Intermittent and blocky demand sales prediction method based on multi-dimensional aggregation

Abstract

The invention belongs to the technical field of sales volume prediction, and particularly relates to an intermittent and blocky demand sales volume prediction method based on multi-dimensional aggregation, which comprises the following steps: (1) data preparation: the data comprises the real number of stores, the number of products and corresponding daily sales data in a preset observation period; and (2) respectively carrying out aggregation processing on the data based on three dimensions of products, stores and time, fusing an exponential smoothing model to generate a plurality of groups of predicted values, constructing adjustment coefficient correction and demand mode adaptation correction through combination of the prediction values, and finally outputting an SKU-level accurate prediction result. The method can effectively break through the data sparse limitation, solves the problems of intermittency of multiple stores and multiple products in a retail scene and low precision of SKU sales prediction, and has a good market application prospect.

Description

Technical Field

[0001] This invention belongs to the field of sales forecasting technology, specifically relating to a method for forecasting intermittent and block demand based on multi-dimensional aggregation. Background Technology

[0002] In the retail industry, in sales scenarios involving multiple stores and multiple products, a large number of SKUs exhibit intermittent or block demand characteristics. Intermittent demand is characterized by zero sales most of the time and sales records only on rare occasions, while block demand combines the characteristics of large sales fluctuations with long intervals between zero demand.

[0003] In recent decades, intermittent demand forecasting has become one of the core research directions in the retail industry. However, existing sales forecasting methods, such as simple exponential smoothing or exponentially weighted moving averages, are mostly designed based on continuous and stable sales data, making it difficult to address the data sparsity problem caused by intermittent and block demand. This results in limited forecasting capabilities and poor reference effectiveness. Furthermore, current research in the retail industry has not fully emphasized the application value of time aggregation and contemporaneous aggregation of time series. Therefore, this application provides a multi-dimensional aggregation-based sales forecasting method for intermittent and block demand to overcome the limitations of existing forecasting techniques for such specific demand goods and solve the problem of low accuracy in forecasting sales of intermittent and block demand SKUs across multiple stores and products in retail scenarios.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting sales of intermittent and bulk demand based on multi-dimensional aggregation, so as to solve the problem of low accuracy in predicting sales of SKUs with intermittent and bulk demand in retail scenarios involving multiple stores and multiple products.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for forecasting intermittent and block demand based on multi-dimensional aggregation includes the following steps: S1. Data Preparation Phase: Acquiring data in retail scenarios Each store Use the daily sales data of the products to construct the original time-series dataset; S2, Product Dimension Aggregation Forecast: Aggregate SKU sales data from all stores by product dimension, use the exponential smoothing (ES) model to predict aggregated demand by product dimension, calculate weights based on store sales ratio, allocate aggregated forecast values ​​through weights, and obtain the initial forecast value for each SKU. S3, Store-level Aggregated Prediction: A sequence is constructed based on the daily total sales of each store, and an exponential smoothing (ES) model is used to predict the store's future sales. Total sales for the day; based on the past sales of each SKU in the corresponding stores. The total sales volume is used to calculate the weight, and the total forecast value is allocated through the weight to obtain the secondary forecast value of each SKU; S4. Time-Dimensional Aggregation Forecasting: Daily sales of SKUs are aggregated into weekly sales, and future sales are predicted using an exponential smoothing (ES) model. ( Weekly sales; S5, Time Dimension Constraint Correction: The weekly sales forecast value generated by S4 is used to correct the average forecast value generated by S2 and S3 to avoid extreme overestimation of sales forecast value. S6. Demand Pattern Adaptation and Correction: Identify the demand pattern of store SKU sequences. If it is intermittent or lumpy, and all store SKU data still maintains these two patterns after product-dimensional aggregation and prediction, then based on past... ( The forecast is revised using historical sales data from the weekday dimension. Specifically, if a certain day of the week (such as Monday, Tuesday, etc.) was historically... If the daily sales for the week are all 0, then the forecast for that date will be adjusted to 0.

[0007] Preferably, the observation period of the original time series dataset constructed in S1 is [missing information]. Heaven; among them, the former The data was used as the training set, and then... Tian data was used as the test set; each product corresponded to multiple SKUs, forming a total. SKU-level time series.

[0008] As a preferred approach, product dimension aggregation prediction is performed in S2, specifically including the following steps: S21. For each product At any time For all The daily sales data of this product in each store are aggregated concurrently to obtain the product... In time The aggregation requirement is calculated using the following formula: ; in, Indicates product The Each SKU in time The actual daily sales volume; , , ; S22. Employing an exponential smoothing (ES) model, based on the products within the training set. Aggregated demand sequence Predicting the product's future ( The demand for time (using days as the unit of time) is denoted as Its prediction function has the following form: ; S23, Computing Products correspond Weight vector of each SKU The weight Based on the products in the training set The recent The total sales revenue of Tian Guan is calculated using the following formula: ; S24, Product The aggregated demand forecast is obtained through a weighted vector. Assigned to the corresponding SKU, and obtain the future of each SKU The preliminary predicted value for time is calculated using the following formula: .

[0009] As a preferred approach, S3 performs store-level aggregated prediction, which specifically includes the following steps: S31, Targeting Stores At any time All of the stores The daily total sales data of these products are aggregated concurrently to obtain the store's data. In time The aggregation requirement is calculated using the following formula: ; in, Indicates store All products In time The actual daily sales volume; S32. Employ the exponential smoothing (ES) model, based on the stores in the training set. Aggregated demand sequence Predicting the store's future ( The demand for time (using days as the unit of time) is denoted as Its prediction function has the following form: ; S33, Calculate Stores correspond Weight vector of each SKU The weight Based on training centralized stores recent The total sales revenue of Tian Guan is calculated using the following formula: ; S34, Store The aggregated demand forecast is obtained through a weighted vector. Assigned to the corresponding SKU, and obtain the future of each SKU The predicted value at time t is calculated using the following formula: .

[0010] As a preferred approach, time-dimensional aggregation prediction is performed in S4, specifically including the following steps: S41. Aggregate the daily sales of each SKU along the time dimension to generate weekly sales, forming a new dataset. The calculation formula is as follows: ; in, This represents the weekly sales data for each SKU. S42. An exponential smoothing (ES) model is used, based on the historical observation sequence within the training set. Predicting the future Weekly sales at any given time, denoted as ;in, , The time unit is the week.

[0011] As a preferred option, S5 performs time dimension constraint correction, specifically including the following steps: S51, Predicted values ​​obtained from S2 and S3 and The formula for calculating the arithmetic mean and generating the predicted mean is as follows: ; S52. Construct the adjustment coefficient, the formula is as follows: ; S53. Based on the adjustment coefficients obtained in S52, the predicted values ​​obtained in S51 are corrected to obtain the values ​​of each SKU at time [time value missing]. The final predicted value is calculated using the following formula: .

[0012] As a preferred option, S6 performs demand pattern adaptation and modification, specifically including the following steps: S61. Demand Pattern Recognition: Calculating the Squared Coefficient of Variation The formula is as follows: ; in, The variance of the demand data, This represents the average of the demand data; The formula for calculating the Average Demand Interval (ADI) is as follows: ; in, For quantities with zero demand, The number of intervals; Demand patterns are categorized as follows: Smooth: satisfies CV 2 <0.49 and ADI<1.32; Intermittent: satisfies CV 2 <0.49 and ADI≥1.32; Erratic: satisfies CV 2 ≥0.49 and ADI<1.32; Lumpy type: satisfies CV 2 ≥0.49 and ADI≥1.32; S62. Retrieve Product At the store recent ( Historical sales data for the past 10 days, broken down by week. ( For example, Monday is Summing is performed, denoted as The calculation formula is as follows: ; S63, If the product At the store If the demand pattern is either intermittent or lumpy, and the product dimension aggregation still maintains these two patterns, then the sales forecast value is revised to obtain the final forecast value. The calculation formula is as follows: ; If the daily sales data for the corresponding day of the week (such as Monday, Tuesday, etc.) are all 0, then the predicted value for that prediction day (corresponding to the same day of the week) is determined to be 0.

[0013] Compared with the prior art, the present invention has the following beneficial effects: (1) The intermittent and block demand sales forecasting method based on multi-dimensional aggregation of the present invention aggregates scattered SKU-level sparse data into a more regular aggregated demand sequence through the aggregation processing of product, store and time three dimensions, providing effective input for the forecasting model and greatly expanding the applicability of the forecasting method.

[0014] (2) The intermittent and block demand sales forecasting method based on multi-dimensional aggregation of the present invention generates forecast values ​​by product and store dimensions respectively, and then makes a correction by combining the weekly forecast results of time dimension to avoid extreme overestimation of sales; subsequently, a second correction is made based on demand pattern recognition to accurately match the characteristics of intermittent and block demand, and the double guarantee makes the SKU-level forecast results more in line with the actual sales situation.

[0015] (3) The intermittent and block demand sales forecasting method based on multi-dimensional aggregation of the present invention focuses the weight calculation on the sales amount observed in the training set in the last 28 days, which can reflect the recent sales proportion trend, and make the aggregated forecast value tilted towards SKUs with more stable sales performance. The allocation result is more in line with the actual sales pattern of the retail scenario.

[0016] (4) The intermittent and bulk demand sales forecasting method based on multi-dimensional aggregation of the present invention outputs accurate SKU-level forecast results, which can directly provide data support for retail enterprises' inventory management and replenishment decisions, help enterprises optimize inventory structure, reduce capital occupation and stockout losses, and improve operational efficiency. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 These are characteristics of sales data from a large chain convenience store in Beijing; Figure 3 It represents the predictive performance of each model under the rolling prediction test. Detailed Implementation

[0018] The technical solution of this invention patent will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0019] See attached document Figure 1 This invention provides a method for forecasting intermittent and block demand sales based on multi-dimensional aggregation, comprising the following steps: S1. Data Preparation Phase: Acquiring data in retail scenarios Each store The daily sales data of the products were used to construct the original time series dataset; the observation period of the original time series dataset was... Heaven; among them, the former The data was used as the training set, and then... Tian data was used as the test set; each product corresponded to multiple SKUs, forming a total. SKU-level time series S2. Product-Dimensional Aggregated Forecasting: Aggregate SKU sales data from all stores by product dimension, use the exponential smoothing (ES) model to predict aggregated demand for each product dimension, then calculate weights based on store sales share, and allocate aggregated forecast values ​​according to these weights to obtain initial forecast values ​​for each SKU; the specific steps are as follows: S21. For each product ( ), at any time ( ), for all The daily sales data of this product in each store are aggregated concurrently to obtain the product... In time The aggregation requirement is calculated using the following formula: ; in, Indicates product The ( ) SKUs in time The actual daily sales volume; S22. Employing an exponential smoothing (ES) model, based on the products within the training set. Aggregated demand sequence Predicting the product's future ( The demand for time (using days as the unit of time) is denoted as Its prediction function has the following form: ; S23, Computing Products correspond Weight vector of each SKU The weight Based on the products in the training set The recent ( The total sales revenue is calculated based on the observations of the day, using the following formula: ; S24, Product The aggregated demand forecast is obtained through a weighted vector. Assigned to the corresponding SKU, and obtain the future of each SKU The preliminary predicted value for time is calculated using the following formula: ; S3, Store-level Aggregated Prediction: A sequence is constructed based on the daily total sales of each store, and an exponential smoothing (ES) model is used to predict the store's future sales. Total sales for the day; based on the past sales of each SKU in the corresponding stores. ( The total sales volume of each day is weighted according to its percentage, and the total forecast value is allocated based on the weights to obtain the secondary forecast value for each SKU. The specific steps are as follows: S31, Targeting Stores ( ), at any time ( ), for all of the stores The daily total sales data of these products are aggregated concurrently to obtain the store's data. In time The aggregation requirement is calculated using the following formula: ; in, Indicates store All products In time The actual daily sales volume; S32. Employ the exponential smoothing (ES) model, based on the stores in the training set. Aggregated demand sequence Predicting the store's future ( The demand for time (using days as the unit of time) is denoted as Its prediction function has the following form: ; S33, Calculate Stores correspond Weight vector of each SKU The weight Based on training centralized stores recent ( The total sales revenue observed over the days is calculated using the following formula: ; S34, Store The aggregated demand forecast is obtained through a weighted vector. Assigned to the corresponding SKU, and obtain the future of each SKU The predicted value at time t is calculated using the following formula: ; S4. Time-Dimensional Aggregation Forecasting: Daily sales of SKUs are aggregated into weekly sales, and future sales are predicted using an exponential smoothing (ES) model. ( Weekly sales figures; specific steps are as follows: S41. Aggregate the daily sales of each SKU along the time dimension to generate weekly sales, forming a new dataset. The calculation formula is as follows: ; in, This represents the weekly sales data for each SKU. S42. An exponential smoothing (ES) model is used, based on the historical observation sequence within the training set. Predicting the future ( , Weekly sales at any given time (using a week as the unit of time) are denoted as ,in ; S5. Time Constraint Correction: The average forecast value generated in S2 and S3 is corrected using the weekly sales forecast value generated in S4 to avoid extreme overestimation of sales forecast values; the specific steps are as follows: S51, Predicted values ​​obtained from S2 and S3 and The formula for calculating the arithmetic mean and generating the predicted mean is as follows: ; S52. Construct the adjustment coefficient, the formula is as follows: ; S53. Based on the adjustment coefficients obtained in S52, the predicted values ​​obtained in S51 are corrected to obtain the values ​​of each SKU at time [time value missing]. The final predicted value is calculated using the following formula: ; S6. Demand Pattern Adaptation and Correction: Identify the demand pattern of store SKU sequences. If it is intermittent or lumpy, and all store SKU data still maintains these two patterns after product-dimensional aggregation and prediction, then based on past... ( The forecast is revised using historical sales data from the weekday dimension. Specifically, if a certain day of the week (such as Monday, Tuesday, etc.) was historically... If the daily sales for the week are all 0, then the forecast for that date will be adjusted to 0.

[0020] The specific steps for adapting and modifying the demand pattern are as follows: S61. Demand Pattern Recognition: Calculating the Squared Coefficient of Variation The formula is as follows: ; in, The variance of the demand data, This represents the average of the demand data; The formula for calculating the Average Demand Interval (ADI) is as follows: ; in, For quantities with zero demand, The number of intervals; Demand patterns are categorized as follows: Smooth: satisfies CV 2 <0.49 and ADI<1.32; Intermittent: satisfies CV 2 <0.49 and ADI≥1.32; Erratic: satisfies CV 2 ≥0.49 and ADI<1.32; Lumpy type: satisfies CV 2 ≥0.49 and ADI≥1.32; S62. Retrieve Product At the store recent ( Historical sales data for the past 10 days, broken down by week. ( For example, Monday is Summing is performed, denoted as The calculation formula is as follows: ; S63, If the product At the store If the demand pattern is either intermittent or lumpy, and the product dimension aggregation still maintains these two patterns, then the sales forecast value is revised to obtain the final forecast value. The calculation formula is as follows: ; If the daily sales data for the corresponding day of the week (such as Monday, Tuesday, etc.) are all 0, then the predicted value for that prediction day (corresponding to the same day of the week) is determined to be 0.

[0021] To verify the feasibility of the method of the present invention, sales data from a large chain convenience store in Beijing from August 29, 2022 to June 19, 2023 were selected as data samples, and an original time series dataset was constructed according to the method of the present invention. The observation period of the original time series dataset is 295 days, of which the data of the first 267 days is used as the training set and the data of the last 28 days is used as the test set. It includes 152 stores and 173 kinds of low-temperature dairy beverages, forming a total of 14,242 time series.

[0022] Using the method described in S61 of this invention, the original data is simultaneously aggregated to identify the demand patterns of the data samples. Figure 2 The data features of a large chain convenience store in Beijing show that, after being aggregated by week (7 days), there are 5 sequences that show non-zero sales only once.

[0023] As attached Figure 2 As shown in the third column, after simultaneous aggregation, the proportions of discontinuous demand and block demand in the original time series decreased to 28.32% and 12.72%, respectively, indicating that the stability of the aggregated time series was relatively improved; Appendix Figure 2 The fourth column shows that the proportion of intermittent demand also decreased after time aggregation, further proving the stabilizing effect of aggregation on demand patterns.

[0024] Three sets of prediction experiments were conducted on the data samples using the seasonal naive model (SNaïve), the Croston method, the exponential smoothing method (ES), the exponential smoothing contemporaneous aggregation model (ES-C), and the method described in this invention. The mean absolute error (MAE) and root mean square error (RMSE) were used for comprehensive evaluation. Notably, the exponential smoothing contemporaneous aggregation model (ES-C) only employs a contemporaneous aggregation strategy, while the method described in this invention integrates contemporaneous aggregation and time aggregation strategies. Model performance was evaluated under different prediction periods (including one or more rounds). (Appendix) Figure 3 This section presents the prediction performance under the rolling prediction test, showcasing the prediction results for the test dataset. The number of observations is calculated as: number of items × prediction period × number of rolling rounds. The first three columns represent the average values.

[0025] See attached document Figure 3 Overall, the multidimensional aggregation correction model described in this invention performs optimally, outperforming other benchmark models: when the forecast period is 7 days, compared with the exponential smoothing (ES) method, the multidimensional aggregation correction model improves the MAE by 9.77% and the RMSE by 5.44%; when the forecast period is extended to 14 days and 28 days, the multidimensional aggregation correction model still maintains excellent performance, outperforming the exponential smoothing (ES) method and other benchmark models. This robust performance verifies the effectiveness of the multidimensional aggregation correction model in forecasting discontinuous and blocky demand, proving that it can handle complex demand patterns.

[0026] The exponential smoothing contemporaneous aggregation model (ES-C) outperforms the exponential smoothing method (ES), but falls short of the multidimensional aggregation correction model, indicating that combining contemporaneous aggregation and time aggregation strategies can further improve the predictive performance of the exponential smoothing model. Therefore, the multidimensional aggregation correction model described in this invention performs optimally, outperforming the exponential smoothing contemporaneous aggregation model (ES-C), the exponential smoothing method (ES), and other benchmark models. This demonstrates that combining contemporaneous aggregation and time aggregation on time series data not only simplifies the prediction process but also yields more accurate prediction results. The method of this invention is simple and robust. Although it lacks complex statistical characteristics, it is highly practical, providing reasonably accurate prediction results. It effectively overcomes data sparsity limitations, improves prediction accuracy, and is suitable for scenarios such as inventory management and replenishment decisions in retail enterprises.

[0027] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for forecasting intermittent and block demand sales based on multi-dimensional aggregation, characterized in that, Includes the following steps: S1. Data Preparation Phase: Acquiring data in retail scenarios Each store Use the daily sales data of the products to construct the original time-series dataset; S2, Product Dimension Aggregation Forecast: Aggregate SKU sales data from all stores by product dimension, use a smoothing exponential model to predict the aggregated demand for the product dimension, calculate the weights based on the store sales ratio, and allocate the aggregated forecast values ​​through weight allocation to obtain the initial forecast values ​​for each SKU. S3, Store-level Aggregated Prediction: A sequence is constructed based on the daily total sales of each store, and an exponential smoothing model is used to predict the store's future sales. Total sales for the day; based on the past sales of each SKU in the corresponding stores. The total sales volume of each day is used to calculate the weight, and the total forecast value is allocated through the weight to obtain the secondary forecast value of each SKU. S4. Time-Dimensional Aggregation Forecasting: Daily sales of SKUs are aggregated into weekly sales, and future sales are predicted using an exponential smoothing model. Weekly sales; S5, Time Dimension Constraint Correction: The weekly sales forecast value generated by S4 is used to correct the average forecast value generated by S2 and S3 to avoid extreme overestimation of sales forecast value. S6. Demand Pattern Adaptation and Correction: Identify the demand patterns of store SKU sequences. If the demand patterns are intermittent or block-like, and all store SKU data still maintain these two patterns after product-dimensional aggregation and prediction, then based on past... Historical sales data from the week's perspective are used to adjust the forecast.

2. The method for forecasting intermittent and block demand sales based on multi-dimensional aggregation according to claim 1, characterized in that, The observation period of the original time series dataset constructed in S1 is... Heaven; among them, the former The data was used as the training set, and then... Tian data was used as the test set; each product corresponded to multiple SKUs, forming a total. SKU-level time series.

3. The method for forecasting intermittent and block demand sales based on multi-dimensional aggregation according to claim 1, characterized in that, In S2, product-dimensional aggregation prediction is performed, specifically including the following steps: S21. For each product At any time For all The daily sales data of this product in each store are aggregated concurrently to obtain the product... In time The aggregation requirement is calculated using the following formula: ； in, Indicates product The Each SKU in time The actual daily sales volume; , , ; S22. Employ an exponential smoothing model based on products within the training set. Aggregated demand sequence Predicting the product's future The need for time, denoted as Its prediction function has the following form: ； S23, Computing Products correspond Weight vector of each SKU The weight Based on the products in the training set The recent The total sales revenue of Tian Guan is calculated using the following formula: ； S24, Product The aggregated demand forecast is obtained through a weighted vector. Assigned to the corresponding SKU, and obtain the future of each SKU The preliminary predicted value for time is calculated using the following formula: 。 4. The method for forecasting intermittent and block demand sales based on multi-dimensional aggregation according to claim 1, characterized in that, The store-level aggregation prediction in S3 includes the following steps: S31, Targeting Stores At any time All of the stores The daily total sales data of these products are aggregated concurrently to obtain the store's data. In time The aggregation requirement is calculated using the following formula: ； in, Indicates store All products In time The actual daily sales volume; S32. Employ an exponential smoothing model based on stores within the training set. Aggregated demand sequence Predicting the store's future The need for time, denoted as Its prediction function has the following form: ； S33, Calculate Stores correspond Weight vector of each SKU The weight Based on training centralized stores recent The total sales revenue of Tian Guan is calculated using the following formula: ； S34, Store The aggregated demand forecast is obtained through a weighted vector. Assigned to the corresponding SKU, and obtain the future of each SKU The predicted value at time t is calculated using the following formula: 。 5. The method for forecasting intermittent and block demand sales based on multi-dimensional aggregation according to claim 1, characterized in that, The time-dimensional aggregation prediction in S4 includes the following steps: S41. Aggregate the daily sales of each SKU along the time dimension to generate weekly sales, forming a new dataset. The calculation formula is as follows: ； in, This represents the weekly sales data for each SKU. S42. An exponential smoothing model is adopted, based on the historical observation sequence within the training set. Predicting the future Weekly sales at any given time, denoted as ;in, , The time unit is the week.

6. The method for forecasting intermittent and block demand sales based on multi-dimensional aggregation according to claim 1, characterized in that, S5 performs time dimension constraint correction, specifically including the following steps: S51, Predicted values ​​obtained from S2 and S3 and The formula for calculating the arithmetic mean and generating the predicted mean is as follows: ； S52. Construct the adjustment coefficient, the formula is as follows: ； S53. Based on the adjustment coefficients obtained in S52, the predicted values ​​obtained in S51 are corrected to obtain the values ​​of each SKU at time [time value missing]. The final predicted value is calculated using the following formula: 。 7. The method for forecasting intermittent and block demand sales based on multi-dimensional aggregation according to claim 1, characterized in that, S6 performs requirement pattern adaptation and correction, specifically including the following steps: S61. Demand Pattern Recognition: Calculating the Squared Coefficient of Variation The formula is as follows: ； in, The variance of the demand data, This represents the average of the demand data; The formula for calculating the Average Demand Interval (ADI) is as follows: ； in, For quantities with zero demand, The number of intervals; Demand patterns are categorized as follows: Stable type: satisfies CV 2 <0.49 and ADI<1.32; Intermittent type: satisfies CV 2 <0.49 and ADI≥1.32; Fluctuating type: satisfies CV 2 ≥0.49 and ADI<1.32; Block type: satisfies CV 2 ≥0.49 and ADI≥1.32; S62. Retrieve Product At the store recent Historical sales data for the day, broken down by week. Summation is performed, denoted as , The calculation formula is as follows: ； S63, If the product At the store If the demand pattern is intermittent or blocky, and the product dimension aggregation still maintains these two patterns, then the sales forecast value is revised to obtain the final forecast value. The calculation formula is as follows: ； If the daily sales data for the corresponding weekday dimension are all 0, then the predicted value for that prediction day is determined to be 0.

Images