Energy future price prediction method based on quantile feature reduction technology

Abstract

The invention discloses an energy future price prediction method based on a quantile feature reduction technology, and the method comprises the steps: capturing the heterogeneity influence of a high-dimensional tail risk factor on the condition distribution of the energy future price through a quantile regression (LQAR) model based on an LASSO algorithm, and achieving the prediction of the uncertainty of the energy future price. The limitation that the traditional OLS only predicts the mean effect is broken through; meanwhile, an LASSO algorithm is introduced to dynamically screen high-dimensional tail risk factors, the problems of noise interference and overfitting are solved, and the robustness of a prediction model is improved; in addition, by constructing a high-frequency high-dimensional tail risk factor set, the defect that a traditional method depends on low-frequency fundamental plane data is overcome, and therefore the accuracy of energy future price prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of forecasting methods, and more specifically, to an energy futures price forecasting method based on quantile feature reduction technology. Background Art

[0002] Sharp fluctuations in energy prices can pose a systemic threat to the normal functioning of the global macroeconomy and the stability of financial markets. Therefore, accurately forecasting energy futures prices is crucial for market participants in formulating risk management plans, investment decisions, and long-term strategic planning. However, the internal and external environments of energy markets are constantly evolving, and the factors influencing them are becoming increasingly diverse and complex. Notably, black swan events such as extreme climate events, pandemics, geopolitical conflicts, and cyclical economic downturns have had a significant impact on energy markets. Against this backdrop, effectively identifying the drivers of energy futures prices and accurately forecasting their uncertainty have become key technical challenges that urgently need to be addressed.

[0003] In the field of energy futures price forecasting, existing techniques generally rely on the ordinary least squares (OLS) forecasting paradigm to estimate the mean effect of risk factors on energy futures prices. This approach fails to capture the tail risks that induce extreme volatility and also struggles to identify heterogeneous influences dependent on market conditions. Empirical evidence shows that energy futures prices generally exhibit heavy-tailed distributions (e.g., oil and natural gas prices are subject to geopolitical and natural disaster shocks), resulting in significantly higher probability of extreme prices than expected by a normal distribution. Furthermore, a brief energy price spike can trigger systemic risks such as cash flow disruptions and supply chain disruptions, resulting in macroeconomic shocks far exceeding the mean. Furthermore, existing techniques often forecast energy futures prices based on low-frequency macroeconomic and energy market fundamentals. They lack the ability to effectively integrate high-dimensional tail risk information from multiple markets, such as commodities, foreign exchange, stocks, and bonds, at high frequencies (e.g., daily), making it difficult to capture the dynamic correlations between cross-market risks. Traditional methods, lacking efficient feature reduction techniques for multi-market high-dimensional tail risk data, are susceptible to noise, leading to model overfitting and information redundancy, further weakening forecast robustness. Summary of the Invention

[0004] In order to solve at least one of the above technical problems, the present invention provides an energy futures price forecasting method based on quantile feature reduction technology, which can effectively capture the heterogeneous impact of tail risk factors on the conditional distribution of energy futures prices, thereby realizing nonlinear modeling and conditional distribution forecasting of energy futures prices; by introducing feature reduction technology into the quantile regression model, it can efficiently integrate high-frequency and high-dimensional tail risk information, and improve the forecast accuracy and robustness of energy futures price uncertainty.

[0005] The present invention provides an energy futures price forecasting method based on quantile feature reduction technology, comprising the following steps:

[0006] Get raw asset price data information;

[0007] Preprocess the original asset price data information to obtain the logarithmic return series;

[0008] Send the logarithmic return series to the preset conditional autoregressive value-at-risk model to obtain a high-dimensional tail risk factor set;

[0009] Divide the high-dimensional tail risk factor set into a high-dimensional tail risk factor training set and a high-dimensional tail risk factor test set;

[0010] The high-dimensional tail risk factor training set is input into the quantile autoregressive model based on the LASSO algorithm to obtain the trained LQAR prediction model;

[0011] The high-dimensional tail risk factor test set is input into the trained LQAR forecasting model to obtain the energy futures price forecast under different conditional probability distributions for the next trading day.

[0012] The original asset price data information includes at least energy futures prices.

[0013] In this solution, the step of sending the logarithmic return series to a preset conditional autoregressive value-at-risk model to obtain a high-dimensional tail risk factor set specifically includes:

[0014] Number the logarithmic yield series in chronological order to obtain the variable vectors at different time nodes;

[0015] The variable vectors at different time nodes are sequentially input into the four-alternative conditional autoregressive risk value model at a given confidence level to obtain the tail risk estimates based on the four-alternative models at different time nodes.

[0016] According to the tail risk estimates based on the four replacement models at different time points, the DQ statistic, %Hits value and RQ loss value are obtained;

[0017] The best tail risk fitting model is selected based on three statistics: DQ statistic, %Hits value and RQ loss value;

[0018] Based on the optimal tail risk fitting model, the value at risk of each asset at a given confidence level is estimated to obtain a high-dimensional tail risk factor set.

[0019] In this solution, the basic formula of the conditional autoregressive value-at-risk model is specifically: where f t(β) represents the asset return rate of the τth (0<τ<1) quantile at time node t, β is the p-dimensional parameter to be estimated, β0 is a constant term, β i and β j They represent the estimated parameters of lag i and lag j respectively, q and r represent the maximum lag order respectively, and satisfy p = q + r + 1, x t Represents the variable vector at time node t, and L is a finite lag function.

[0020] In this solution, the four alternative conditional autoregressive value-at-risk models include an adaptive model, a symmetric absolute value model, an asymmetric model, and an indirect GARCH model; the formula in the adaptive model is specifically:

[0021] f t (β1)=f t―1 (β1)+β1{[1+exp(G[y t―1 ―f t―1 (β1)])] ―1 ―τ};

[0022] The formula in the symmetric absolute value model is:

[0023] f t (β)=β1+β2f t―1 (β)+β3|y t―1 |;

[0024] The formula in the asymmetric model is:

[0025] f t (β)=β1+β2f t―1 (β)+β3(y t―1 ) + +β4(y t―1 ) ― ;

[0026] The formula in the indirect GARCH model is:

[0027] f t (β)=(β1+β2f 2 t―1 (β)+β3y 2 t―1 ) 1 / 2 ; Among them, β1, β2, β3, β4 are the coefficients to be estimated, G represents a finite positive number, y t―1 represents the lagged first-order rate of return, (y t―1 ) + and (y t―1 ) ― denote positive and negative lagged returns, respectively.

[0028] In this solution, the step of selecting the best tail risk fitting model based on the three statistics of DQ statistic, %Hits value and RQ loss value is specifically as follows: determining the P value based on the DQ statistic; if the P value is greater than the preset significance level, the model corresponding to the DQ statistic is retained; if the P value is less than or equal to the preset significance level, the model corresponding to the DQ statistic is deleted; based on the %Hits value, calculating the absolute deviation between the %Hits and the theoretical exception ratio corresponding to the tail risk confidence level, and setting the model with the smallest absolute deviation as the best tail risk fitting model; if there are multiple models with the smallest absolute deviation, selecting the model with the smallest RQ loss value among the multiple models with the smallest absolute deviation as the best tail risk fitting model.

[0029] In this solution, the quantile autoregressive model based on the LASSO algorithm is specifically:

[0030] where Q τ (r t+1 │I t ) represents the information set I at a given time node t t The energy futures price return rate r at the τth quantile at the next t+1 time node t+1 , TR i,t represents the i-th tail risk factor at time node t, β0(τ) and β1(τ) are the estimated coefficients of the constant term and autoregressive term of the τ-th quantile, respectively, and V i (τ) is the estimated coefficient of the i-th tail risk factor under the τ-th quantile, k is the dimension of the tail risk factor set, ε t+1 (τ) is the error term, and the parameter estimation is performed by minimizing the objective function with LASSO penalty. The parameter estimation formula is:

[0031] in Represents the unknown coefficient estimate of the tail risk factor, T represents the length of the input data set interval, represents the LASSO penalty term, λ is the regularization parameter, and the optimal λ is selected through cross-validation.

[0032] In this solution, the formula for predicting energy futures prices under different conditional probability distributions for the next trading day is specifically:

[0033] in is the energy futures price return forecast value of the τth quantile at time node t+1, and (τ) are the unknown coefficient estimates of the τth quantile constant and the autoregressive term, The τth quantile i-th tail risk factor TRi,t The estimated values ​​of the unknown coefficients.

[0034] This plan also includes:

[0035] Based on the benchmark model and the LQAR model, the τth quantile energy futures price return forecast values ​​are obtained, which are set as and

[0036] Will and Perform comparative analysis to obtain the prediction result evaluation index quantile out-of-sample R 2 , the formula is where ρ τ (u)=u(τ―1), and u<0, u is the corresponding weight coefficient, S represents the Sth observation of the out-of-sample test set; s′ represents the length of the out-of-sample test set, f S represents the Sth actual value of the out-of-sample test set;

[0037] When the prediction result evaluation index quantile out-of-sample R 2 When it is less than or equal to zero, the LQAR model does not have better out-of-sample forecast performance than the benchmark model;

[0038] When the prediction result evaluation index quantile out-of-sample R 2 When is greater than zero, the LQAR model has better out-of-sample forecasting performance than the benchmark model.

[0039] The present invention discloses an energy futures price forecasting method based on quantile feature reduction technology. The method captures the heterogeneous impact of high-dimensional tail risk factors on the conditional distribution of energy futures prices through a quantile regression (LQAR) model based on the LASSO algorithm, thereby realizing the prediction of energy futures price uncertainty and breaking through the limitation of traditional OLS that only predicts the mean effect. At the same time, the LASSO algorithm is introduced to dynamically screen high-dimensional tail risk factors to solve the problems of noise interference and overfitting, thereby improving the robustness of the forecasting model. In addition, by constructing a high-frequency and high-dimensional tail risk factor set, the deficiency of traditional methods that rely on low-frequency fundamental data is compensated, thereby improving the accuracy of energy futures price forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flow chart of an energy futures price forecasting method based on quantile feature reduction technology according to the present invention is shown;

[0041] Figure 2 An example table showing the out-of-sample forecast performance of the LQAR model of the present invention and common competing models, the PCQAR model and the PQAR model;

[0042] Figure 3An example table showing out-of-sample forecast performance under the combined forecasting method of the present invention;

[0043] Figure 4 The heat map of the risk factor coefficients of the LQAR model is shown. DETAILED DESCRIPTION

[0044] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0046] Figure 1 A flow chart of an energy futures price forecasting method based on quantile feature reduction technology according to the present invention is shown.

[0047] like Figure 1 As shown, the present invention discloses a flow chart of an energy futures price forecasting method based on quantile feature reduction technology, including:

[0048] S101, obtaining original asset price data information;

[0049] S102, preprocessing the original asset price data to obtain a logarithmic yield sequence;

[0050] S103, sending the logarithmic return series to a preset conditional autoregressive value-at-risk model to obtain a high-dimensional tail risk factor set;

[0051] S104, dividing the high-dimensional tail risk factor set into a high-dimensional tail risk factor training set and a high-dimensional tail risk factor test set;

[0052] S105 inputs the high-dimensional tail risk factor training set into the quantile autoregressive model based on the LASSO algorithm to obtain the trained LQAR prediction model;

[0053] S106, input the high-dimensional tail risk factor test set into the trained LQAR prediction model to obtain the energy futures price forecast under different conditional probability distributions for the next trading day.

[0054] According to an embodiment of the present invention, the raw asset price data includes at least energy futures prices; the energy futures prices are selected from the daily closing prices of WTI crude oil futures, which are one of the world's important oil price benchmarks. Furthermore, considering the dual commodity and financial attributes of crude oil futures, the raw asset price data also includes at least price data corresponding to commodity futures, foreign exchange, stock market indices, and bond markets. Commodity futures include diesel, natural gas, gold, silver, etc.; foreign exchange includes Canadian dollar, British pound, Japanese yen, etc.; stock market indices include the S&P 500 Index and the Shanghai Composite Index; and bond markets include 10-year Treasury bond yields in the United States and the United Kingdom. The raw asset price data is standardized, logarithmically transformed, and differentially transformed to obtain a stable logarithmic yield series.

[0055] According to an embodiment of the present invention, the step of sending the logarithmic return sequence to a preset conditional autoregressive value-at-risk model to obtain a high-dimensional tail risk factor set specifically includes:

[0056] Number the logarithmic yield series in chronological order to obtain the variable vectors at different time nodes;

[0057] The variable vectors at different time nodes are sequentially input into the four-alternative conditional autoregressive risk value model at a given confidence level to obtain the tail risk estimates based on the four-alternative models at different time nodes.

[0058] According to the tail risk estimates based on the four replacement models at different time points, the DQ statistic, %Hits value and RQ loss value are obtained;

[0059] The best tail risk fitting model is selected based on three statistics: DQ statistic, %Hits value and RQ loss value;

[0060] Based on the optimal tail risk fitting model, the value at risk of each asset at a given confidence level is estimated to obtain a high-dimensional tail risk factor set.

[0061] It should be noted that, for example, if the tail risk confidence level is set to 99%, the variable vectors at different time nodes are sequentially input into the four-replacement conditional autoregressive risk value model at the 99% confidence level.

[0062] According to an embodiment of the present invention, the basic formula of the conditional autoregressive value-at-risk model is specifically: where f t (β) represents the asset return rate of the τth (0<τ<1) quantile at time node t, β is the p-dimensional parameter to be estimated, β0 is a constant term, β i and β jThey represent the estimated parameters of lag i and lag j respectively, q and r represent the maximum lag order respectively, and satisfy p = q + r + 1, x t Represents the variable vector at time node t, and L is a finite lag function.

[0063] According to an embodiment of the present invention, the four alternative conditional autoregressive value-at-risk models include an adaptive model, a symmetric absolute value model, an asymmetric model, and an indirect GARCH model; wherein the formula in the adaptive model is specifically:

[0064] f t (β1)=f t―1 (β1)+β1{[1+exp(G[y t―1 ―f t―1 (β1)])] ―1 ―τ};

[0065] The formula in the symmetric absolute value model is:

[0066] f t (β)=β1+β2f t―1 (β)+β3|y t―1 |;

[0067] The formula in the asymmetric model is:

[0068] f t (β)=β1+β2f t―1 (β)+β3(y t―1 ) + +β4(y t―1 ) ― ;

[0069] The formula in the indirect GARCH model is:

[0070] f t (β)=(β1+β2f 2 t―1 (β)+β3y 2 t―1 ) 1 / 2 ; Among them, β1, β2, β3, β4 are the coefficients to be estimated, G represents a finite positive number, y t―1 represents the lagged first-order rate of return, (y t―1 ) + and (y t―1 ) ― denote positive and negative lagged returns, respectively.

[0071] It should be noted that the conditional autoregressive value-at-risk (CAViaR) model is a semi-parametric model. Since it does not require residual distribution assumptions, it improves estimation efficiency and reduces model specification bias.

[0072] According to an embodiment of the present invention, it is characterized in that the step of selecting the best tail risk fitting model based on the three statistics of DQ statistic, %Hits value and RQ loss value is specifically as follows: determining the P value based on the DQ statistic; if the P value is greater than the preset significance level, the model corresponding to the DQ statistic is retained; if the P value is less than or equal to the preset significance level, the model corresponding to the DQ statistic is deleted; based on the %Hits value, calculating the absolute deviation between the %Hits and the theoretical exception ratio corresponding to the tail risk confidence level, and setting the model with the smallest absolute deviation as the best tail risk fitting model; if there are multiple models with the smallest absolute deviation, selecting the model with the smallest RQ loss value among the multiple models with the smallest absolute deviation as the best tail risk fitting model.

[0073] It should be noted that, for example, if the significance level is set to 10%, then when the P value is greater than 10%, the model corresponding to the DQ statistic is set to insignificant; if the P value is less than or equal to 10%, it means that the corresponding model has dynamic failure and the corresponding model will be deleted; the theoretical exception ratio = 1-confidence level. If the confidence level is set to 99%, the corresponding theoretical exception ratio is 1%; the absolute deviation degree = |actual %Hits value-theoretical exception ratio|.

[0074] According to an embodiment of the present invention, the quantile autoregressive model based on the LASSO algorithm is specifically:

[0075] where Q τ (r t+1 │I t ) represents the information set I at a given time node t t The energy futures price return rate r at the τth quantile at the next t+1 time node t+1 , TR i,t represents the i-th tail risk factor at time node t, β0(τ) and β1(τ) are the estimated coefficients of the constant term and autoregressive term of the τ-th quantile, respectively, and V i (τ) is the estimated coefficient of the i-th tail risk factor under the τ-th quantile, k is the dimension of the tail risk factor set, ε t+1 (τ) is the error term, and the parameter estimation is performed by minimizing the objective function with LASSO penalty. The parameter estimation formula is:

[0076] in Represents the unknown coefficient estimate of the tail risk factor, T represents the length of the input data set interval, represents the LASSO penalty term, λ is the regularization parameter, and the optimal λ is selected through cross-validation.

[0077] According to an embodiment of the present invention, a recursive window method is used to train and predict the model; specifically, the initial training set is set to 1000 daily observations, and in the initial training set {r1, r2, ..., r 1000}Train the model and predict The conditional distribution of the next true value r 1001 Incorporated into the training set, the expanded training set is {r1,r2,...,r 1001}, retrain the model and predict the next observation This is repeated recursively until all out-of-sample data points are predicted. Through this sample segmentation and prediction method, the predictive ability of the model can be evaluated in the time dimension, so as to better understand its performance in practical applications. The predicted variable is the crude oil futures price yield sequence, the input prediction factor is a high-dimensional tail risk factor set, and the training model is a quantile autoregression (LQAR) model based on the LASSO algorithm. The LASSO algorithm introduces L1 regularization, which compresses the prediction factor coefficients that are not important to the predicted variable to 0, thereby achieving automatic feature reduction, avoiding the overfitting problem of the model caused by the input of high-dimensional prediction factors, and improving the robustness of the prediction. The present invention introduces the LASSO algorithm into the quantile regression model, which can effectively identify the risk factor set under different conditional distributions of energy futures prices, and improve the accuracy and robustness of energy futures price uncertainty prediction.

[0078] It should be noted that the output of the training model in this embodiment is the conditional distribution of crude oil futures price returns, with quantiles selected as τ = {0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95}. By training the LQAR model at different quantiles, the optimal parameter estimates of the model are obtained.

[0079] According to an embodiment of the present invention, the formula for obtaining the energy futures price forecast under different conditional probability distributions for the next trading day is specifically:

[0080] in is the energy futures price return forecast value of the τth quantile at time node t+1, and (τ) are the unknown coefficient estimates of the τth quantile constant and the autoregressive term, is the τth quantile i-th tail risk factor TR i,t The estimated values ​​of the unknown coefficients.

[0081] According to an embodiment of the present invention, the further embodiment includes:

[0082] Based on the benchmark model and the LQAR model, the τth quantile energy futures price return forecast values ​​are obtained, which are set as and

[0083] Will and Perform comparative analysis to obtain the prediction result evaluation index quantile out-of-sample R 2 , the formula is quantile out-of-sample where ρ τ (u)=u(τ―1), and u<0, u is the corresponding weight coefficient, S represents the Sth observation of the out-of-sample test set; s′ represents the length of the out-of-sample test set, f S represents the Sth actual value of the out-of-sample test set;

[0084] When the prediction result evaluation index quantile out-of-sample R 2 When it is less than or equal to zero, the LQAR model does not have better out-of-sample forecast performance than the benchmark model;

[0085] When the prediction result evaluation index quantile out-of-sample R 2 When is greater than zero, the LQAR model has better out-of-sample forecasting performance than the benchmark model.

[0086] Figure 2 An example table showing the out-of-sample prediction performance of the LQAR model of the present invention and common competing models PCQAR model and PQAR model.

[0087] like Figure 2 As shown in Figure 2, the superior performance of the LQAR model selected by the present invention is shown compared with similar models based on other feature reduction techniques. The competing models are the PCQAR model and the PQAR model. Panels A, B, and C report the quantile out-of-sample R values ​​of the PCQAR model, the PQAR model, and the LQAR model of the present invention compared with the benchmark model, respectively. 2Results: Columns (1)-(7) represent the prediction performance of the conditional distribution at each quantile level. The bolded columns indicate that the CW statistic is significant at the 5% significance level, which means that the model has better prediction performance than the benchmark model at the 5% statistical significance level. Among them, the benchmark model selects the quantile autoregressive (QAR) model, that is, the quantile regression model that also includes autoregressive terms. The PCQAR model refers to the inclusion of the principal components extracted from the high-dimensional tail risk factor set by principal component analysis (PCA) on the basis of the QAR model, while the PQAR model refers to the inclusion of the high-dimensional tail risk factor index constructed by partial least squares (PLS) on the basis of the QAR model. Furthermore, according to the tail risk factor sets of the four asset classes of commodities, foreign exchange, stocks and bonds, the PCQAR_COM, PCQAR_FX, PCQAR_STK, PCQAR_BOND models and the corresponding PQAR_COM, PQAR_FX, PQAR_STK, PQAR_BOND models are constructed respectively. Due to the quantile out-of-sample R 2 The larger the value, the better the out-of-sample prediction performance compared to the benchmark model. Figure 2 The results show that the LQAR model quantile out-of-sample R 2 The prediction performance at all quantile levels except τ = 0.5 is better than that of other similar competing models.

[0088] Figure 3 An example table showing out-of-sample forecast performance under the combined forecasting method of the present invention is shown.

[0089] like Figure 2 As shown in the figure, the simple average combination has a relatively good out-of-sample forecast performance, but the LQAR model's out-of-sample forecast performance is still far better than all the combined forecast methods. This superiority is attributed to the LASSO algorithm's ability to dynamically filter out poorly performing predictors, resulting in a lower forecast error variance.

[0090] Figure 4 The heat map of the risk factor coefficients of the LQAR model is shown.

[0091] like Figure 4 As shown in the figure, based on the parameter estimation results of the LQAR model, four representative quantile levels, namely τ = {0.1, 0.25, 0.75, 0.9}, are selected to plot the dynamic selection heat maps of risk factors with coefficients not equal to 0 obtained by the LASSO algorithm under different conditional probability distributions. The gray shading of the bar area at each time point represents the relative contribution of the predictor variable. The darker the color, the higher the relative contribution. If the risk factor coefficient is equal to 0, the bar area is white, which means that the risk factor is eliminated.

[0092] The present invention discloses an energy futures price forecasting method based on quantile feature reduction technology. The method captures the heterogeneous impact of high-dimensional tail risk factors on the conditional distribution of energy futures prices through a quantile regression (LQAR) model based on the LASSO algorithm, thereby realizing the prediction of energy futures price uncertainty and breaking through the limitation of traditional OLS that only predicts the mean effect. At the same time, the LASSO algorithm is introduced to dynamically screen high-dimensional tail risk factors, solve the problems of noise interference and overfitting, and improve the robustness of the forecasting model. In addition, by constructing a high-dimensional risk factor set, the deficiency of traditional methods that rely on low-frequency fundamental data is compensated, thereby improving the accuracy of energy futures price forecasting.

[0093] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0094] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0095] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0096] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0097] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

Claims

1. A method for predicting energy futures prices based on quantile feature reduction technology, characterized in that: The steps include: Get raw asset price data information; Preprocess the original asset price data information to obtain the logarithmic return series; Send the logarithmic return series to the preset conditional autoregressive value-at-risk model to obtain a high-dimensional tail risk factor set; Divide the high-dimensional tail risk factor set into a high-dimensional tail risk factor training set and a high-dimensional tail risk factor test set; The high-dimensional tail risk factor training set is input into the quantile autoregressive model based on the LASSO algorithm to obtain the trained LQAR prediction model; The high-dimensional tail risk factor test set is input into the trained LQAR forecasting model to obtain the energy futures price forecast under different conditional probability distributions for the next trading day. The original asset price data information includes at least energy futures prices.

2. The energy futures price forecasting method based on quantile feature reduction technology according to claim 1 is characterized in that: The step of sending the logarithmic return sequence to a preset conditional autoregressive value-at-risk model to obtain a high-dimensional tail risk factor set specifically includes: Number the logarithmic yield series in chronological order to obtain the variable vectors at different time nodes; The variable vectors at different time nodes are sequentially input into the four-alternative conditional autoregressive risk value model at a given confidence level to obtain the tail risk estimates based on the four-alternative models at different time nodes. According to the tail risk estimates based on the four replacement models at different time points, the DQ statistic, %Hits value and RQ loss value are obtained; The best tail risk fitting model is selected based on three statistics: DQ statistic, %Hits value and RQ loss value; Based on the optimal tail risk fitting model, the value at risk of each asset at a given confidence level is estimated to obtain a high-dimensional tail risk factor set.

3. The energy futures price forecasting method based on quantile feature reduction technology according to claim 2 is characterized in that: The basic formula of the conditional autoregressive value-at-risk model is: where f t (β) represents the asset return rate at the τth (0<τ<1) quantile at time node t, β is the p-dimensional parameter to be estimated, β0 is a constant term, and β i and β j They represent the estimated parameters of lag i and lag j respectively, q and r represent the maximum lag order respectively, and satisfy p = q + r + 1, x t Represents the variable vector at time node t, and L is a finite lag function.

4. According to the energy futures price forecasting method based on quantile feature reduction technology of claim 3, the four alternative conditional autoregressive value-at-risk models include an adaptive model, a symmetric absolute value model, an asymmetric model, and an indirect GARCH model; wherein the formula in the adaptive model is specifically: f t (β1)=f t―1 (β1)+β1{[1+exp(G[y t―1 ―f t―1 (b1)])] ―1 ―τ}; The formula in the symmetric absolute value model is: f t (β)=β1+β2f t―1 (β)+β3|y t―1 |; The formula in the asymmetric model is: f t (β)=β1+β2f t―1 (β)+β3(y t―1 ) + +β4(y t―1 ) ― ; The formula in the indirect GARCH model is: f t (β)=(β1+β2f 2 t―1 (β)+β3y 2 t―1 ) 1 / 2 ; Among them, β1, β2, β3, β4 are the coefficients to be estimated, G represents a finite positive number, y t―1 represents the lagged first-order rate of return, (y t―1 ) + and (y t―1 ) ― denote positive and negative lagged returns, respectively.

5. The energy futures price forecasting method based on quantile feature reduction technology according to claim 2 is characterized in that: The step of selecting the best tail risk fitting model based on the three statistics of DQ statistic, %Hits value and RQ loss value is specifically as follows: determining the P value based on the DQ statistic; if the P value is greater than the preset significance level, retaining the model corresponding to the DQ statistic; if the P value is less than or equal to the preset significance level, deleting the model corresponding to the DQ statistic; calculating the absolute deviation between the %Hits and the theoretical exception ratio corresponding to the tail risk confidence level based on the %Hits value, and setting the model with the smallest absolute deviation as the best tail risk fitting model; if there are multiple models with the smallest absolute deviation, selecting the model with the smallest RQ loss value among the multiple models with the smallest absolute deviation as the best tail risk fitting model.

6. The energy futures price forecasting method based on quantile feature reduction technology according to claim 1 is characterized in that: The quantile autoregressive model based on the LASSO algorithm is specifically: where Q τ (r t+1 │I t ) represents the information set I at a given time node t t The energy futures price return rate r at the τth quantile at the next t+1 time node t+1 , TR i,t represents the i-th tail risk factor at time node t, β0(τ) and β1(τ) are the estimated coefficients of the constant term and autoregressive term of the τ-th quantile, respectively, and V i (τ) is the estimated coefficient of the i-th tail risk factor under the τ-th quantile, k is the dimension of the tail risk factor set, ε t+1 (τ) is the error term, and the parameter estimation is performed by minimizing the objective function with LASSO penalty. The parameter estimation formula is: Represents the unknown coefficient estimate of the tail risk factor, T represents the length of the input data set interval, represents the LASSO penalty term, λ is the regularization parameter, and the optimal λ is selected through cross-validation.

7. The energy futures price forecasting method based on quantile feature reduction technology according to claim 1 is characterized in that: The formula for predicting energy futures prices under different conditional probability distributions for the next trading day is specifically: in is the energy futures price return forecast value of the τth quantile at time node t+1, and (τ) are the unknown coefficient estimates of the τth quantile constant and the autoregressive term, The τth quantile i-th tail risk factor TR i,t The estimated values ​​of the unknown coefficients.

8. The energy futures price forecasting method based on quantile feature reduction technology according to claim 1 is characterized in that: Also includes: Based on the benchmark model and the LQAR model, the τth quantile energy futures price return forecast values ​​are obtained, which are set as and Will and Perform comparative analysis to obtain the prediction result evaluation index quantile out-of-sample R 2 , the formula is where ρ τ (u)=u(τ―1), and u<0, u is the corresponding weight coefficient, S represents the Sth observation of the out-of-sample test set; s ′ represents the length of the out-of-sample test set, f S represents the Sth actual value of the out-of-sample test set; When the prediction result evaluation index quantile out-of-sample R 2 When it is less than or equal to zero, the LQAR model does not have better out-of-sample forecast performance than the benchmark model; When the prediction result evaluation index quantile out-of-sample R 2 When is greater than zero, the LQAR model has better out-of-sample forecasting performance than the benchmark model.

Images