a training module 320, where:

[0118] In an embodiment of the present application, the acquisition module 310 is specifically configured to: acquire a plurality of sample data, where the sample data includes first-class target samples, second-class target samples, and third-class target samples. The first-class target samples are feature data samples obtained based on monthly frequency, the second-class target samples are feature data samples obtained based on daily frequency, and the third-class target samples are feature data samples obtained based on the daily values of rise or non-rise.

[0119] The plurality of sample data are specifically classified into first-class target samples, second-class target samples, and third-class target samples.

[0120] The first-class target samples are feature data samples obtained based on monthly frequency and used as data to be input into the model for training. For the monthly frequency target, 38 important features are extracted according to the impulse response of financial macroeconomic features and the optimization process of the pre-VAR model, mainly including features such as fundamental, policy, sentiment, and capital aspects. At the same time, the training set and the test set are set according to the total number of different months.

[0121] The second-class target samples are feature data samples obtained based on daily frequency and used as data to be input into the model for training. For the daily frequency target, on the basis of the feature selection of the monthly frequency target, in order to effectively expand the feature range, according to the search heat and relevance of key information in the industry, the features are increased to 221, including variable features such as sentiment factors. There are unstable phenomena in some indicators. Based on the classical methods for time series problems, differential transformation is performed on features such as R007, M2, SR, CPI, PPI, and the trading volume of inter-bank pledged repurchase. At the same time, the training set and the test set are set according to the total number of different dates.

[0122] The third-class target samples are feature data samples obtained based on the daily values of rise or non-rise. In the samples, when the increase rate of the 10-year financial interest rate on the current day compared with the previous day exceeds 0, it is regarded as a positive sample, and when it is less than or equal to 0, it is regarded as a negative sample. In this way, 1,629 positive examples and 3,363 negative examples are constructed, and the ratio of positive and negative samples is 1:2.08.

[0123] It can be understood that the specific indicators of the above first-class target samples, second-class target samples, and third-class target samples are only examples and are not used to limit the protection scope in the embodiment of the present application.

[0124] In one embodiment of the present application, the training module 320 is specifically configured to: train a preset model according to each sample data to obtain a prediction model, and the prediction model is used to predict the financial time series monthly and predict the increase or non-increase of financial indicators daily.

[0125] One of the objectives of the prediction model includes a regression task. For example, predicting the next 2-year, 3-year, and 10-year financial interest rate values monthly, involving a total of 3 models. Another objective of the prediction model includes a regression task. For example, predicting the morphological characteristics of the financial interest rate curve (the curve composed of different maturities and corresponding interest rates at the same time point) monthly, involving a total of 3 models. The third objective of the prediction model includes a classification task. For example, predicting the increase or non-increase of the 10-year financial interest rate on the next day daily, involving a total of 1 model.

[0126] The preset models selected when training the preset model according to each sample data include, but are not limited to, NS (classical econometric model), VAR (classical econometric model), LSTM (machine learning model), Transformer (large language model), etc., and the above models are effectively combined to make full use of their respective advantages to improve the performance and efficiency of the model, and finally achieve the prediction of the target.

[0127] In one embodiment of the present application, the obtaining module 310 is further configured to

[0128] Based on the VAR model and financial characteristics, obtain monthly frequency target samples, and the monthly frequency includes at least one of the following: 2-year financial interest rate, 3-year financial interest rate, 10-year financial interest rate;

[0129] Based on the selection of monthly frequency target features and the search popularity and relevance of the industry for key information, obtain daily frequency target samples, and the daily frequency includes the monthly average values of the three parameters β0, β1, and β2 in the NS model;

[0130] Taking the increase of the current day compared to the previous day exceeding 0 as a positive sample and less than or equal to 0 as a negative sample, obtain daily value target samples, and the daily value includes the 10-year financial interest rate.

[0131] In one embodiment of the present application, the obtaining module 310 is further configured to set the following model objectives:

[0132] The first objective is to predict the next 2-year, 3-year, and 10-year financial interest rate values monthly based on a regression task, and select the 2-year, 3-year, and 10-year financial interest rates as the analysis targets;

[0133] The second objective is to select the shift factor, slope factor, and curvature factor of the financial interest rate curve as the analysis targets based on a regression task and predict the morphological characteristics of the curve composed of different maturities and corresponding interest rates of the financial interest rate curve at the same time point on a monthly basis;

[0134] The third objective is to predict the increase or non-increase of the 10-year financial interest rate on the next day on a daily basis based on a classification task, select the 10-year financial interest rate as the target with a daily frequency, and the target output is whether it increases or not.

[0135] In an embodiment of the present application, the training module 320 is further configured to

[0136] Combine the NS model, VAR model, LSTM machine learning model, and Transformer large language model to predict the first objective, the second objective, and the third objective.

[0137] Among them,

[0138] For the first objective, the features selected by the VAR model are used as inputs to train the LSTM model, and the model structures of LSTM and Transformer are used for predictive analysis;

[0139] For the second objective, the NS model is used to construct the level, slope, and curvature factors, and the model structures of LSTM and Transformer are used for predictive analysis;

[0140] For the third objective, the model structures of LSTM and Transformer are used for predictive analysis.

[0141] In an embodiment of the present application, the LSTM and Transformer model structures include:

[0142] Two layers of LSTM, two layers of Encoder, and two layers of Decoder;

[0143] The bottom layer uses two layers of unidirectional LSTM models. The results of the LSTM models are input into the Transformer network, and MSE LOSS is added to the last layer, and the activation function used is LeakyReLU.

[0144] In an embodiment of the present application, the training module 320 is further configured to

[0145] Perform translation and scaling optimization on the LSTM and Transformer models. By using different learning rates for different network layers of the LSTM model and the Transformer respectively, and setting the bias and weight_decay during the model training process to 0 or an extremely small value.

[0146] It can be understood that the above model training device can implement each step of the model training method provided in the foregoing embodiments. The relevant explanations regarding the model training method are applicable to the model training device and will not be elaborated here.

[0147] Figure 4 It is a schematic structural diagram of an electronic device according to an embodiment of the present application. Please refer to Figure 4 , at the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and a memory. Among them, the memory may include a memory, such as a high-speed random access memory (Random-Access Memory, RAM), and may also include a non-volatile memory, such as at least one disk memory, etc. Of course, the electronic device may also include other hardware required for other services.

[0148] The processor, network interface, and memory can be interconnected through an internal bus. The internal bus can be an ISA (Industry Standard Architecture, industrial standard architecture) bus, a PCI (Peripheral Component Interconnect, peripheral component interconnect standard) bus, or an EISA (Extended Industry Standard Architecture, extended industrial standard architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, Figure 4 only a two-way arrow is used in

[0149] The memory is used to store programs. Specifically, the program may include program code, and the program code includes computer operation instructions. The memory may include a memory and a non-volatile memory, and provide instructions and data to the processor.

[0150] The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it, forming a model training device at the logical level. The processor executes the program stored in the memory and is specifically used to perform the following operations:

[0151] Obtain multiple sample data, where the sample data includes first-class target samples, second-class target samples, and third-class target samples. The first-class target samples are feature data samples obtained based on monthly frequency, the second-class target samples are feature data samples obtained based on daily frequency, and the third-class target samples are feature data samples obtained based on the daily values of rising or non-rising.

[0152] Train a preset model according to each of the sample data to obtain a prediction model, which is used to predict the financial time series monthly and predict the rising or non-rising of financial indicators daily.

[0153] The above as in this application Figure 1 The method executed by the model training device disclosed in the embodiments shown in this application can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the hardware of the processor or the instructions in the form of software. The above processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of this application can be directly embodied as being executed and completed by the hardware decoding processor, or executed and completed by the combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method.

[0154] The electronic device can also execute Figure 1 the method executed by the model training device in Figure 1 the embodiments shown and implement the functions of the model training device in

[0155] An embodiment of the present application also provides a computer-readable storage medium storing one or more programs, where the one or more programs include instructions that, when executed by an electronic device including a plurality of application programs, enable the electronic device to execute Figure 1 the method executed by the model training device in the illustrated embodiment, and specifically used to execute:

[0156] Obtain a plurality of sample data, where the sample data includes first-class target samples, second-class target samples, and third-class target samples. The first-class target samples are feature data samples obtained based on monthly frequencies, the second-class target samples are feature data samples obtained based on daily frequencies, and the third-class target samples are feature data samples obtained based on daily values of increase or non-increase;

[0157] Train a preset model according to each of the sample data to obtain a prediction model, where the prediction model is used to predict financial time series on a monthly basis and the increase or non-increase of financial indicators on a daily basis.

[0158] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of an all-hardware embodiment, an all-software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0159] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0160] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1The functions specified in one or more boxes.

[0161] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide for implementing the steps of the functions specified in one Figure 1 one process or more processes and / or boxes Figure 1 the steps of the functions specified in one box or more boxes.

[0162] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0163] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). Memory is an example of computer-readable media.

[0164] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0165] It should also be noted that the term "comprises", "comprising", or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity, or device that includes a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, commodity, or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the process, method, commodity, or device that includes the element.

[0166] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0167] The above are only the embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A model training method, wherein: The training method comprises: Acquire multiple sample data, the sample data including a first type of target sample, a second type of target sample and a third type of target sample, the first type of target sample is a characteristic data sample obtained based on a monthly frequency, the second type of target sample is a characteristic data sample obtained based on a daily frequency, and the third type of target sample is a characteristic data sample obtained based on an increasing or non-increasing daily value; The preset model is trained according to each of the sample data to obtain a prediction model, and the prediction model is used to predict the financial time series on a monthly basis and to predict the increase or non-increase of financial indicators on a daily basis.

2. The method of claim 1, wherein: The acquiring of a plurality of sample data, wherein the sample data includes a first type of target sample, a second type of target sample, and a third type of target sample, comprises: Based on the VAR model and financial characteristics, a monthly frequency target sample is obtained, wherein the monthly frequency includes at least one of the following: a 2-year financial interest rate, a 3-year financial interest rate, and a 10-year financial interest rate; Based on the monthly frequency target feature selection and the industry's search popularity and relevance for key information, a daily frequency target sample is obtained. The daily frequency includes the monthly average values ​​of the three parameters β0, β1, and β2 in the NS model. The daily value target sample is obtained by taking the increase of more than 0 compared with the previous day as the positive sample and the increase of less than or equal to 0 as the negative sample, and the daily value includes the 10-year financial interest rate.

3. The method according to claim 2, further comprising: Set up the following model target: The first goal is to predict the next 2-year, 3-year and 10-year financial interest rates on a monthly basis based on the regression task, and select 2-year, 3-year and 10-year financial interest rates as the analysis targets; The second goal is to predict the morphological characteristics of the financial interest rate curve composed of different maturities and corresponding interest rates at the same time point on a monthly basis based on the regression task, and select the translation factor, slope factor and curvature factor of the financial interest rate curve as the analysis target; The third goal is to predict on a daily basis whether the 10-year financial interest rate will rise or not on the next day based on the classification task. The 10-year financial interest rate is selected to set a daily frequency target, and the target output is whether it will rise or not.

4. The method of claim 3, wherein: The step of training a preset model according to each sample data to obtain a prediction model includes: The NS model, the VAR model, the LSTM machine learning model and the Transformer large language model are combined to predict the first target, the second target and the third target. in, For the first goal, the features selected by the VAR model are used as input to the training of the LSTM model, and the model structure of LSTM and Transformer is used for prediction analysis; For the second objective, the NS model is used to construct the level, slope and curvature factors, and the LSTM and Transformer model structures are used for predictive analysis; For the third goal, LSTM and Transformer model structures are used for predictive analysis.

5. The method of claim 4, wherein: The LSTM and Transformer model structures include: Two layers of LSTM, two layers of encoders and two layers of decoders; The bottom layer uses a two-layer unidirectional LSTM model. The results of the LSTM model are input into the Transformer network. MSE LOSS is added to the last layer, and the activation function uses LeakyReLU.

6. The method according to claim 5, further comprising: The LSTM and Transformer models are optimized for translation and scaling by using different learning rates for different network layers of the LSTM model and Transformer, and the bias and weight_decay during model training are set to 0 or a minimum value.

7. A method for determining an index value of a financial product, wherein: The determination method comprises: Obtain source data of financial products and user asset allocation and planning parameters; The source data of the financial product and the user asset configuration and planning parameters are input into a prediction model to obtain indicator values ​​of different frequencies corresponding to the financial product, wherein the prediction model is trained according to the model training method described in any one of claims 1-6.

8. A model training device, wherein: The training device comprises: An acquisition module, used to acquire multiple sample data, the sample data including a first type of target sample, a second type of target sample and a third type of target sample, the first type of target sample is a characteristic data sample obtained based on a monthly frequency, the second type of target sample is a characteristic data sample obtained based on a daily frequency, and the third type of target sample is a characteristic data sample obtained based on an increasing or non-increasing daily value; The training module is used to train the preset model according to each of the sample data to obtain a prediction model, and the prediction model is used to predict the financial time series on a monthly basis and predict the increase or non-increase of financial indicators on a daily basis.

9. An electronic device, comprising: processor; as well as A memory arranged to store computer executable instructions, which when executed cause the processor to perform the method of any one of claims 1 to 6.

10. A computer-readable storage medium storing one or more programs, which, when executed by an electronic device including a plurality of application programs, causes the electronic device to execute any one of the methods of claims 1 to 6.

Images