Method, device and equipment for bias compensation of time series prediction data, and medium

Abstract

The application relates to the technical field of data processing, and discloses a bias compensation method and device for time series prediction data, equipment and a medium. By constructing a three-layer architecture of basic prediction-residual compensation-Bayesian dynamic fusion, first, the initial prediction value is generated by using the differentiated benchmark and cross-period correction of working days and holidays, then the residual sequence is predicted from three dimensions of time series trend, hierarchical structure and category structure, and finally the optimal weight is adaptively determined by a Bayesian fusion algorithm to realize multi-dimensional bias compensation. It can be widely applied to time series prediction scenes in the fields of financial technology, medical health and the like. The problems of coarse granularity, lack of correction mechanism and inability to adaptively fuse of the traditional same period mean method are effectively solved. The whole-link optimization from bias detection, tracing to accurate compensation is realized, and the accuracy, stability and interpretability of the prediction result are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, device, and medium for offset compensation of time series prediction data. Background Technology

[0002] Currently, forecasting based on historical data often employs the average method, which directly takes the average of multiple indicators from the same historical period as the forecast result. This method is widely used in scenarios such as insurance, finance, and healthcare, but it generally suffers from coarse-grained forecasting. That is, it is difficult to distinguish the different impacts of different date types on time-series indicators, it cannot effectively trace and correct forecast biases, and it lacks an adaptive fusion mechanism for multi-source information, resulting in systematic biases in forecast results across different periods, making it difficult to meet the accurate forecasting needs of complex business scenarios.

[0003] Taking the fintech sector as an example, the traditional mean-of-contemporaneous method cannot distinguish the temporal impact of weekdays and holidays on business capacity, nor can it trace whether the deviation stems from changes in manpower, product structure adjustments, or seasonal fluctuations. This results in insufficient forecast accuracy and makes it difficult to support refined resource allocation and business decisions. In the healthcare sector, the traditional mean-of-contemporaneous method ignores the fluctuation differences in demand for medical services between holidays and weekdays, and lacks a mechanism for tracing and correcting deviations, affecting the accuracy of medical resource allocation, medical staff scheduling, and drug inventory management. Summary of the Invention

[0004] This invention provides a method, apparatus, computer equipment, and medium for bias compensation of time series forecast data, in order to solve the technical problem that the coarse granularity of the prediction by the contemporaneous mean method and the lack of bias correction for the prediction results lead to insufficient prediction accuracy.

[0005] Firstly, a method for compensating for biases in time-series forecast data is provided, including: Obtain the preset forecast period and the cumulative numerical indicators to be predicted; Acquire multi-source time series data from multiple historical periods that coincide with the preset prediction period. The multi-source time series data includes the actual values ​​of cumulative numerical indicators in each historical period, date attribute information for each historical period, and the contribution value of each detailed data item corresponding to the cumulative numerical indicators in each historical period. Based on multi-source time series data, basic forecast parameters are determined, including at least one of the following: the first historical baseline value corresponding to weekday periods, the second historical baseline value corresponding to holiday periods, and the cross-period effect correction parameter; Based on the basic prediction parameters and the date attribute information of each historical period, back-forward predictions are made for each historical period to generate back-forward prediction values ​​of cumulative numerical indicators for each historical period. A residual sequence is generated based on the difference between the actual value and the backtested predicted value for each historical period. Based on multi-source time series data and date attribute information of the preset prediction period, residual prediction is performed on the residual sequence according to multiple preset dimensions to obtain multiple residual prediction values. Among them, the multiple preset dimensions include time series dimension, hierarchical structure dimension and category structure dimension. Based on the basic forecast parameters and the date attribute information of the preset forecast period, the initial forecast value of the cumulative numerical index is generated for the preset forecast period. By using a Bayesian fusion algorithm, multiple residual prediction values ​​are used as deviation compensation amounts to compensate for the deviation of the initial prediction value, thereby generating the target prediction value of the cumulative numerical index within the preset prediction period.

[0006] Secondly, a deviation compensation device for time series prediction data is provided, comprising: The first acquisition module is used to acquire the preset forecast period and the cumulative numerical indicators to be predicted. The second acquisition module is used to acquire multi-source time series data of multiple historical periods that are concurrent with the preset prediction period. The multi-source time series data includes the actual value of the cumulative numerical indicator in each historical period, the date attribute information of each historical period, and the contribution value of each detailed data corresponding to the cumulative numerical indicator in each historical period. The determination module is used to determine the basic forecast parameters based on multi-source time series data. The basic forecast parameters include at least one of the following: the first historical benchmark value corresponding to the weekday period, the second historical benchmark value corresponding to the holiday period, and the cross-period effect correction parameter. The first generation module is used to perform back-forward prediction for each historical period based on the basic prediction parameters and the date attribute information of each historical period, and generate the back-forward prediction value of the cumulative numerical index for each historical period. The second generation module is used to generate a residual sequence based on the difference between the actual value and the backtracked predicted value for each historical period. The third generation module is used to perform residual prediction on the residual sequence based on multi-source time series data and date attribute information of the preset prediction period, according to multiple preset dimensions, to obtain multiple residual prediction values. Among them, the multiple preset dimensions include time series dimension, hierarchical structure dimension and category structure dimension. The fourth generation module is used to generate the initial forecast value of the cumulative numerical index for the preset forecast period based on the basic forecast parameters and the date attribute information of the preset forecast period. The fifth generation module is used to compensate for the deviation of the initial prediction value by using multiple residual prediction values ​​as deviation compensation amounts through a Bayesian fusion algorithm, and to generate the target prediction value of the cumulative numerical index for the preset prediction period.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for compensating for deviations in timing prediction data.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described method for compensating for deviations in timing prediction data.

[0009] The aforementioned scheme, encompassing the method, apparatus, computer equipment, and storage medium for bias compensation of time-series forecast data, constructs a three-layer architecture: basic forecasting, residual compensation, and Bayesian dynamic fusion. It first generates initial forecast values ​​using differentiated benchmarks for weekdays and holidays and inter-period corrections. Then, it predicts residual sequences from three dimensions: time-series trend, hierarchical structure, and category structure. Finally, a Bayesian fusion algorithm adaptively determines the optimal weights, achieving multi-dimensional bias compensation. This effectively solves the problems of coarse granularity, lack of correction mechanisms, and inability to adaptively fuse traditional contemporaneous mean methods. It achieves end-to-end optimization from bias detection and tracing to precise compensation, significantly improving the accuracy, stability, and interpretability of forecasts. In financial scenarios, it helps improve the accuracy of performance forecasting and capital allocation, supporting adjustments to investment strategies; in medical scenarios, it effectively improves the stability of forecasting medical demand. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an application environment for a deviation compensation method for time series prediction data according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for compensating for deviations in time-series prediction data according to an embodiment of the present invention; Figure 3 This is a schematic block diagram of a deviation compensation method for time series prediction data in one embodiment of the present invention; Figure 4 This is a schematic diagram of the monthly NBEV production capacity distribution at different time periods in one embodiment of the present invention; Figure 5 This is a schematic block diagram of residual prediction for the agent hierarchy dimension in one embodiment of the present invention; Figure 6This is a schematic block diagram of residual prediction for the product category structure dimension in one embodiment of the present invention; Figure 7 This is a schematic diagram of a deviation compensation device for time series prediction data in one embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 9 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] The deviation compensation method for time series prediction data provided in this embodiment of the invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can obtain a preset forecast period and the cumulative numerical indicators to be predicted; obtain multi-source time-series data from multiple historical periods contemporaneous with the preset forecast period, where the multi-source time-series data includes the actual values ​​of the cumulative numerical indicators in each historical period, the date attribute information of each historical period, and the contribution value of each detailed data point corresponding to the cumulative numerical indicators in each historical period; based on the multi-source time-series data, determine the basic forecast parameters, where the basic forecast parameters include at least one of the following: a first historical baseline value corresponding to weekday periods, a second historical baseline value corresponding to holiday periods, and a cross-period effect correction parameter; based on the basic forecast parameters and the date attribute information of each historical period, perform backtracking forecasts for each historical period respectively. The system generates retrospective predicted values ​​of cumulative numerical indicators for each historical period. Based on the difference between the actual values ​​and the retrospective predicted values ​​for each historical period, a residual sequence is generated. Based on multi-source time-series data and date attribute information of the preset prediction period, residual prediction is performed on the residual sequence according to multiple preset dimensions to obtain multiple residual predicted values. These preset dimensions include time-series dimension, hierarchical structure dimension, and category structure dimension. Based on basic prediction parameters and date attribute information of the preset prediction period, an initial predicted value of the cumulative numerical indicator for the preset prediction period is generated. Through a Bayesian fusion algorithm, the initial predicted value is compensated for the deviation using multiple residual predicted values ​​as deviation compensation amounts, generating the target predicted value of the cumulative numerical indicator for the preset prediction period. By constructing a three-layer architecture of basic prediction, residual compensation, and Bayesian dynamic fusion, the system solves the technical problems of coarse granularity, lack of deviation correction mechanism, and inability to adaptively fuse traditional contemporaneous mean method predictions. It achieves end-to-end optimization from deviation detection and tracing to precise compensation, significantly improving the accuracy and stability of prediction results and enhancing the interpretability of prediction deviations, providing reliable data support for resource allocation and business decision-making. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0014] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the deviation compensation method for time series prediction data provided in this embodiment of the invention includes the following steps: S10: Obtain the preset forecast period and the cumulative numerical indicators to be predicted.

[0015] In this step, the preset forecast period refers to the target time interval for numerical forecasting, such as a specific month in a particular year (denoted as month N). The selection of this period depends on the actual needs of business forecasting and is usually a statistical period that has not yet occurred or is currently underway but has not yet ended. The cumulative numerical indicator refers to the core business metric for which the summary results need to be estimated within the preset forecast period. The specific type of this indicator depends on the application scenario. By clearly defining the preset forecast period and the cumulative numerical indicator, basic input conditions are provided for the subsequent acquisition of historical data from the same period and the construction of the forecasting model.

[0016] For example, insurance companies need to forecast the monthly New Business Embedded Value (NBEV) to support performance target setting, capital allocation and business strategy adjustment. The cumulative value to be forecasted is the monthly NBEV, and the preset forecast period is the month to be forecasted.

[0017] S20: Obtain multi-source time series data from multiple historical time periods that coincide with the preset prediction time period.

[0018] The multi-source time-series data includes the actual values ​​of cumulative numerical indicators in each historical period, the date attribute information of each historical period, and the contribution value of each detailed data point corresponding to the cumulative numerical indicators in each historical period.

[0019] In this step, multiple historical time periods and the preset prediction time period belong to the same relative time position. For example, if the preset prediction time period is the month to be predicted, then the multiple historical time periods are the same month in each year in history that was the month to be predicted. Multi-source time series data is time series data that integrates various types of information. It includes the actual value of the cumulative numerical indicator corresponding to each historical time period, date attribute information, and the contribution value of each detailed data point. The actual value of the cumulative numerical indicator refers to the sum of the indicator values ​​that have actually occurred within each historical time period. The date attribute information is the date type information corresponding to each day within each historical time period, including weekdays and holidays. The contribution value of each detailed data point is the actual contribution of each detailed data point to the cumulative numerical indicator; the sum of the contribution values ​​of all detailed data points constitutes the actual value of that historical time period.

[0020] For example, an insurance company's cumulative numerical indicator is NBEV (New Business Expectations Exceeded). This indicator is a core value benchmark for measuring the insurance company's long-term profitability and business growth potential. The preset forecast period is M years and N months. Then, multiple historical periods are the N months of each year that belong to the same month as M years and N months in history. The number of these periods can be determined based on data availability and business needs, generally 3 to 5. The actual value is the total NBEV value actually calculated for each historical N months. The date attribute information is the workday or holiday marking for each day in each historical N months. The detailed data is for each policy, and the contribution value is the NBEV contribution value corresponding to each policy. In the healthcare field, if the cumulative numerical indicator is the monthly outpatient visits of a hospital, this indicator is a direct business representation reflecting the fluctuations in hospital service volume and patient demand. The preset forecast period is M years and N months. Then, multiple historical periods are M-1 years and N months, M-2 years and N months, and M-3 years and N months. The actual value is the total number of outpatient visits actually counted in each historical N months. The date attribute information is the weekday or holiday mark corresponding to each day in each historical N months. The detailed data is the visit record for each case. The contribution value is the contribution of each visit to the total number of visits (i.e., 1 visit).

[0021] S30: Determine the basic prediction parameters based on multi-source time series data.

[0022] The basic forecast parameters include at least one of the following: the first historical baseline value corresponding to the weekday period, the second historical baseline value corresponding to the holiday period, and the cross-period effect correction parameter.

[0023] In this step, based on multi-source time-series data, fundamental forecasting parameters are determined to generate initial and retrospective forecasts. Specifically, the first historical baseline value corresponding to weekday periods reflects the average contribution level of the weekday category in historical periods; the second historical baseline value corresponding to holiday periods reflects the average contribution level of the holiday category in historical periods. By separately determining the historical baseline values ​​for weekdays and holidays, the differentiated impact of different date types on indicator values ​​can be distinguished, improving forecast granularity and accuracy. Furthermore, the inter-period effect correction parameter reflects the inter-period effect generated by detailed data across different statistical periods, correcting for this effect in the forecast calculation, eliminating interference from inter-period effects on the forecast values, and making the forecast results more accurate.

[0024] In one embodiment of this application, a specific scheme for determining basic prediction parameters is provided. Step S30, namely determining the basic prediction parameters based on multi-source time series data, specifically includes the following steps S31-S38: S31: Based on the date attribute information of each historical period, the dates within each historical period are divided into weekday periods and holiday periods.

[0025] In this step, due to the significant differences in production capacity between weekdays and holidays, it is necessary to classify dates into two types for separate statistical analysis. Date attribute information includes the corresponding weekday or holiday markers for each day within each historical period. Using the date attribute information for each historical period, dates within that period are divided into weekday periods and holiday periods.

[0026] S32: Obtain the effective time information and recording time information of each detailed data in each historical time period.

[0027] S33: Determine whether the effective time information and recording time information of each detailed data in each historical period belong to the same statistical period.

[0028] For steps S32-S33, detailed data refers to each underlying data record that constitutes the cumulative numerical indicator. For example, in the insurance and finance sector, detailed data could be a policy underwriting record, with the recording time information being the signing date and the effective date being the policy effective date. In the healthcare sector, detailed data could be an outpatient registration record, with the recording time information being the registration date and the effective date being the actual date of visit.

[0029] Furthermore, the statistical period refers to the basic time unit used to summarize cumulative numerical indicators. For example, if the statistical period is one month, it is determined whether the effective date and record date of each detailed data point are within the same month. This is to determine whether there is a cross-period effect, and thus decide whether a cross-period effect correction value needs to be calculated when calculating the baseline value.

[0030] S34: If the effective time information and record time information of all detailed data belong to the same statistical period, the cross-period effect correction parameter is determined to be zero.

[0031] In this step, if the effective time information and the recording time information of all detailed data belong to the same statistical period, it means that the actual contribution time of all detailed data is consistent with its recording time, and there is no cross-period effect.

[0032] For example, in the insurance and finance sector, if the signing date and effective date of all policies are within the same month, it indicates that all policies signed in that month become effective in that month, and there are no cases of policies becoming effective across months. In this case, when calculating the benchmark value, there is no need to remove the contribution value of any detailed data, nor is there any need to set an additional cross-period correction amount. Therefore, the cross-period effect correction parameter is set to zero, and the subsequent initial forecast value is only composed of the sum of the products of the number of days in each period and the corresponding benchmark value.

[0033] S35: Obtain the sum of the first contribution values ​​of all detailed data belonging to weekday periods within multiple historical time periods, and the sum of the second contribution values ​​of all detailed data belonging to holiday periods.

[0034] In this step, assuming there is no cross-period effect, the contribution values ​​of detailed data corresponding to all working days in multiple historical periods are summed to obtain the sum of the first contribution values ​​of the detailed data belonging to the working day period. At the same time, the contribution values ​​of detailed data corresponding to all holidays in multiple historical periods are summed to obtain the sum of the second contribution values ​​of the detailed data belonging to the holiday period.

[0035] S36: Divide the sum of the first contribution values ​​by the total number of working days to obtain the first historical baseline value for the working day period.

[0036] In this step, the total number of working days across multiple historical periods is calculated. The sum of the first contribution values ​​for working days across all historical periods is then divided by the total number of working days across all historical periods to obtain a unified first historical baseline value, reflecting the average contribution level of working days after comprehensive statistics across multiple historical periods.

[0037] S37: Divide the sum of the second contribution values ​​by the total number of days in the holiday period to obtain the second historical baseline value for the holiday period.

[0038] In this step, the total number of holiday days across multiple historical periods is calculated. The sum of the second contribution values ​​for holiday periods across all historical periods is then divided by the total number of holiday days across all historical periods to obtain a unified second historical benchmark value, reflecting the average contribution level of holidays after comprehensive statistics across multiple historical periods.

[0039] S38: If the effective time information and record time information of any detailed data do not belong to the same statistical period, based on the date attribute information of each historical period and the contribution value of each detailed data, calculate the first historical benchmark value for the weekday period, the second historical benchmark value for the holiday period, and the cross-period effect correction parameter respectively.

[0040] In this step, if the effective date and record date of any detailed data do not belong to the same statistical period, it indicates the existence of a cross-period effect. For example, in the insurance business scenario, there are cross-month policies with signing dates and effective dates that are not in the same month. In this case, it is necessary to calculate the first historical benchmark value for weekday periods, the second historical benchmark value for holiday periods, and the cross-period effect correction parameter separately. This correction parameter is used to reflect the pulling effect of the cross-period effect on the indicator value. When generating the initial forecast value, the cross-period effect correction parameter is added as an independent correction item to the benchmark value forecast result to restore the true contribution of the cross-period effect to the indicator value, thereby improving the forecast accuracy.

[0041] In one embodiment of this application, a scheme for determining basic forecast parameters is provided when there is a specific cross-period effect. In step S38, based on the date attribute information of each historical period and the contribution value of each detailed data, the first forecast parameter for weekdays, the second forecast parameter for holidays, and the cross-period effect correction parameter are calculated respectively. Specifically, this includes the following steps S381-S386: S381: Based on the effective time information and record time information of each detailed data, identify the detailed data that have cross-period effects.

[0042] In this step, the cross-period effectiveness impact refers to situations where the effective date information and the recording date information of detailed data do not belong to the same statistical period. For example, in an insurance business scenario, if the signing date of a policy is June 30, 2025, and the effective date is July 1, 2025, then the signing date and the effective date of the policy do not belong to the same month, which means that the detailed data has a cross-period effectiveness impact. Detailed data whose recording time and effective time do not belong to the same statistical period are marked as detailed data with cross-period effectiveness impact.

[0043] S382: Remove the contribution values ​​of detailed data that have cross-period effects from the contribution values ​​of detailed data during weekday periods, and remove the contribution values ​​of detailed data that have cross-period effects from the contribution values ​​of detailed data during holiday periods.

[0044] In this step, since the cross-period effect may exist simultaneously during both weekdays and holidays, it is necessary to perform elimination processing in both time periods to ensure that the contribution values ​​participating in the statistics in each time period are pure contribution values ​​without cross-period effects. Specifically, based on the identified markers indicating whether each detailed data point has a cross-period effect, the contribution values ​​of detailed data points with cross-period effects are eliminated from the detailed data points corresponding to the weekday and holiday periods, respectively, so that they are not included in the subsequent calculation of the baseline value.

[0045] S383: Obtain the sum of the first contribution values ​​of all detailed data corresponding to the weekday time period after removal, and the sum of the second contribution values ​​of all detailed data corresponding to the holiday time period after removal.

[0046] In this step, after removing the contribution values ​​marked as having cross-period effects from all detailed data belonging to the working day period in all historical periods, the contribution values ​​of all remaining detailed data of the working day period are summed to obtain the sum of the first contribution values.

[0047] Similarly, in all detailed data belonging to holiday periods throughout all historical periods, after removing the contribution values ​​marked as having cross-period effects, the contribution values ​​of all remaining detailed data for holiday periods are summed to obtain the sum of the second contribution values.

[0048] S384: Divide the sum of the first contribution values ​​by the total number of days in the working day period to obtain the first historical baseline value for the working day period.

[0049] In this step, the total number of working days across multiple historical periods is calculated. The sum of the first contribution values ​​for working days across all historical periods is then divided by the total number of working days across all historical periods to obtain a unified first historical baseline value, reflecting the average contribution level of working days after comprehensive statistics across multiple historical periods.

[0050] S385: Divide the sum of the second contribution values ​​by the total number of days in the holiday period to obtain the second historical baseline value for the holiday period.

[0051] In this step, the total number of holiday days across multiple historical periods is calculated. The sum of the second contribution values ​​for holiday periods across all historical periods is then divided by the total number of holiday days across all historical periods to obtain a unified second historical benchmark value, reflecting the average contribution level of holidays after comprehensive statistics across multiple historical periods.

[0052] For example, if an insurance company's cumulative NBEV (New Business Value) indicator is used as an example, taking June 2023 as an example, the date attribute information shows that the month had 22 working days and 8 holidays (including weekends and the Dragon Boat Festival). The transaction details record the NBEV amount, signing date, and effective date of each policy in that month. First, based on the date attribute information, June 2023 is divided into two time periods: working days and holidays. The NBEV contribution value for the working day period and the NBEV contribution value for the holiday period are calculated separately. However, during the calculation process, by comparing the signing date and effective date of each policy, those cross-month designated effective policies with signing dates in June but effective dates on July 1st are identified. For example, policies signed between June 28th and 30th, 2023, and designated as effective on July 1st, have a total NBEV of 120 million yuan. Although these policies were signed in June, their NBEV value is not the true legal value for June. Therefore, the RMB 120 million of cross-month NBEVs is deducted from the total NBEVs for working days in June to accurately reflect the actual production capacity of each category during different periods, avoiding overestimation of historical daily average benchmarks by including NBEVs that should belong to the following month. Subsequently, the total NBEVs after deduction are divided by the number of working days in that month to obtain the corrected daily average NBEV for working days in June 2023. The same treatment is applied to holiday periods. If some cross-month policies were signed on the holiday of June 22, they are also deducted before calculating the corrected daily average NBEV for the holiday. The same calculation process is repeated for other historical months. The data from multiple months are aggregated and averaged to obtain the two basic forecast parameters: the benchmark daily average NBEV for working days and the benchmark daily average NBEV for holidays.

[0053] S386: Obtain the sum of the third contribution values ​​of the detailed data of the cross-period effect within each historical period, and use the sum of the third contribution values ​​as the correction parameter for the cross-period effect of each historical period.

[0054] In this step, for each historical period, all detailed data with inter-period effects are identified one by one, and the contribution value corresponding to each detailed data is extracted. The contribution values ​​of all inter-period detailed data are summed to obtain the sum of the third contribution values ​​for that historical period. The sum of the third contribution values ​​is directly used as the correction parameter for the inter-period effects of that historical period, and is then superimposed on the baseline prediction result when generating the retrospective prediction value for that historical period.

[0055] For example, in an insurance business scenario, if there are policies that take effect across months in June 2025, the sum of their NBEV contribution values ​​is the correction parameter for the cross-month effect in June 2025. This correction parameter is calculated independently of the correction parameters for other historical periods when calculating the retrospective forecast value for June 2025, and they do not affect each other.

[0056] S40: Based on the basic forecast parameters and the date attribute information of each historical period, back-forward forecasts are performed for each historical period to generate back-forward forecast values ​​of cumulative numerical indicators for each historical period.

[0057] In this step, retrospective forecasting refers to simulating and forecasting a historical period that has already occurred, using the same forecasting logic and the same set of basic forecasting parameters as the forecast period, to obtain the simulated forecast value corresponding to that historical period. Specifically, for any historical period, based on the date attribute information of that historical period, the number of days corresponding to weekdays and holidays is counted, multiplied by the first historical baseline value and the second historical baseline value respectively, and the cross-period effect correction parameter is added to obtain the retrospective forecast value of that historical period.

[0058] In one embodiment of this application, a specific backtracking prediction value generation scheme is provided. In step S40, based on the basic prediction parameters and the date attribute information of each historical period, backtracking prediction is performed on each historical period to generate the backtracking prediction value of the cumulative numerical index in each historical period. Specifically, this includes the following steps S41-S44: S41: Based on the date attribute information of each historical period, determine the number of the first day of the weekday period and the number of the second day of the holiday period in each historical period.

[0059] In this step, for any historical period, the date type label corresponding to each day in each historical period is determined based on the date attribute information. Based on the date type label of each day in the historical period, the number of days marked as working days and days marked as holidays in the historical period are counted respectively.

[0060] S42: Multiply the number of the first day of the weekday period in each historical period by the first historical baseline value to obtain the first contribution value estimate corresponding to the weekday period in each historical period.

[0061] In this step, for each historical period, the number of the first day is multiplied by the first historical baseline value to obtain the first contribution estimate of all working days in that historical period. This estimate reflects the contribution of the working day portion of that historical period to the cumulative numerical indicator in the absence of intertemporal effects.

[0062] S43: Multiply the number of the second day that belongs to the holiday period in each historical period by the second historical benchmark value to obtain the second contribution value estimate corresponding to the holiday period in each historical period.

[0063] In this step, for each historical period, the number of the second day is multiplied by the second historical baseline value to obtain the second contribution estimate of all holiday periods in that historical period. This estimate reflects the contribution of the holiday portion of that historical period to the cumulative numerical indicator without intertemporal effects.

[0064] S44: Add the first contribution value estimate, the second contribution value estimate, and the inter-period impact correction parameter for each historical period to generate the retrospective forecast value corresponding to each historical period.

[0065] In this step, for each historical period, the first contribution value estimate reflects the estimated contribution of the weekday portion to the cumulative numerical indicator within that period, the second contribution value estimate reflects the estimated contribution of the holiday portion to the cumulative numerical indicator within that period, and the inter-period effect correction parameter reflects the actual pull of the inter-period detailed data on the indicator value within that historical period. Adding these three together yields the retrospective forecast value for that historical period. This retrospective forecast value is a simulated estimate of the indicator value for that historical period based on the basic forecast parameters.

[0066] S50: Generate a residual sequence based on the difference between the actual value and the backtested predicted value for each historical period.

[0067] In this step, since the basic forecast parameters are fixed benchmark values ​​derived from the comprehensive statistics of multiple historical periods, they reflect the average production capacity level of these periods and cannot capture individual differences caused by factors such as changes in the business environment, market fluctuations, and policy adjustments. Therefore, when making backtesting forecasts for any historical period, there will inevitably be a deviation between the backtesting forecast value calculated based on this fixed benchmark value and the actual value of that historical period. Subtracting the corresponding backtesting forecast value from the actual value of each historical period yields the residual value for each historical period. Subsequently, the residual values ​​of each historical period are arranged in chronological order to generate a residual sequence. This residual sequence reflects the systematic forecasting deviation of the basic forecast parameters in different historical periods and serves as the data foundation for subsequent multi-dimensional residual forecasting.

[0068] S60: Based on multi-source time series data and date attribute information of the preset prediction period, residual prediction is performed on the residual sequence according to multiple preset dimensions to obtain multiple residual prediction values.

[0069] Among them, several preset dimensions include time sequence dimension, hierarchical structure dimension and category structure dimension.

[0070] In this step, residual prediction refers to using the residual sequence as the prediction object and employing a time series prediction model to predict the future trend of the residual sequence, thereby obtaining an estimate of the residual during the prediction period. When predicting the residual sequence, it is not done from a single perspective, but rather from multiple preset dimensions to comprehensively capture the deviation characteristics under different dimensions.

[0071] Specifically, the time-series dimension refers to the trend and periodicity of the residual sequence over time. Time-series modeling of the residual sequence is performed based on date attribute information to predict trend deviations in future periods. The hierarchical structure dimension refers to the decomposability of the residual sequence at the organizational or business level. Based on the hierarchical structure, the residual sequence is decomposed into multiple hierarchical components and predicted separately to capture micro-deviations at each level. The category structure dimension refers to the categoricalability of the residual sequence in terms of business categories. Based on the category structure, the residual sequence is decomposed into multiple category components and predicted separately to capture structural deviations within each category. Through predictions along these three dimensions, the resulting multiple residual prediction values ​​characterize the systematic deviations of the basic prediction parameters from different perspectives. These can then be used as deviation compensation amounts to correct the initial prediction values, achieving multi-dimensional and comprehensive deviation compensation.

[0072] In one embodiment of this application, a specific residual prediction value generation scheme is provided. In step S60, based on multi-source time series data and date attribute information of a preset prediction period, residual prediction is performed on the residual sequence according to multiple preset dimensions to obtain multiple residual prediction values. Specifically, this includes the following steps S61-S66: S61: Using the residual sequence and the date attribute information of each historical period as training data, train the time series residual prediction model corresponding to the time series dimension to obtain the trained time series residual prediction model.

[0073] In this step, the temporal residual prediction model is used to predict the residual sequence from a temporal dimension. Its training data includes two parts: training input and training labels. Specifically, the date attribute information corresponding to each historical period is used as the training input feature, and the residual value corresponding to each historical period is used as the training label. The residual value is the difference between the true value and the backtracked predicted value for each historical period. By constructing a training sample set with the date attribute information and the corresponding residual values ​​for each historical period, the temporal residual prediction model is trained, enabling the model to learn the mapping relationship between date attribute information and residuals.

[0074] During training, the time-series residual prediction model analyzes the correspondence between date attribute information and residual values ​​in each historical period, learning the influence of different date characteristics on prediction bias, thereby establishing a prediction function from date attribute information to residual values. After training, the time-series residual prediction model has the ability to predict the residual values ​​for a preset prediction period based on the date attribute information.

[0075] S62: Based on the data structure characteristics of the hierarchical structure dimension, the residual sequence is decomposed into multiple hierarchical components. After aligning each hierarchical component with the date attribute information of each historical period, a time series prediction model is trained for each hierarchical component to obtain the time series prediction model corresponding to each hierarchical component after training.

[0076] In this step, the hierarchical structure dimension's data structure characteristic refers to the tree-like decomposability of cumulative numerical indicators at the organizational or business level; that is, the value of a higher-level node is equal to the sum of the values ​​of all its lower-level child nodes. Based on this data structure characteristic, the residual sequence is decomposed level by level along a preset hierarchical path to obtain the residual components corresponding to each level node. For example, in an insurance business scenario, the residual of the entire system is first decomposed into the residuals of each institution, and then the residuals of each institution are decomposed into the drill-down residuals of each agent, forming a complete set of hierarchical components from the root node to the leaf nodes.

[0077] After decomposition, each hierarchical component is aligned with the date attribute information of each historical time period. That is, for each hierarchical component, its residual value in each historical time period is paired with the corresponding date attribute information of that historical time period. This ensures a one-to-one correspondence between the input features (date attribute information of that historical time period) and the training labels (residual values ​​of that hierarchical component in that historical time period), guaranteeing that the model can learn the mapping relationship between the two. After alignment, for each hierarchical component, its residual value in each historical time period is used as the training label, and its corresponding date attribute information of that historical time period is used as the training input feature. A time-series prediction model specific to that hierarchical component is then trained. Since the residual variation patterns of different hierarchical nodes may differ, training an independent prediction model for each hierarchical component allows the model to better adapt to the unique patterns of each hierarchical component, improving prediction accuracy. After training, each hierarchical component corresponds to a pre-trained time-series prediction model.

[0078] S63: Based on the data structure characteristics of the category structure dimension, the residual sequence is decomposed into multiple category components. After aligning each category component with the date attribute information of each historical period, a time series prediction model is trained for each category component to obtain the time series prediction model corresponding to each category component after training.

[0079] In this step, the data structure characteristic of the category structure dimension refers to the fact that the cumulative numerical indicators have categorical decomposability in terms of business type. That is, the overall value is equal to the sum of the values ​​of each parallel category, and the categories do not overlap and are combined to form a whole. According to this data structure characteristic, the residual sequence is decomposed into multiple category components according to preset classification rules. For example, in the insurance business scenario, the residual of the entire system is decomposed into wealth category residuals, health category residuals, accident category residuals, and life insurance category residuals.

[0080] After decomposition, each category component is aligned with the date attribute information of each historical time period. That is, for each category component, its residual value in each historical time period is paired with the corresponding date attribute information of that historical time period, ensuring a one-to-one correspondence between the input features in the training data (the date attribute information of that historical time period) and the training labels (the residual values ​​of that category in that historical time period), guaranteeing that the model can learn the mapping relationship between the two. After alignment, for each category component, its residual value in each historical time period is used as the training label, and its corresponding date attribute information of each historical time period is used as the training input feature, training a time-series prediction model specifically for that category component. Since the residual variation patterns of different categories may differ, training an independent prediction model for each category component allows the model to better adapt to the unique patterns of each category, improving prediction accuracy. After training, each category component corresponds to a pre-trained time-series prediction model.

[0081] S64: Input the date attribute information of the preset prediction period into the trained time-series residual prediction model to obtain the first residual prediction value.

[0082] In this step, the time-series residual prediction model has learned the mapping relationship between the date attribute information of historical time periods and the residual values ​​during the training phase. The date attribute information of the preset prediction period is used as input features and fed into the trained time-series residual prediction model. Based on the learned time-series patterns, the model outputs the residual prediction value for that period in the time-series dimension, denoted as the first residual prediction value. This residual prediction value reflects the potential trend deviation of the basic prediction parameters in the preset prediction period within the time-series dimension.

[0083] S65: Input the date attribute information of the preset prediction period into the time series prediction model corresponding to each level component after training, and merge the prediction results to obtain the second residual prediction value.

[0084] In this step, the date attribute information of the preset prediction period is used as input features and input into the time series prediction model corresponding to each level component. Each model independently outputs the residual prediction value of that level component in the preset prediction period. For example, in an insurance business scenario, the date attribute information of the prediction month is input into the time series prediction model corresponding to each agency and each level agent to obtain the residual prediction value for each agency and each level.

[0085] Furthermore, after obtaining the residual prediction values ​​for each level component, the prediction results of each level component are merged level by level according to the data structure characteristics of the hierarchical structure dimension. The resulting hierarchical residual prediction value for the entire system after merging is denoted as the second residual prediction value. This prediction value reflects the micro-deviations that the basic prediction parameters may produce during the preset prediction period under the hierarchical structure dimension.

[0086] Optionally, the prediction results of leaf nodes (such as agents at each drill level) are first merged to their parent node (such as the organization), and then the prediction results of each organization level are merged to the root node (such as the entire system), finally obtaining the residual prediction value of the entire system under the hierarchical structure dimension. This merging process is the reverse operation of the decomposition process during the training phase, ensuring that the correlation between components at each level remains consistent.

[0087] S66: Input the date attribute information of the preset prediction period into the time series prediction model corresponding to each category component after training, and merge the prediction results to obtain the third residual prediction value.

[0088] In this step, the date attribute information of the preset prediction period is used as input features and input into the time series prediction model corresponding to each category component. Each model independently outputs the residual prediction value of that category component in the preset prediction period. For example, in the insurance business scenario, the date attribute information of the prediction month is input into the time series prediction models corresponding to wealth, health, accident, and life insurance categories respectively to obtain the residual prediction value for each product category.

[0089] Furthermore, after obtaining the residual predicted values ​​for each category component, the prediction results for each category component are merged according to the data structure characteristics of the category structure dimension. The merged system-wide category residual predicted value is denoted as the third residual predicted value. This predicted value reflects the structural bias that the basic prediction parameters may produce during the preset prediction period under the category structure dimension.

[0090] Optionally, since the categories are parallel and do not overlap, the merging method involves directly adding the residual prediction values ​​of each category to obtain the residual prediction value of the entire system under the category structure dimension. This merging process is the inverse operation of the decomposition process during the training phase, ensuring that the components of each category remain structurally consistent with the overall residual sequence after merging.

[0091] S70: Based on the basic forecast parameters and the date attribute information of the preset forecast period, generate the initial forecast value of the cumulative numerical index for the preset forecast period.

[0092] In this step, the date attribute information of the preset forecast period is obtained from the calendar. Before the actual value of the forecast period is obtained, the number of days corresponding to weekdays and holidays within the preset forecast period is calculated based on the date attribute information. These are then multiplied by their respective historical baseline values ​​and the cross-period effect correction parameter is added to obtain the initial forecast value.

[0093] In one embodiment of this application, a specific initial prediction value generation scheme is provided. In step S70, that is, based on the basic prediction parameters and the date attribute information of the preset prediction period, the initial prediction value of the cumulative numerical index in the preset prediction period is generated, which specifically includes the following steps S71-S75: S71: Based on the date attribute information of the preset prediction period, determine the third day that belongs to the weekday period and the fourth day that belongs to the holiday period within the preset prediction period.

[0094] In this step, based on the date type label of each day in the date attribute information of the preset prediction period, the number of days marked as working days and the number of days marked as holidays in that period are counted respectively.

[0095] S72: Multiply the third day number by the first historical baseline value to obtain the estimate of the third contribution value corresponding to the weekday period.

[0096] In this step, the third day number is multiplied by the first historical benchmark value to obtain the third contribution value estimate for all working days within the preset forecast period, which is used to reflect the contribution estimate of the working day portion of the period to the cumulative numerical index.

[0097] S73: Multiply the fourth day number by the second historical baseline value to obtain the fourth contribution value estimate corresponding to the holiday period.

[0098] In this step, the fourth day is multiplied by the second historical baseline value to obtain the fourth contribution value estimate for all holiday periods within the preset prediction period, which is used to reflect the contribution estimate of the holiday portion of the period to the cumulative numerical index.

[0099] Optionally, it can be determined in advance whether the cross-period effect correction parameter is zero. If the cross-period effect correction parameter is not zero, the number of days is multiplied by the historical benchmark value with cross-period effect; if the cross-period effect correction parameter is zero, the number of days is multiplied by the historical benchmark value without cross-period effect.

[0100] S74: If the cross-period effect correction parameter is not zero, calculate the average value of the cross-period effect correction parameter for multiple historical periods.

[0101] In this step, when inter-period effects exist in historical time periods, each historical time period corresponds to an independent inter-period effect correction parameter, the value of which is equal to the sum of the contribution values ​​of all detailed data with inter-period effects within that historical time period. Since the composition of inter-period detailed data differs across historical time periods, the inter-period effect correction parameters may vary. To reasonably estimate the magnitude of the inter-period effect in the forecast period, the inter-period effect correction parameters from multiple historical time periods need to be aggregated, and their arithmetic mean calculated as a benchmark reference value for the inter-period effect correction in the forecast period. This average value reflects the average contribution level of the inter-period detailed data across multiple historical time periods, providing a reasonable estimate of the inter-period effect in the forecast period when the inter-period detailed data cannot be known in advance. By using this average value as the inter-period effect correction parameter for the forecast period, the inter-period correction portion in the initial forecast value can be kept consistent with the average level of inter-period effects in historical time periods, thereby achieving reasonable compensation for the inter-period effect in a statistical sense and improving the accuracy of the initial forecast value.

[0102] S75: Sum the third contribution value estimate, the fourth contribution value estimate, and the average value to generate the initial forecast value of the cumulative numerical index for the preset forecast period.

[0103] In this step, the third contribution estimate reflects the estimated contribution of the weekday portion of the forecast period to the cumulative numerical index, the fourth contribution estimate reflects the estimated contribution of the holiday portion of the forecast period to the cumulative numerical index, and the average value reflects the estimated average contribution level of the inter-period impact within the forecast period. Summing these three values ​​yields the initial forecast value of the cumulative numerical index for the preset forecast period.

[0104] S80: Using a Bayesian fusion algorithm, multiple residual prediction values ​​are used as deviation compensation amounts to compensate for the deviation of the initial prediction value, generating the target prediction value of the cumulative numerical index within the preset prediction period.

[0105] In this step, the initial predicted value and multiple residual predicted values ​​are used as input features. A Bayesian linear regression model is used to determine the weight coefficients corresponding to each input feature. These weight coefficients measure the importance of each input feature to the final prediction result. A larger weight coefficient indicates a greater contribution of the input feature to the final prediction result; a smaller weight coefficient indicates a smaller contribution. By assigning appropriate weights to the initial predicted value and each residual predicted value, complementary advantages among different input features can be achieved. Subsequently, a weighted fusion is performed according to the weight coefficients to obtain the target predicted value for the preset prediction period.

[0106] In one embodiment of this application, a specific initial prediction value deviation compensation scheme is provided. In step S80, that is, using a Bayesian fusion algorithm, multiple residual prediction values ​​are used as deviation compensation amounts to compensate for the deviation of the initial prediction value, generating the target prediction value of the cumulative numerical index in the preset prediction period. Specifically, it includes the following steps S81-S84: S81: Obtain the scene identifier information of the preset prediction time period.

[0107] In this step, the scenario identification information is used to indicate the type of business scenario to which the preset prediction period belongs. The historical performance of each residual prediction model varies under different business scenarios, therefore different fusion strategies need to be adapted according to the scenario type.

[0108] Optionally, the scene identification information can be automatically determined based on the date attribute information of the preset prediction period, or it can be obtained based on the scene tags preset by the user. For example, business scene types include peak sales periods, off-season periods, normal periods, and end-of-quarter sales push periods.

[0109] S82: Based on the scene identification information, determine the prior weight distribution parameters corresponding to each residual prediction value.

[0110] In this step, the prior weight distribution parameters refer to the initial weight distribution pre-set for each input feature (i.e., each residual prediction value and initial prediction value) based on historical experience before Bayesian fusion. Different prediction models exhibit different historical performances in different business scenarios, therefore, different prior weight distributions need to be set for different scenarios. Based on the scenario identification information, the prior weight distribution parameters corresponding to the current scenario are matched from the pre-set parameters. These prior weight distribution parameters include a mean parameter and a variance parameter. The mean parameter reflects the expected contribution of the input feature to the final prediction result in the corresponding scenario; a larger mean value indicates a higher base weight assigned to the input feature during fusion. The variance parameter reflects the stability of the input feature in historical predictions; a smaller variance value indicates higher reliability of the input feature.

[0111] For example, during peak sales periods, if the historical performance of the time-series residual prediction model is better than other models, a higher mean parameter is assigned to it; conversely, during off-peak periods, if the historical performance of the hierarchical structure dimension residual prediction model is better, a higher mean parameter is assigned to it. By configuring prior weight distribution parameters for different scenarios, the initial constraints of subsequent Bayesian fusion can be adapted to the data characteristics of the current scenario.

[0112] S83: Based on the initial predicted value, the first residual predicted value, the second residual predicted value, the third residual predicted value, and the prior weight distribution parameters, the posterior distribution is sampled using the Markov chain Monte Carlo algorithm to determine the weight coefficients corresponding to the initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value, respectively.

[0113] In this step, the prior weight distribution parameters are used as the initial constraints for the weight coefficients of each input feature. The initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value are used as input features. The posterior distribution is sampled using the Markov chain Monte Carlo algorithm, and the weight parameters corresponding to each input feature are iteratively updated until the sampling results converge to a stable posterior distribution. The initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value together constitute the input feature set of Bayesian fusion. Each input feature reflects its contribution to the final prediction result from different perspectives. Specifically, the initial predicted value is the basic prediction result generated based on the basic prediction parameters; the first residual predicted value reflects the trend bias estimate under the time series dimension; the second residual predicted value reflects the micro bias estimate under the hierarchical structure dimension; and the third residual predicted value reflects the structural bias estimate under the category structure dimension. Subsequently, the posterior expected value corresponding to each input feature is extracted from the stable posterior distribution and used as the optimal weight coefficient of each input feature, so that the weight coefficients corresponding to the initial predicted value, the first residual predicted value, the second residual predicted value and the third residual predicted value are determined respectively.

[0114] S84: According to each weight coefficient, the initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value are weighted and summed, and the weighted sum is used as the target predicted value of the cumulative numerical index during the prediction period.

[0115] In this step, the initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value are multiplied by their respective weighting coefficients and then summed. These three residual predicted values ​​are then added to the initial predicted value as bias compensation, achieving multi-dimensional bias compensation for the initial predicted value. Since each weighting coefficient is adaptively determined based on Bayesian posterior probability, the weighted summation result can integrate the advantages of each input feature, resulting in a final output target predicted value with higher accuracy and stability compared to a single prediction method.

[0116] The following example, using the prediction of an insurance company's NBEV in month N of year M, illustrates the deviation compensation method for time-series prediction data proposed in this application, specifically including the following: like Figure 3The diagram illustrates a method for compensating for deviations in time-series forecast data. First, historical NBEV data for each month N from year M-3 to M-1 is acquired, along with daily workday and holiday markers for each month, and the signing date, effective date, and NBEV contribution value for each policy. Based on historical production capacity patterns, the average daily NBEV for all workdays and holidays in month N of each year is calculated, and then aggregated to obtain the average daily benchmark values ​​for workdays and holidays. Simultaneously, policies with signing and effective dates not belonging to the same month are identified, their historical proportion is calculated, and the cross-month contribution coefficient is obtained through exponential smoothing. Using the base forecaster, the initial forecast value X0 is calculated based on the number of workdays, holidays, and cross-month effective correction parameters for month N in year M. Subsequently, the actual NBEV for each historical month N is subtracted from the backtested forecast value for that month by the base forecaster, generating a residual sequence. By employing three parallel residual prediction dimensions—SARIMAX (Seasonal Autoregressive Integrated Moving Average with eXogenous regressors) time-series residual prediction, agent-level residual decomposition, and product-dimensional residual decomposition—the model uses the residual sequence of backtracking predicted values ​​and historical true values ​​as common input. It outputs residual prediction values ​​ε1, ε2, and ε3 from three dimensions: time-series trend, micro-level human resources, and product structure. Specifically, the SARIMAX time-series residual prediction model predicts residuals from the time-series trend dimension, compensating for trend bias; the agent-level residual decomposition model decomposes residuals level by level according to institutional hierarchy to agents at each level, predicts them separately, and then back-aggregates them to compensate for biases at the micro-level human resources level; and the product-dimensional residual decomposition model decomposes residuals into product categories such as wealth management, health, accident, and life insurance, predicts them separately, and then merges them to compensate for biases caused by product structure adjustments. Subsequently, based on prior information from business scenarios such as the "opening sales season," a priori weight distribution for each residual prediction value is set. The posterior distribution is then determined using MCMC (Markov Chain Monte Carlo) sampling combined with Bayesian linear regression to determine the initial prediction value X0 and the optimal weights for each of the three residual prediction values. Specifically, the prior distribution N (0.3, 0.05) of the residual weights during the "opening sales season" is introduced. The posterior distribution is sampled using the Markov Chain Monte Carlo algorithm to solve for the optimal weight combination ω0, ω1, ω2, ω3, satisfying ω0 + ω1 + ω2 + ω3 = 1. Finally, after weighted fusion, the final target prediction value for NBEV in year M and month N is output as ω0X0 + ω1ε1 + ω2ε2 + ω3ε3.Through the coordinated operation of the three-layer architecture of basic prediction, residual prediction, and Bayesian fusion, a closed-loop full-link system is achieved, which includes basic value → residual calculation → correction value → fused prediction value. This ensures that the final prediction result retains the basic production capacity pattern of specific periods while incorporating refined compensation from multi-dimensional residuals. Furthermore, by optimizing the dynamic weights of Bayesian posterior probabilities, the shortcomings of traditional fixed-weight fusion in responding flexibly under specific periods are effectively overcome. At the same time, the contribution of each input feature is traceable, ensuring the interpretability of the prediction results.

[0117] like Figure 4 The diagram illustrates the monthly NBEV production capacity distribution across different time periods. Based on the average production capacity for each period using historical data from the past three years, significant time-specific variations in NBEV production capacity are observed within each month. Production capacity is lowest on the 1st of the month; it remains stable from the 10th to the 25th; and peaks in production capacity are concentrated in the last five days of the month, representing the peak period for month-end sales pushes. Production capacity is relatively stable on weekdays, significantly lower on weekends, and further declines during holidays. The proportion of policies taking effect across months increases nearly tenfold during the "opening sales" period, significantly boosting production capacity at the beginning of the month. This time-specific production capacity distribution pattern is used in the basic forecast, combining the actual number of weekdays and holidays in the forecast month to allocate production capacity, allowing the basic forecast values ​​to adapt to the business rhythm of different months.

[0118] like Figure 5 The diagram illustrates the residual prediction process at the agent hierarchy level. In the agent drill-level residual prediction process, the entire system's residual sequence is first divided into 44 institutional nodes based on the capacity share of each institution. Then, the residuals of each institution are further subdivided according to agent drill levels (including non-drill, single drill, silver drill, gold drill, and double gold drill), obtaining the residual subsequences corresponding to each drill level. Finally, the residuals at each drill level are decomposed into three interpretable core indicators: effective activity rate residual, average NBEV residual, and number of employees residual. After predicting these three indicators using the SARIMAX algorithm, the residuals are back-collected according to the reconciliation relationship of drill-level residual = number of employees residual × effective activity rate residual × average NBEV residual, and then rolled up level by level to obtain the predicted residual values ​​for the entire system. Through this closed-loop process of hierarchical decomposition and back-collection, micro-level deviation compensation is achieved. When the final residual prediction value is abnormal, the cross-reference relationship can be used to trace back to the changes in core indicators such as the number of agents and activity rate of a certain diamond-level agent in a specific institution. This provides a precise direction for business optimization, rather than relying solely on a rough judgment at the overall level.

[0119] like Figure 6The diagram shown illustrates the residual prediction process for the product category structure dimension. During the product dimension residual prediction process, a pie chart is used to display the residual percentage distribution and residual volatility characteristics of the four major product categories: pension, security, wealth, and other. Specifically, the residual volatility of wealth products is 30%-40%, significantly affected by holidays (residual volatility greater than 0.9); the residual volatility of pension and security products is 15%-20%, with relatively moderate impact from holidays (residual volatility less than 0.8); and the residual volatility of other product categories is only 5%-10%, with virtually no special influence (residual volatility less than 0.1). To address the bias in the basic prediction caused by insufficient product structure adaptation, a product dimension residual prediction model is used for specific compensation. The residual sequence is decomposed according to product category, and each category's residual value is independently predicted using a SARIMAX model. This ensures that product categories with higher volatility (such as wealth) receive differentiated prediction parameters, avoiding averaging by categories with lower volatility. Then, the residual prediction values ​​for each category are summed to obtain the total residual prediction value ε3 for the product dimension, which is then input into the Bayesian fusion layer to compensate for structural biases caused by product structure adjustments. This achieves accurate capture and differentiated compensation for changes in product structure, effectively improving prediction accuracy in scenarios with frequent product structure adjustments.

[0120] As can be seen, the above scheme, by constructing a three-layer architecture of basic prediction, residual compensation, and Bayesian dynamic fusion, first generates initial predicted values ​​using differentiated benchmarks for weekdays and holidays and inter-period corrections. Then, it predicts residual sequences from three dimensions: time series trend, hierarchical structure, and category structure. Finally, the optimal weights are adaptively determined through a Bayesian fusion algorithm to achieve multi-dimensional bias compensation. This effectively solves the problems of coarse granularity, lack of correction mechanisms, and inability to adaptively fuse traditional contemporaneous mean methods. It achieves end-to-end optimization from bias detection and tracing to precise compensation, significantly improving the accuracy, stability, and interpretability of predictions. In financial scenarios, it helps improve the accuracy of performance forecasting and capital allocation, supporting adjustments to investment strategies; in medical scenarios, it can effectively improve the stability of medical demand forecasting.

[0121] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0122] In one embodiment, a deviation compensation device for time series prediction data is provided, which corresponds one-to-one with the deviation compensation method for time series prediction data in the above embodiments. For example... Figure 7As shown, the time-series prediction data deviation compensation device 100 includes a first acquisition module 101, a second acquisition module 102, a determination module 103, a first generation module 104, a second generation module 105, a third generation module 106, a fourth generation module 107, and a fifth generation module 108. Detailed descriptions of each functional module are as follows: The first acquisition module 101 is used to acquire the preset prediction period and the cumulative numerical indicators to be predicted; The second acquisition module 102 is used to acquire multi-source time series data of multiple historical periods that are concurrent with the preset prediction period. The multi-source time series data includes the actual value of the cumulative numerical indicator in each historical period, the date attribute information of each historical period, and the contribution value of each detailed data corresponding to the cumulative numerical indicator in each historical period. The determination module 103 is used to determine the basic forecast parameters based on multi-source time series data. The basic forecast parameters include at least one of the following: the first historical benchmark value corresponding to the weekday period, the second historical benchmark value corresponding to the holiday period, and the cross-period effect correction parameter. The first generation module 104 is used to perform back-forward prediction for each historical period based on the basic prediction parameters and the date attribute information of each historical period, and generate the back-forward prediction value of the cumulative numerical index in each historical period. The second generation module 105 is used to generate a residual sequence based on the difference between the actual value and the backtracked predicted value for each historical period. The third generation module 106 is used to perform residual prediction on the residual sequence based on multi-source time series data and date attribute information of the preset prediction period according to multiple preset dimensions to obtain multiple residual prediction values. Among them, the multiple preset dimensions include time series dimension, hierarchical structure dimension and category structure dimension. The fourth generation module 107 is used to generate the initial forecast value of the cumulative numerical index in the preset forecast period based on the basic forecast parameters and the date attribute information of the preset forecast period. The fifth generation module 108 is used to compensate for the deviation of the initial prediction value by using multiple residual prediction values ​​as deviation compensation amounts through a Bayesian fusion algorithm, and to generate the target prediction value of the cumulative numerical index in the preset prediction period.

[0123] In one embodiment, the determining module 103 is specifically used for: Based on the date attribute information of each historical period, the dates within each historical period are divided into weekday periods and holiday periods; Obtain the effective time and recording time information for each detailed data item within each historical time period; Determine whether the effective time information and recording time information of each detailed data in each historical period belong to the same statistical period; If the effective time information and record time information of all detailed data belong to the same statistical period, the correction parameter for cross-period effective impact is determined to be zero. Obtain the sum of the first contribution values ​​of all detailed data belonging to weekday periods within multiple historical time periods, and the sum of the second contribution values ​​of all detailed data belonging to holiday periods; Divide the sum of the first contribution values ​​by the total number of working days to obtain the first historical baseline value for the working day period; Divide the sum of the second contribution values ​​by the total number of days in the holiday period to obtain the second historical baseline value for the holiday period; If the effective time information and record time information of any detailed data do not belong to the same statistical period, the first historical benchmark value for weekday periods, the second historical benchmark value for holiday periods, and the cross-period effect correction parameter are calculated based on the date attribute information of each historical period and the contribution value of each detailed data.

[0124] In one embodiment, the determining module 103 is further configured to: Based on the effective time and record time information of each detailed data item, identify the detailed data items that have cross-period effects; Remove the contribution values ​​of detailed data that have cross-period effects from the contribution values ​​of detailed data during weekday periods, and remove the contribution values ​​of detailed data that have cross-period effects from the contribution values ​​of detailed data during holiday periods; Get the sum of the first contribution values ​​of all detailed data corresponding to the weekday time period after removal, and the sum of the second contribution values ​​of all detailed data corresponding to the holiday time period after removal; Divide the sum of the first contribution values ​​by the total number of working days to obtain the first historical baseline value for the working day period; Divide the sum of the second contribution values ​​by the total number of days in the holiday period to obtain the second historical baseline value for the holiday period; The sum of the third contribution values ​​of the detailed data on inter-period impacts within each historical period is obtained, and the sum of the third contribution values ​​is used as the correction parameter for the inter-period impacts of each historical period.

[0125] In one embodiment, the first generation module 104 is specifically used for: Based on the date attribute information of each historical period, the number of the first day of the weekday period and the number of the second day of the holiday period are determined respectively. Multiply the number of the first day of the weekday period in each historical period by the first historical baseline value to obtain the first contribution value estimate corresponding to the weekday period in each historical period; Multiply the number of the second day that belongs to the holiday period in each historical period by the second historical baseline value to obtain the second contribution value estimate corresponding to the holiday period in each historical period; The first contribution value estimate, the second contribution value estimate, and the intertemporal impact correction parameters for each historical period are added together to generate the retrospective forecast value for each historical period.

[0126] In one embodiment, the third generation module 106 is specifically used for: Using the residual sequence and date attribute information of each historical period as training data, the time series residual prediction model corresponding to the time series dimension is trained to obtain the trained time series residual prediction model. Based on the data structure characteristics of the hierarchical structure dimension, the residual sequence is decomposed into multiple hierarchical components. After aligning each hierarchical component with the date attribute information of each historical period, a time series prediction model is trained for each hierarchical component to obtain the time series prediction model corresponding to each hierarchical component after training. Based on the data structure characteristics of the category structure dimension, the residual sequence is decomposed into multiple category components. After aligning each category component with the date attribute information of each historical period, a time series prediction model is trained for each category component to obtain the time series prediction model corresponding to each category component after training. Input the date attribute information of the preset prediction period into the trained time-series residual prediction model to obtain the first residual prediction value; The date attribute information of the preset prediction period is input into the time series prediction model corresponding to each level component after training, and the prediction results are merged to obtain the second residual prediction value. The date attribute information of the preset prediction period is input into the time series prediction model corresponding to each category component after training, and the prediction results are merged to obtain the third residual prediction value.

[0127] In one embodiment, the fourth generation module 107 is specifically used for: Based on the date attribute information of the preset prediction period, determine the third day that belongs to the weekday period and the fourth day that belongs to the holiday period within the preset prediction period; Multiplying the third day number by the first historical baseline value yields the estimate of the third contribution value corresponding to the weekday period; Multiplying the fourth day number by the second historical baseline value yields the estimate of the fourth contribution value corresponding to the holiday period; If the cross-period effect correction parameter is not zero, calculate the average value of the cross-period effect correction parameter for multiple historical periods; The third contribution value estimate, the fourth contribution value estimate, and the average value are summed to generate the initial predicted value of the cumulative numerical index for the preset prediction period.

[0128] In one embodiment, the fifth generation module 108 is specifically used for: Obtain the scene identifier information of the preset prediction time period; Based on the scene identification information, determine the prior weight distribution parameters corresponding to each residual prediction value; Based on the initial predicted value, the first residual predicted value, the second residual predicted value, the third residual predicted value, and the prior weight distribution parameters, the posterior distribution is sampled using the Markov chain Monte Carlo algorithm to determine the weight coefficients corresponding to the initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value, respectively. According to each weight coefficient, the initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value are weighted and summed, and the weighted sum is used as the target predicted value of the cumulative numerical index during the prediction period.

[0129] This invention provides a bias compensation device for time-series forecast data. It constructs a three-layer architecture: basic forecasting, residual compensation, and Bayesian dynamic fusion. First, it generates initial forecast values ​​using differentiated benchmarks for weekdays and holidays and inter-period corrections. Then, it forecasts residual sequences from three dimensions: time-series trend, hierarchical structure, and category structure. Finally, a Bayesian fusion algorithm adaptively determines the optimal weights, achieving multi-dimensional bias compensation. This effectively solves the problems of coarse granularity, lack of correction mechanisms, and inability to adaptively fuse traditional contemporaneous mean methods. It achieves end-to-end optimization from bias detection and tracing to precise compensation, significantly improving the accuracy, stability, and interpretability of forecasts. In financial scenarios, it helps improve the accuracy of performance forecasting and capital allocation, supporting adjustments to investment strategies; in medical scenarios, it effectively improves the stability of forecasting medical needs.

[0130] Specific limitations regarding the deviation compensation device for time series forecast data can be found in the limitations of the deviation compensation method for time series forecast data described above, and will not be repeated here. Each module in the aforementioned deviation compensation device for time series forecast data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0131] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a time-series prediction data deviation compensation method on the server side.

[0132] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of a time-series forecast data deviation compensation method.

[0133] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain the preset forecast period and the cumulative numerical indicators to be predicted; Acquire multi-source time series data from multiple historical periods that coincide with the preset prediction period. The multi-source time series data includes the actual values ​​of cumulative numerical indicators in each historical period, date attribute information for each historical period, and the contribution value of each detailed data item corresponding to the cumulative numerical indicators in each historical period. Based on multi-source time series data, basic forecast parameters are determined, including at least one of the following: the first historical baseline value corresponding to weekday periods, the second historical baseline value corresponding to holiday periods, and the cross-period effect correction parameter; Based on the basic prediction parameters and the date attribute information of each historical period, back-forward predictions are made for each historical period to generate back-forward prediction values ​​of cumulative numerical indicators for each historical period. A residual sequence is generated based on the difference between the actual value and the backtested predicted value for each historical period. Based on multi-source time series data and date attribute information of the preset prediction period, residual prediction is performed on the residual sequence according to multiple preset dimensions to obtain multiple residual prediction values. Among them, the multiple preset dimensions include time series dimension, hierarchical structure dimension and category structure dimension. Based on the basic forecast parameters and the date attribute information of the preset forecast period, the initial forecast value of the cumulative numerical index is generated for the preset forecast period. By using a Bayesian fusion algorithm, multiple residual prediction values ​​are used as deviation compensation amounts to compensate for the deviation of the initial prediction value, thereby generating the target prediction value of the cumulative numerical index within the preset prediction period.

[0134] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain the preset forecast period and the cumulative numerical indicators to be predicted; Acquire multi-source time series data from multiple historical periods that coincide with the preset prediction period. The multi-source time series data includes the actual values ​​of cumulative numerical indicators in each historical period, date attribute information for each historical period, and the contribution value of each detailed data item corresponding to the cumulative numerical indicators in each historical period. Based on multi-source time series data, basic forecast parameters are determined, including at least one of the following: the first historical baseline value corresponding to weekday periods, the second historical baseline value corresponding to holiday periods, and the cross-period effect correction parameter; Based on the basic prediction parameters and the date attribute information of each historical period, back-forward predictions are made for each historical period to generate back-forward prediction values ​​of cumulative numerical indicators for each historical period. A residual sequence is generated based on the difference between the actual value and the backtested predicted value for each historical period. Based on multi-source time series data and date attribute information of the preset prediction period, residual prediction is performed on the residual sequence according to multiple preset dimensions to obtain multiple residual prediction values. Among them, the multiple preset dimensions include time series dimension, hierarchical structure dimension and category structure dimension. Based on the basic forecast parameters and the date attribute information of the preset forecast period, the initial forecast value of the cumulative numerical index is generated for the preset forecast period. By using a Bayesian fusion algorithm, multiple residual prediction values ​​are used as deviation compensation amounts to compensate for the deviation of the initial prediction value, thereby generating the target prediction value of the cumulative numerical index within the preset prediction period.

[0135] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0136] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0137] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0138] It should be noted that any neural network models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been obtained with the authorization (knowledge and consent) of the relevant parties or with full trust from all parties, and the executing entity can obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0139] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for compensating for deviations in time series forecast data, characterized in that, include: Obtain the preset forecast period and the cumulative numerical indicators to be predicted; Acquire multi-source time series data for multiple historical periods concurrent with the preset prediction period. The multi-source time series data includes the actual value of the cumulative numerical indicator in each historical period, the date attribute information of each historical period, and the contribution value of each detailed data item corresponding to the cumulative numerical indicator in each historical period. Based on the multi-source time series data, basic prediction parameters are determined, wherein the basic prediction parameters include at least one of the following: a first historical baseline value corresponding to weekday periods, a second historical baseline value corresponding to holiday periods, and a cross-period effect correction parameter; Based on the aforementioned basic prediction parameters and the date attribute information of each historical period, backtracking predictions are performed for each historical period to generate backtracking prediction values ​​of cumulative numerical indicators for each historical period. A residual sequence is generated based on the difference between the actual value and the backtracked predicted value for each historical period. Based on the multi-source time series data and the date attribute information of the preset prediction period, residual prediction is performed on the residual sequence according to multiple preset dimensions to obtain multiple residual prediction values. The multiple preset dimensions include time series dimension, hierarchical structure dimension and category structure dimension. Based on the basic prediction parameters and the date attribute information of the preset prediction period, the initial predicted value of the cumulative numerical index is generated in the preset prediction period. By using a Bayesian fusion algorithm, multiple residual prediction values ​​are used as deviation compensation amounts to compensate for the deviation of the initial prediction value, thereby generating the target prediction value of the cumulative numerical index within the preset prediction period.

2. The deviation compensation method for time series prediction data as described in claim 1, characterized in that, The step of determining the basic prediction parameters based on the multi-source time series data specifically includes: Based on the date attribute information of each historical period, the dates within each historical period are divided into weekday periods and holiday periods; Obtain the effective time and recording time information for each detailed data item within each historical time period; Determine whether the effective time information and recording time information of each detailed data in each historical period belong to the same statistical period; If the effective time information and record time information of all detailed data belong to the same statistical period, the correction parameter for cross-period effective impact is determined to be zero. Obtain the sum of the first contribution values ​​of all detailed data belonging to weekday periods within multiple historical time periods, and the sum of the second contribution values ​​of all detailed data belonging to holiday periods; Divide the sum of the first contribution values ​​by the total number of working days to obtain the first historical baseline value for the working day period; Divide the sum of the second contribution values ​​by the total number of days in the holiday period to obtain the second historical baseline value for the holiday period; If the effective time information and record time information of any detailed data do not belong to the same statistical period, the first historical benchmark value for weekday periods, the second historical benchmark value for holiday periods, and the cross-period effect correction parameter are calculated based on the date attribute information of each historical period and the contribution value of each detailed data.

3. The deviation compensation method for time series prediction data as described in claim 2, characterized in that, The steps for calculating the first historical baseline value for weekday periods, the second historical baseline value for holiday periods, and the cross-period effect correction parameter based on the date attribute information of each historical period and the contribution value of each detailed data point specifically include: Based on the effective time and record time information of each detailed data item, identify the detailed data items that have cross-period effects; Remove the contribution values ​​of detailed data that have cross-period effects from the contribution values ​​of detailed data during weekday periods, and remove the contribution values ​​of detailed data that have cross-period effects from the contribution values ​​of detailed data during holiday periods; Get the sum of the first contribution values ​​of all detailed data corresponding to the weekday time period after removal, and the sum of the second contribution values ​​of all detailed data corresponding to the holiday time period after removal; Divide the sum of the first contribution values ​​by the total number of working days to obtain the first historical baseline value for the working day period; Divide the sum of the second contribution values ​​by the total number of days in the holiday period to obtain the second historical baseline value for the holiday period; The sum of the third contribution values ​​of detailed data on inter-period impacts within each historical period is obtained, and the sum of the third contribution values ​​is used as the inter-period impact correction parameter for each historical period.

4. The method for compensating for deviations in time-series prediction data as described in claim 1, characterized in that, The step of generating backtracking predictions for each historical period based on the basic prediction parameters and the date attribute information of each historical period, and generating backtracking prediction values ​​for the cumulative numerical indicators in each historical period, specifically includes: Based on the date attribute information of each historical period, the number of the first day of the weekday period and the number of the second day of the holiday period are determined respectively. Multiply the number of the first day of the weekday period in each historical period by the first historical baseline value to obtain the first contribution value estimate corresponding to the weekday period in each historical period; Multiply the number of the second day that belongs to the holiday period in each historical period by the second historical baseline value to obtain the second contribution value estimate corresponding to the holiday period in each historical period; The first contribution value estimate, the second contribution value estimate, and the intertemporal impact correction parameters for each historical period are added together to generate the retrospective forecast value for each historical period.

5. The method for compensating for deviations in time-series prediction data as described in claim 1, characterized in that, The step of performing residual prediction on the residual sequence according to multiple preset dimensions based on the multi-source time series data and the date attribute information of the preset prediction period to obtain multiple residual prediction values ​​specifically includes: Using the residual sequence and date attribute information of each historical period as training data, the time series residual prediction model corresponding to the time series dimension is trained to obtain the trained time series residual prediction model. Based on the data structure characteristics of the hierarchical structure dimension, the residual sequence is decomposed into multiple hierarchical components. After aligning each hierarchical component with the date attribute information of each historical period, a time series prediction model is trained for each hierarchical component to obtain the time series prediction model corresponding to each hierarchical component after training. Based on the data structure characteristics of the category structure dimension, the residual sequence is decomposed into multiple category components. After aligning each category component with the date attribute information of each historical period, a time series prediction model is trained for each category component to obtain the time series prediction model corresponding to each category component after training. Input the date attribute information of the preset prediction period into the trained time-series residual prediction model to obtain the first residual prediction value; The date attribute information of the preset prediction period is input into the time series prediction model corresponding to each level component after training, and the prediction results are merged to obtain the second residual prediction value. The date attribute information of the preset prediction period is input into the time series prediction model corresponding to each category component after training, and the prediction results are merged to obtain the third residual prediction value.

6. The method for compensating for deviations in time-series prediction data as described in claim 1, characterized in that, The step of generating the initial predicted value of the cumulative numerical index for the preset prediction period based on the basic prediction parameters and the date attribute information of the preset prediction period specifically includes: Based on the date attribute information of the preset prediction period, determine the third day that belongs to the weekday period and the fourth day that belongs to the holiday period within the preset prediction period; Multiplying the third day number by the first historical baseline value yields the estimate of the third contribution value corresponding to the weekday period; Multiplying the fourth day number by the second historical baseline value yields the estimate of the fourth contribution value corresponding to the holiday period; If the cross-period effect correction parameter is not zero, calculate the average value of the cross-period effect correction parameter for multiple historical periods; The third contribution value estimate, the fourth contribution value estimate, and the average value are summed to generate the initial predicted value of the cumulative numerical index for the preset prediction period.

7. The method for compensating for deviations in time-series prediction data as described in claim 1, characterized in that, The step of using a Bayesian fusion algorithm to compensate for the deviation of the initial prediction value with multiple residual prediction values ​​as deviation compensation amounts, and generating the target prediction value of the cumulative numerical index within a preset prediction period specifically includes: Obtain the scene identifier information of the preset prediction time period; Based on the scene identification information, determine the prior weight distribution parameters corresponding to each residual prediction value; Based on the initial predicted value, the first residual predicted value, the second residual predicted value, the third residual predicted value, and the prior weight distribution parameters, the posterior distribution is sampled using the Markov chain Monte Carlo algorithm to determine the weight coefficients corresponding to the initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value, respectively. According to each weight coefficient, the initial predicted value, the first residual predicted value, the second residual predicted value, and the third residual predicted value are weighted and summed, and the weighted sum is used as the target predicted value of the cumulative numerical index during the prediction period.

8. A deviation compensation device for time series prediction data, characterized in that, include: The first acquisition module is used to acquire the preset forecast period and the cumulative numerical indicators to be predicted. The second acquisition module is used to acquire multi-source time series data of multiple historical periods that are concurrent with the preset prediction period. The multi-source time series data includes the actual value of the cumulative numerical indicator in each historical period, the date attribute information of each historical period, and the contribution value of each detailed data corresponding to the cumulative numerical indicator in each historical period. The determination module is used to determine basic prediction parameters based on the multi-source time series data, wherein the basic prediction parameters include at least one of the following: a first historical benchmark value corresponding to the weekday period, a second historical benchmark value corresponding to the holiday period, and a cross-period effect correction parameter; The first generation module is used to perform back-forward prediction for each historical period based on the basic prediction parameters and the date attribute information of each historical period, and generate the back-forward prediction value of the cumulative numerical index in each historical period. The second generation module is used to generate a residual sequence based on the difference between the actual value and the backtracked predicted value for each historical period. The third generation module is used to perform residual prediction on the residual sequence based on the multi-source time series data and the date attribute information of the preset prediction period according to multiple preset dimensions to obtain multiple residual prediction values. The multiple preset dimensions include time series dimension, hierarchical structure dimension and category structure dimension. The fourth generation module is used to generate the initial predicted value of the cumulative numerical index in the preset prediction period based on the basic prediction parameters and the date attribute information of the preset prediction period. The fifth generation module is used to compensate for the deviation of the initial prediction value by using multiple residual prediction values ​​as deviation compensation amounts through a Bayesian fusion algorithm, and to generate the target prediction value of the cumulative numerical index for the preset prediction period.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the deviation compensation method for time series prediction data as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the deviation compensation method for time series prediction data as described in any one of claims 1 to 7.

Images