Deep learning-based sports market demand prediction method and device, and medium

Abstract

The invention discloses a sports market demand prediction method and device based on deep learning and a medium, and relates to the technical field of market demand prediction, and the method comprises the steps: collecting sports demand data, and carrying out the preprocessing; performing relation mining and graph structure learning on the preprocessed sports demand data through a graph attention space-time network to generate a macroscopic demand potential energy graph; performing potential area identification on the macroscopic demand potential energy diagram by adopting a pre-trained sports market simulation model, and outputting local demand prediction data; performing weighted fusion and error correction on the macroscopic demand potential energy map and the local demand prediction data, and outputting a sports demand prediction score; and making a sports market demand strategy according to the sports demand prediction score and the multi-granularity demand prediction report, and transmitting the sports market demand strategy to an enterprise manager through an enterprise decision support interface. According to the method, multi-level accurate prediction and decision support of sports market demands are realized through dual-mechanism cooperation of the graph attention space-time network and the space-time convolution.

Description

Technical Field

[0001] This invention relates to the field of market demand forecasting technology, and in particular to a method, device and medium for forecasting sports market demand based on deep learning. Background Technology

[0002] In recent years, sports market demand forecasting technology has rapidly evolved from traditional statistical analysis to intelligent decision-making based on multi-source data fusion. Key advancements include: extrapolating trends from historical sales data using time series models (such as ARIMA and LSTM); integrating macroeconomic indicators and consumer attribute data using ensemble learning algorithms like random forests; analyzing potential relationships between product categories using association rule mining techniques; the initial application of graph neural networks in building brand-consumer interaction networks to identify potential customer groups; deep learning frameworks weighting multi-dimensional features through attention mechanisms; and cloud computing platforms providing computational support for the parallel processing of massive market data. This has resulted in a data-driven market forecasting methodology.

[0003] However, existing methods have shortcomings in extracting spatiotemporal dynamic features and providing multi-granularity decision support. Traditional models struggle to effectively characterize the spatiotemporal transmission effects of dynamic factors such as sporting events and regional policies on market demand, and lack the ability to explicitly model complex relationships between market entities. This makes it difficult for macro-level forecasts to guide specific operational decisions. Furthermore, existing solutions generally lack reliable quantification mechanisms for forecast results and do not establish simulation verification links linked to marketing strategies, thus limiting their application value in real-world business scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for predicting sports market demand to address the shortcomings in spatiotemporal dynamic feature extraction and multi-granularity decision support.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, this invention provides a deep learning-based method for predicting sports market demand, comprising: collecting sports demand data and preprocessing it; the sports demand data includes marketing strategy data, historical sales data, market trend data, and policy environment data; performing relationship mining and graph structure learning on the preprocessed sports demand data through a graph attention spatiotemporal network to output a dynamic knowledge graph, and using spatiotemporal convolution to extract spatiotemporal patterns from the dynamic knowledge graph to generate a macro-demand potential energy map; using a pre-trained sports market simulation model to identify potential areas on the macro-demand potential energy map, and loading marketing strategy data to simulate the consumer decision-making process to output local demand prediction data; performing weighted fusion and error correction on the macro-demand potential energy map and local demand prediction data to form a multi-granularity demand prediction report, and performing confidence assessment on the multi-granularity demand prediction report to output a sports demand prediction score; formulating a sports market demand strategy based on the sports demand prediction score and the multi-granularity demand prediction report, and transmitting the sports market demand strategy to enterprise managers through an enterprise decision support interface.

[0007] As a preferred embodiment of the deep learning-based sports market demand prediction method of the present invention, the step of performing relationship mining and graph structure learning on the preprocessed sports demand data through a graph attention spatiotemporal network to output a dynamic knowledge graph includes the following specific steps: Different sports entities in the preprocessed sports demand data are defined as market entity nodes. Business relationship edges are established according to the business logic between sports entities, and integrated to form an initial structure diagram. The dynamic influence weights between market entity nodes are obtained by using a graph attention spatiotemporal network, and the dynamic influence weights are added to the initial structure graph to generate a weighted structure graph. Feature extraction is performed on the weighted structure graph in the spatiotemporal dimension, and the spatiotemporal feature matrix is ​​output. Based on the spatiotemporal feature matrix, the market entity nodes of the weighted structure graph are updated, and a dynamic knowledge graph is output.

[0008] As a preferred embodiment of the deep learning-based sports market demand prediction method of the present invention, the step of using spatiotemporal convolution to extract spatiotemporal patterns from a dynamic knowledge graph to generate a macroscopic demand potential graph includes the following specific steps: Perform spatiotemporal structuring transformation on dynamic knowledge graphs and output spatiotemporal feature tensors; Spatiotemporal feature tensors are subjected to pattern recognition and feature diffusion analysis through spatiotemporal convolution, and spatiotemporal feature maps are output. Regional feature aggregation and geographic grid mapping are performed on the spatiotemporal feature map to generate a macroscopic demand potential energy map.

[0009] As a preferred embodiment of the deep learning-based sports market demand forecasting method of the present invention, the specific steps of using a pre-trained sports market simulation model to identify potential regions in the macro demand potential map are as follows: Regional grid demand intensity analysis is performed on the macro demand potential energy map using a grid clustering algorithm to obtain regional demand intensity data. Based on regional demand intensity data, a pre-trained sports market simulation model is used to assess the regional potential of sports market demand and output a list of regional potential assessments. The list of regions with potential is sorted in reverse order and intensity is selected to output a list of potential regions.

[0010] As a preferred embodiment of the deep learning-based sports market demand forecasting method described in this invention, the specific steps of loading marketing strategy data to simulate the consumer decision-making process and outputting local demand forecast data are as follows: The marketing strategy data is linked and matched with the regional potential assessment list to form a strategy-region configuration set; Perform consumer decision simulation on the strategy-region configuration set, generate a simulated decision dataset, summarize the demand on the simulated decision dataset, and output the initial demand forecast results; The initial demand forecast results are aggregated and calibrated across regions to form local demand forecast data for each potential region.

[0011] As a preferred embodiment of the deep learning-based sports market demand forecasting method described in this invention, the specific steps for weighted fusion and error correction of the macro-level demand potential map and local demand forecasting data to form a multi-granularity demand forecasting report are as follows: Align and match the macro demand potential map with local demand forecast data to generate a demand fusion dataset; Based on a preset weighting strategy, the demand fusion dataset is weighted and statistically analyzed to generate sports market demand data. The sports market demand data is integrated and formatted into a multi-granularity demand forecast report.

[0012] As a preferred embodiment of the deep learning-based sports market demand forecasting method of the present invention, the specific steps for evaluating the confidence level of the multi-granularity demand forecasting report and outputting a sports demand forecasting score are as follows: The predicted values ​​in the multi-granularity demand forecast report are compared with historical sales data point by point to generate a residual sequence. Autocorrelation analysis and white noise test are performed on the residual sequence to generate a time series stability index; Based on market trend data and policy environment data, this study uses event correlation analysis to assess the impact of market trend data and policy environment data on sports market demand forecasts and generates event impact coefficients. The time series stability index and event impact coefficient are scored and fused to output a sports demand prediction score.

[0013] As a preferred embodiment of the deep learning-based sports market demand forecasting method described in this invention, the steps of formulating a sports market demand strategy based on sports demand forecasting scores and multi-granularity demand forecasting reports, and transmitting the sports market demand strategy to enterprise managers through an enterprise decision support interface, are as follows: The sports demand forecasting score and multi-granularity demand forecasting report are matched with a pre-set set of strategy templates to output a draft sports demand strategy. The draft sports demand strategy is validated for completeness and feasibility, and a sports market demand strategy is output. The sports market demand strategy is encapsulated into data messages through the enterprise decision support interface and transmitted to enterprise managers via the HTTPS protocol to complete the sports market demand forecast.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the deep learning-based sports market demand forecasting method described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the deep learning-based sports market demand forecasting method described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By synergistically employing both graph attention spatiotemporal networks and spatiotemporal convolution, multi-level accurate prediction and decision support for sports market demand are achieved. The graph attention spatiotemporal network performs relationship mining and graph structure learning on preprocessed sports demand data, dynamically capturing complex relationships between market entities and providing a structured foundation for spatiotemporal pattern analysis. Spatiotemporal convolution extracts spatiotemporal patterns from the dynamic knowledge graph, generating a macro-level demand potential graph to quantify global market demand trends. Combined with a pre-trained sports market simulation model, potential region identification and consumer decision simulation are performed to generate executable strategies, comprehensively improving the granularity, reliability, and practicality of the prediction results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0018] Figure 1 This is a flowchart of a deep learning-based method for predicting sports market demand. Figure 2 A flowchart for outputting a dynamic knowledge graph; Figure 3 To output a flowchart of local demand forecast data; Figure 4 This is a flowchart for a multi-granularity demand forecasting report. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Reference Figures 1-4 This is one embodiment of the present invention, which provides a deep learning-based method for predicting sports market demand, comprising the following steps: S1. Collect sports demand data and perform preprocessing.

[0023] Sports demand data includes marketing strategy data, historical sales data, market trend data, and policy environment data; Specifically, marketing strategy data is collected by connecting to the company's internal marketing activity management data. This includes structured and semi-structured data such as promotional activity plans, price adjustment records, advertising plans, channel promotion strategies, and membership benefits. The collection process extracts and stores the strategy parameters of the marketing strategy as standardized data entries at fixed time intervals. Historical sales data is obtained from the company's transaction database and retail terminal information, covering sports goods sales records from different regions, stores, and e-commerce platforms. This includes information such as product number, sales quantity, transaction price, transaction time, and geographical location. Data aggregation is completed through periodic batch export or real-time interface synchronization. Market trend data comes from publicly available industry research reports, third-party published sports consumption indices, statistics on trending topics on social media, and keyword popularity changes. Web crawlers combined with API calls are used to continuously crawl and collect trend information from multiple channels. Policy environment data is collected by accessing official websites at all levels and sports management-related bulletin boards, collecting textual materials such as regulations related to the sports industry, approval information for major events, and national fitness plans. Scheduled scanning and keyword matching are used to automate the acquisition and organization of policy information.

[0024] Preprocessing includes missing value imputation, data formatting, and timestamp alignment; Specifically, for missing or incomplete records in marketing strategy data, historical sales data, market trend data, and policy environment data, average value filling and nearest neighbor node filling methods are used to supplement the missing information points in the marketing strategy data, historical sales data, market trend data, and policy environment data, ensuring the integrity of each record; the time fields in the marketing strategy data, historical sales data, market trend data, and policy environment data are uniformly converted to the international standard date format, and non-daily granular time information in the marketing strategy data, historical sales data, market trend data, and policy environment data is expanded or aligned to specific dates according to rules to ensure the consistency of time expression; the marketing strategy data, historical sales data, market trend data, and policy environment data are also filled with missing or incomplete records. In sales data, market trend data, and policy environment data, various numerical fields such as price, sales volume, and index values ​​are standardized in precision, retaining two decimal places and removing unit symbols, and converted to a pure number format. After format conversion, timestamp alignment is performed, and a unified timeline is constructed with "day" as the smallest time unit. All data entries are mapped to the timeline according to their corresponding time points or time periods. For validity periods or activity cycles that span multiple days, continuous time records are generated by breaking them down by day to ensure that data from different sources are comparable and aligned in the time dimension. Through missing value imputation, data formatting, and timestamp alignment operations, a set of preprocessed sports demand data with a unified time base, standardized field formats, and consistent semantics is output. Furthermore, the unified timeline is constructed using the "day" as the smallest unit of time because daily granularity offers both temporal resolution and data operability in sports market demand analysis. It can capture short-term fluctuations in promotional activities and consumer behavior while avoiding redundancy and noise from hourly or minute-level data. The statistical period for historical sales data, the effective period of marketing strategies, and changes in market trends is based on the day. Daily granularity can naturally align with the original collection frequency of various data sources, reducing information distortion caused by interpolation and aggregation. Sports consumption behavior is significantly affected by weekends, holidays, and event calendars. Using the day as the unit can accurately correlate the periodic factors of weekends, holidays, and event calendars, supporting subsequent aggregation analysis on longer time scales such as weeks and months. Therefore, the day is chosen as the smallest unit of the unified timeline.

[0025] S2. Relationship mining and graph structure learning are performed on the preprocessed sports demand data through a graph attention spatiotemporal network to output a dynamic knowledge graph. Spatiotemporal convolution is then used to extract spatiotemporal patterns from the dynamic knowledge graph to generate a macro-demand potential energy graph.

[0026] Different sports entities in the preprocessed sports demand data are defined as market entity nodes. Business relationship edges are established according to the business logic between sports entities, and integrated to form an initial structure diagram. Specifically, sports entities with independent business significance are extracted from the preprocessed sports demand data, including athletes, sports brands, sports events, product categories, sales regions, and marketing activities. These sports entities are treated as market entity nodes. Business relationship edges are established according to the business logic between sports entities. The connection relationship is determined based on the actual association between sports entities. For example, if an athlete participates in a sports event, a "participation" type business relationship edge is established between the corresponding athlete market entity node and the sports event market entity node. If a sports brand sponsors a sports event, a "sponsorship" type business relationship edge is established between the corresponding sports brand market entity node and the sports event market entity node. If a product category generates sales records in a sales region, a "sales" type business relationship edge is established between the corresponding product category market entity node and the sales region market entity node. All market entity nodes and business relationship edges are stored and organized to form an initial structure diagram containing a set of nodes, a set of edges, and corresponding attributes.

[0027] The dynamic influence weights between market entity nodes are obtained by using a graph attention spatiotemporal network, and the dynamic influence weights are added to the initial structure graph to generate a weighted structure graph. Specifically, the initial structure graph is input into the graph attention spatiotemporal network. The input feature dimension of the graph attention spatiotemporal network is defined as a three-dimensional tensor composed of the multi-dimensional attributes (such as demand intensity, historical sales, and marketing intensity) carried by each market entity node at continuous time steps. The structure of the three-dimensional tensor is the number of market entity nodes × time step length × feature dimension of market entity nodes. The output feature dimension is defined as the three-dimensional structure of the dynamic influence weights generated by the graph attention spatiotemporal network. The three-dimensional structure is the number of market entity nodes × number of adjacent market entity nodes × output time step length, used to represent the mutual influence strength between market entity nodes at different times. The graph attention spatiotemporal network calculates the attention coefficient between market entity nodes by using the features of each market entity node... Vectors are converted into query vectors, key vectors, and value vectors. The dot product similarity between the query vectors and key vectors of the current market entity node and its neighboring market entity nodes is obtained. After scaling and softmax normalization, the attention coefficient is obtained. The importance of each market entity node to its neighboring nodes is dynamically learned and quantified, i.e., dynamic influence weight. The dynamic influence weight reflects the strength of the interaction between market entity nodes. For example, the influence of popular events on the host region is much greater than that of ordinary event nodes. The dynamic influence weight is added as a new attribute to the business relationship edges of the initial structure graph, thereby improving the initial structure graph, which only contains basic connection information, into a weighted structure graph that can accurately characterize the strength of the dynamic interaction between market entity nodes. Furthermore, the training process of the graph attention spatiotemporal network takes the initial structure graph as input and combines it with the time series information in the preprocessed sports demand data to construct dynamic graph training samples containing multiple time steps. Each dynamic graph training sample contains the market entity node status and business relationship edge connections within the corresponding time window. The size of the training samples is determined according to the time span of the preprocessed sports demand data, set to cover daily data from the past two years, totaling 730 time points. A time window of 14 days is then slidably extracted as a single dynamic graph training sample. The time window length is set to 14 to capture both cyclical consumption patterns and... This reflects the market response delay of short-term marketing activities; the batch size is set to 32 to balance training stability and computational efficiency, avoiding memory overflow under limited hardware resources; the learning rate is initialized to 0.001, using the Adam optimizer combined with a learning rate decay strategy, enabling rapid convergence in the early stages of training and fine-tuning parameters later to prevent oscillations; the learnable parameters of the graph attention spatiotemporal network are initialized, and the attention coefficients of each market entity node to its neighboring market entity nodes at different time steps are obtained through forward propagation. This involves linearly transforming the feature vectors of each market entity node at different time steps to generate query vectors and key vectors, and statistically analyzing the current market... The dot product similarity between the query vector of a market entity node and the key vectors of its neighboring market entity nodes is scaled and normalized using softmax to obtain the attention coefficients of each neighboring market entity node. These attention coefficients are then normalized to dynamic influence weights using the softmax function. The dynamic influence weights are used to weighted aggregate the features of neighboring market entity nodes, generating an updated representation of the current market entity node. This updated representation is then passed layer by layer to extract higher-order association features. Known market response results from the preprocessed sports demand data, such as historical sales growth or changes in consumer attention, are used as monitoring signals. A mean squared error loss function is constructed and passed layer by layer using the backpropagation algorithm. The learnable parameters in the graph attention spatiotemporal network are updated by comparing the deviation gradient between the output of the mean squared error loss function and the actual market response, outputting a gradient signal, and propagating the gradient signal forward layer by layer. The optimizer is then used to adjust the learnable parameters of the graph attention spatiotemporal network using gradient descent. This forward propagation, weighted aggregation, and learnable parameter update are repeated until the mean squared error loss function converges (when the mean squared error loss function no longer decreases in consecutive training cycles and the prediction results of the graph attention spatiotemporal network on the validation data are stable, the mean squared error loss function is considered to have converged), resulting in the trained graph attention spatiotemporal network.

[0028] Feature extraction is performed on the weighted structure graph in the spatiotemporal dimension, and the spatiotemporal feature matrix is ​​output. Specifically, at each time step, a feature aggregation operation is performed on each market entity node in the weighted structure graph. By aggregating the features of each market entity node itself and the features of its neighboring nodes, and weighting and summing the features of the neighboring nodes according to the dynamic influence weight, the spatial dependency relationship between market entity nodes is captured. The graph attention spatiotemporal network connects the node representations of different time steps along the time dimension, captures the trend and pattern of market entity nodes changing over time, and outputs a feature matrix that integrates spatiotemporal information, i.e., the spatiotemporal feature matrix.

[0029] Based on the spatiotemporal feature matrix, the market entity nodes of the weighted structure graph are updated, and a dynamic knowledge graph is output. Specifically, the spatiotemporal feature matrix contains new feature representations of each market entity node after learning spatiotemporal patterns. The new feature representations are used to update or replace the original feature vectors of the corresponding market entity nodes in the weighted structure graph. The update process makes the representation of each market entity node not only include static attributes, but also deeply integrate dynamic behavior patterns and the interactive influence of neighboring nodes, outputting a graph structure that reflects the complex spatiotemporal relationships between sports market entities, i.e., a dynamic knowledge graph.

[0030] Perform spatiotemporal structuring transformation on dynamic knowledge graphs and output spatiotemporal feature tensors; Specifically, the feature vector of each market entity node in the dynamic knowledge graph at each time step is bound to the geographic coordinates and timestamp. According to the geographic grid division requirements (the geographic grid division requirements refer to dividing the geographic space into a series of regularly arranged geographic grids of the same size according to the geographic range and spatial resolution of the research area), the market entity nodes with spatial locations are assigned to the corresponding geographic grids. The feature vectors of multiple market entity nodes in the same geographic grid are time-aligned and averaged. Time alignment is to align the feature vectors of each market entity node in the same geographic grid at the same time step to a unified time grid point. For missing time points, the average of the previous and next time steps is used to fill in the missing time points. In the same time step, the feature vectors of all market entity nodes in the geographic grid are averaged element by element to complete the average aggregation and form a spatiotemporal feature tensor. The spatiotemporal feature tensor has a three-dimensional structure with spatial dimension, time dimension and feature dimension. Spatiotemporal feature tensors are subjected to pattern recognition and feature diffusion analysis through spatiotemporal convolution, and spatiotemporal feature maps are output. Spatiotemporal convolution is performed on the spatiotemporal feature tensor. This involves sliding a 3D convolution kernel across the spatial and temporal dimensions to perform weighted summation and nonlinear transformation on the spatiotemporal feature tensor within each local spatiotemporal region, extracting local spatiotemporal patterns. Multiple convolutions (typically two to three layers) are then applied to expand the receptive field layer by layer to capture a wider range of spatiotemporal dependencies, while avoiding gradient vanishing and overfitting issues, ensuring effective feature diffusion. Too few convolutional layers result in insufficient receptive field, making it difficult to capture long-distance spatiotemporal dependencies and limiting feature extraction capabilities; too many layers increase complexity, causing gradient vanishing or exploding, leading to training difficulties. This process diffuses and enhances the spatiotemporal feature tensor in the spatiotemporal domain. The feature mapping results output by the spatiotemporal convolution are then organized according to a spatial grid and time steps to generate a spatiotemporal feature map containing enhanced spatiotemporal pattern information.

[0031] Regional feature aggregation and geographic grid mapping are performed on the spatiotemporal feature map to generate a macro-level demand potential energy map; Specifically, the demand characteristic values ​​within each geographic grid in the spatiotemporal feature map are weighted and summed over time to obtain the comprehensive demand intensity value for each geographic grid. The comprehensive demand intensity values ​​of all geographic grids are arranged into two-dimensional grid data according to geographic coordinates. The two-dimensional grid data is visualized and mapped into a geographic spatial distribution map through color gradients, forming a macro-demand potential energy map with geographic grids as units and color depth representing the expected intensity of demand for sporting goods in a specific future time period.

[0032] S3. A pre-trained sports market simulation model is used to identify potential areas in the macro demand potential map, and marketing strategy data is loaded to simulate the consumer decision-making process, outputting local demand forecast data.

[0033] Regional grid demand intensity analysis is performed on the macro demand potential energy map using a grid clustering algorithm to obtain regional demand intensity data. Specifically, the geographic space covered by the macro-demand potential map is divided into geographic grids according to a fixed resolution (set to 1 km × 1 km, which matches the spatial distribution granularity of city-level sports facilities and consumption hotspots, ensures a moderate number of geographic grids, and avoids the geographic grids from obscuring local demand differences due to being too large or introducing noise due to being too small). Each geographic grid corresponds to a demand intensity value. Using geographic grids as the basic unit of analysis, a grid clustering algorithm is used to cluster adjacent geographic grids based on the similarity of demand intensity (i.e., using the demand intensity value of each geographic grid as a feature, by statistically analyzing the degree of difference in demand intensity between adjacent geographic grids, geographic grids with the same difference and spatial connection are gradually merged). Geographic grids with similar demand intensity and spatial continuity are merged into regional clusters. The arithmetic mean of the demand intensity values ​​of all geographic grids within each regional cluster is calculated to obtain the average demand intensity of the regional cluster. The average demand intensity of all regional clusters is then summarized as regional demand intensity data to form a regional demand intensity dataset covering different geographic areas.

[0034] Based on regional demand intensity data, a pre-trained sports market simulation model is used to assess the regional potential of sports market demand and output a list of regional potential assessments. Specifically, the average demand intensity of each regional cluster in the regional demand intensity data is used as the input feature. This is combined with quantitative indicators from historical sales data, market trend data, and policy environment data corresponding to the sales region, and then input into a pre-trained sports market simulation model. The pre-trained sports market simulation model utilizes internally learned market response patterns to extrapolate the future demand growth trend and market expansion space of each sales region. Through nonlinear transformation and weighting of historical sales data, market trend data, and policy environment data for each sales region, and via multi-dimensional feature interaction and pattern matching, the model extrapolates the demand growth trend and market expansion space of each sales region under the future market environment, obtaining a comprehensive potential score for each sales region. This comprehensive potential score reflects the sports market demand development potential of the sales region under existing market conditions and external environment. All sales region names and their corresponding comprehensive potential scores are grouped into structured data entries, organized by sales region, and output as a regional potential assessment list containing multiple sales regions and their comprehensive potential scores. Furthermore, the pre-trained sports market simulation model is built on a Long Short-Term Memory (LSTM) network framework. The initial training epochs are set to 100 epochs. This initial setting ensures sufficient iterations to fully learn the complex mapping relationship between input features and demand growth rates, while also balancing training efficiency and convergence stability. An early stopping mechanism is used for dynamic termination. The training and validation sets are divided chronologically in an 80% to 20% ratio. This ratio ensures the validation set reflects the model's predictive ability for future time periods, avoiding data skipping. The early stopping mechanism stops training if the validation error does not decrease after 10 consecutive epochs. This 10-epoch setting ensures sufficient learning while avoiding ineffective iterations. Overfitting control is achieved by monitoring the difference between training and validation errors. When the validation error increases while the training error continues to decrease, overfitting is identified, training is immediately terminated, and the model is rolled back. The training process of the pre-trained sports market simulation model is based on historical time... The training sample set is constructed by using regional demand intensity data, historical sales data, market trend data, and policy environment data within a given period as inputs, and the actual demand growth rate for the corresponding future time period as the output label. By adjusting the internal parameters of the sports market simulation model, the model learns the mapping relationship between the input features (i.e., regional demand intensity data, historical sales data, market trend data, and policy environment data) and the actual demand growth rate. A loss function is used to measure the deviation between the potential prediction value (the potential prediction value comes from the comprehensive potential score output by the pre-trained sports market simulation model after forward calculation of regional demand intensity data, historical sales data, market trend data, and policy environment data) and the true value, and the parameters are optimized through backpropagation algorithm. The training process is repeated iteratively until the prediction error of the sports market simulation model converges (when the prediction error of the sports market simulation model on the validation set no longer decreases and tends to stabilize in multiple consecutive training rounds, the prediction error is considered to have converged), thus obtaining a sports market simulation model with regional potential assessment capabilities.

[0035] Sort the regional potential assessment list in reverse order and select the intensity, then output a list of potential regions. Specifically, all potential regions in the regional potential assessment list are arranged in reverse order from highest to lowest according to their corresponding comprehensive potential scores, forming an ordered sequence of regions. The ordered sequence of regions is then screened for potential using a fixed selection ratio (set to the top 20%, based on the Pareto principle (80 / 20 rule), which has been widely verified in practice as an effective ratio for balancing resource focus and coverage breadth, enabling the identification of the core regions with the greatest development potential while avoiding an overly broad candidate range that leads to resource dispersion). The selected potential regions and their corresponding comprehensive potential scores are then reorganized and output as a list of potential regions containing both the region name and its score. The marketing strategy data is linked and matched with the regional potential assessment list to form a strategy-region configuration set; Based on each marketing strategy record in the marketing strategy data, including promotional activity plans, price adjustment records, advertising placement plans, channel promotion strategies, and membership benefits, the applicable product categories, target audiences, and effective time range for each marketing strategy are obtained. Each potential region in the regional potential assessment list is traversed, and based on the market level corresponding to the comprehensive potential score of the potential region, combined with the response of the potential region to the current product category in historical sales data, the marketing strategy most likely to generate a positive demand response is matched. Each marketing strategy is paired with one or more suitable potential regions, forming a combination entry with "marketing strategy-region" as the basic unit. All generated combination entries are summarized to construct a structured data set containing marketing strategies, suitable regions, matching criteria, and expected effective time, i.e., the strategy-region configuration set.

[0036] Perform consumer decision simulation on the strategy-region configuration set, generate a simulated decision dataset, summarize the demand on the simulated decision dataset, and output the initial demand forecast results; Specifically, each "marketing strategy-region" combination in the strategy-region configuration set is used as input. Combined with historical sales data, market trend data, and policy environment data from potential regions, information on demographic characteristics, consumption levels, and behavioral preferences is used to construct a virtual consumer group. Parameters such as price adjustments, promotional efforts, and advertising exposure frequency from the marketing strategy data are applied to the virtual consumer group. A Long Short-Term Memory (LSTM) network is used to extrapolate the probability of each virtual consumer's response to the marketing strategy, recording purchase decision behavior. The decision results of all virtual consumers are statistically analyzed to generate the purchase quantity, purchase time, and product category distribution for each "marketing strategy-region" combination, forming a simulated decision dataset containing multi-dimensional behavioral records. The purchase quantity of each record in the simulated decision dataset is summed according to potential region and product category to obtain the total demand for each potential region under a specific marketing strategy, outputting the initial demand forecast results. Furthermore, the Long Short-Term Memory (LSTM) network constructs training samples based on time-series information from historical sales data, market trend data, and policy environment data. Each training sample contains a sequence of consumer behavior features within a continuous time window as input, and the corresponding actual purchase decision as a label. The LTM network structure employs an embedding layer to process discrete features. This embedding layer maps discrete features (such as region coding, marketing activity type, and consumption level—non-continuous categorical variables) to fixed-length dense vectors through a learnable embedding matrix (which is continuously updated and optimized during LTM training using the backpropagation algorithm based on the gradient of the loss function). This ensures that similar dense vectors have similar representations. This system effectively captures the potential correlations between historical sales data, market trend data, and policy environment data. Stacked long short-term memory (LSM) network layers are used to capture long-term dependencies in behavioral sequences, and fully connected layers output purchase probabilities. That is, by processing and capturing long-term dependencies in behavioral sequences step by step, the system outputs hidden states containing temporal context information and performs linear transformations. The results are then compressed to the 0-1 range using the Sigmoid activation function to output purchase probabilities. During training, the cross-entropy loss function is used to measure the difference between the purchase probability and the true label. The network parameters of the LSM network are iteratively updated using the backpropagation algorithm and the Adam optimizer. Early stopping is used to prevent overfitting, resulting in a trained LSM network that can accurately simulate consumer decision-making behavior.

[0037] The initial demand forecast results are aggregated and calibrated across regions to form local demand forecast data for each potential region. Specifically, the demand forecasts from different "marketing strategy-region" combinations belonging to the same potential region in the initial demand forecast results are accumulated to achieve cross-regional aggregation, obtaining the total demand forecast for the potential region under the influence of multiple marketing strategies. Combining this with historical sales data corresponding to the potential region, the deviation ratio between the current forecast value and the actual sales volume of the same period in history is obtained. This deviation ratio is used to linearly calibrate the aggregated total demand forecast value. During the calibration process, the growth rate trend in market trend data and the regulatory impact in policy environment data are also considered. The calibration coefficient (obtained statistically from the deviation ratio between the current aggregated total demand forecast value and the actual sales volume of the same period in history, with a value between 0.5 and 1.5, determined based on the statistical distribution of historical forecast deviations, covering relevant fluctuation scenarios, preserving trend information while avoiding over-calibration due to excessive historical deviations) is fine-tuned to ensure that the forecast results conform to the direction of macroeconomic changes. The calibrated total demand forecast value is then aggregated by potential region to generate local demand forecast data corresponding to each potential region.

[0038] S4. The macro demand potential energy map and local demand forecast data are weighted, fused, and error-corrected to form a multi-granularity demand forecast report. The confidence level of the multi-granularity demand forecast report is evaluated, and a sports demand forecast score is output.

[0039] Align and match the macro-level demand potential map with local demand forecast data in terms of time scale and geographical level to generate a demand fusion dataset; Specifically, the demand intensity value of each geographic grid in the macro demand potential energy map is expanded by time step to form a daily time series data. Simultaneously, the local demand forecast values ​​for each potential region in the local demand forecast data are interpolated or aggregated at the same time granularity to unify them into a daily-granularity time series. At the geographic level, matching is performed by extracting multiple geographic grids covered by each potential region from the macro demand potential energy map and calculating the average demand intensity of all geographic grids within the potential region as the representative value of the macro demand intensity for that potential region. The representative value of the macro demand intensity corresponding to the same potential region is aligned with the calibrated local demand forecast value daily in chronological order to form a four-dimensional structured data entry. Finally, the four-dimensional structured data entries for all potential regions at all time steps are summarized to generate a demand fusion dataset containing time, geographic, macro, and local forecast information.

[0040] Based on a preset weighting strategy, the demand fusion dataset is weighted and statistically analyzed to generate sports market demand data. Specifically, based on the demand fusion dataset, for each time point and each potential region, the corresponding representative value of macro demand intensity and the local demand forecast value are extracted. According to a preset weighting strategy (which considers the error level and stability between the prediction results of the macro demand potential map in historical periods and the actual historical sales data, combined with the prediction deviation performance of the local demand forecast data within the same time period, and assigns higher weights to data sources with smaller errors and higher stability), the mean square error of the macro demand intensity representative value and the local demand forecast value relative to the actual sales data in historical periods is calculated, and the standard deviation of the mean square error of the macro demand intensity representative value and the local demand forecast value over time is calculated to measure stability. The process involves: normalizing the mean squared error and standard deviation to generate confidence scores for the macroeconomic demand intensity representative value and the local demand forecast value; determining the weight coefficients for the macroeconomic demand intensity representative value and the local demand forecast value based on the proportional weighting of their confidence scores; weighting the macroeconomic demand intensity representative value with its corresponding weight coefficient and the local demand forecast value with its corresponding weight coefficient; combining the two weighted results to obtain the comprehensive demand forecast value for the time point and potential region; performing the same weighted calculation operation on all time points and all potential regions in the demand fusion dataset to generate a set of comprehensive demand forecast values ​​covering multiple time steps and multiple potential regions, i.e., sports market demand data; The formula for calculating the comprehensive demand forecast is as follows: ; in, Identifiers representing each time point in the demand fusion dataset. Identifiers representing potential regions in the demand fusion dataset. Indicates a point in time and potential areas The comprehensive demand forecast, The weighting coefficients represent the values ​​of macroeconomic demand intensity. The weighting coefficients represent the local demand forecasts. Indicates a point in time and potential areas The representative value of macroeconomic demand intensity, Indicates a point in time and potential areas Local demand forecasts.

[0041] Multi-granularity data integration and report formatting of sports market demand data are used to generate multi-granularity demand forecast reports; Specifically, the comprehensive demand forecast values ​​from the sports market demand data are layered and integrated according to different analytical dimensions. In terms of time dimension, aggregation is performed to generate daily, weekly, and monthly forecast summary data. In terms of geographical dimension, it is hierarchically aggregated by city, province, and region to form regional aggregated data. In terms of product dimension, it combines product category information to generate category-specific forecast data. The integrated multi-dimensional forecast results are then structured and formatted to generate a multi-granularity demand forecast report that includes chart generation, trend line drawing, key indicator annotation, and text description filling.

[0042] The predicted values ​​in the multi-granularity demand forecast report are compared with historical sales data point by point to generate a residual sequence. Specifically, the predicted values ​​for historical time periods are extracted from the multi-granularity demand forecast report to ensure that the time granularity and geographical scope are completely consistent with the corresponding historical sales data; the predicted values ​​for the same time point and the same sales region are compared with the actual historical sales data point by point to obtain the prediction error of each data point, i.e., the demand residual; the demand residuals of all matching time points and sales regions are integrated to form a set of residual sequences arranged in chronological order. Autocorrelation analysis and white noise test are performed on the residual sequence to generate a time series stability index; Autocorrelation analysis is performed on the residual series. The autocorrelation coefficients of the demand residuals at multiple lag orders are calculated. This is achieved by analyzing the consistency of fluctuations in the residual values ​​at the current and historical time points under different lag orders, determining whether there is a time dependency between the residual values. If the autocorrelation coefficients are generally close to zero, it indicates no significant autocorrelation. If the autocorrelation coefficients of the residual series deviate from zero at multiple lag orders, significant autocorrelation is considered to exist. Furthermore, a white noise test is performed on the residual series to verify whether it conforms to the characteristics of white noise with zero mean, constant variance, and no serial correlation. If the test passes, it indicates that the prediction error is randomly distributed, and the predicted values ​​in the multi-granularity demand forecast report have good stability. A nonlinear mapping is applied to the autocorrelation analysis results and the white noise test results, transforming the pass rate and autocorrelation strength from both results into a time-series stability index reflecting the stability of the prediction error. The formula for calculating the autocorrelation coefficient is: ; in, This represents the lag order of the residual sequence. Indicates the residual sequence in terms of lag order. The autocorrelation coefficient under the following conditions Identifiers representing each time point in the residual sequence. Indicates the first Residual values ​​at each time point Indicates the first Residual values ​​at each time point This represents the average value of the entire residual sequence. This represents the total length of the residual sequence.

[0043] Based on market trend data and policy environment data, this study uses event correlation analysis to assess the impact of market trend data and policy environment data on sports market demand forecasts and generates event impact coefficients. Specifically, the sliding window detection method is used to identify trend change points from market trend data, such as shifts in consumer preferences and the rise of new sports. Major policy release events, such as sports industry support policies and permits for hosting large-scale events, are extracted from policy environment data. The occurrence time of these major policy release events is aligned with the predicted values ​​in multi-granularity demand forecast reports to analyze the fluctuations in predicted values ​​before and after the events, and to obtain the magnitude of changes in prediction deviations after the events occur. The event correlation analysis method is used to quantify the impact strength of each event on the prediction results. This involves comparing the magnitude of changes in predicted values ​​within the sliding time window before and after the event, and combining this with the duration and significance of the deviation changes to statistically analyze the impact strength of the events on the prediction results, forming an event impact score. All event impact scores are normalized to obtain event impact coefficients, which are used to measure the degree of interference from external environmental factors on the reliability of predictions.

[0044] The time series stability index and event impact coefficient are scored, fused, and mapped to output a sports demand prediction score. Specifically, the time-series stability index and event impact coefficient are used as inputs. These two indices are weighted and combined into a comprehensive score. The weights are set based on historical validation results. Specifically, by analyzing the correlation strength between the time-series stability index and event impact coefficient and the actual prediction error during historical prediction phases, the contribution of the time-series stability index and event impact coefficient to prediction reliability is determined, and corresponding weights are assigned. For example, a weight of 0.7 can be set when stability contributes more, and a weight of 0.3 for event impact coefficients. Specific values ​​are derived through backtesting and optimization using historical data. The weighted result is converted into a standardized score from 0 to 100 using a nonlinear mapping function. A higher score indicates a more reliable prediction result. The standardized score is output as a sports demand prediction score, used to assist in judging the overall credibility of the multi-granularity demand prediction report.

[0045] S5. Develop sports market demand strategies based on sports demand forecasting scores and multi-granularity demand forecasting reports, and transmit these strategies to enterprise managers through the enterprise decision support interface.

[0046] The sports demand forecasting score and multi-granularity demand forecasting report are matched with a pre-set set of strategy templates to output a draft sports demand strategy. Specifically, based on the numerical range of the sports demand forecast score, the corresponding risk level category is searched in a pre-defined strategy template set. The strategy template set contains multiple predefined strategy templates, each corresponding to a different forecast reliability level, such as high, medium, and low. At the same time, comprehensive demand forecast values, main growth areas, key product categories, and time distribution trends for each potential region are extracted from the multi-granularity demand forecast report. The extracted information is then compared item by item with the condition fields in the currently matched strategy template, and blank items such as regional configuration, product launch, marketing rhythm, and resource allocation suggestions are filled in. After the filling is completed, a preliminary structured strategy plan, namely the sports demand strategy draft, is generated. Furthermore, the strategy template set is a structured collection of strategies summarized from historical patterns of sports market demand changes and successful past response strategies of enterprises. Each strategy template corresponds to a specific market response model, including high demand growth response plans, low demand early warning response plans, regional expansion plans, and category-focused promotion plans. The strategy template content includes fields such as target area selection rules, recommended product combinations, marketing campaign launch timing, resource investment intensity suggestions, and expected achievement goals. All strategy templates are classified and archived according to the applicable sports demand forecast scoring range, and maintain field compatibility with the geographical level, time range, and product dimension in the multi-granularity demand forecast report to ensure dynamic matching and filling based on the latest forecast results.

[0047] The draft sports demand strategy is validated for completeness and feasibility, and a sports market demand strategy is output. Specifically, the draft sports demand strategy is checked to ensure it contains all necessary fields, such as target region, recommended products, execution timeframe, resource allocation recommendations, and expected target values. If any are missing, they are supplemented based on the corresponding data in the multi-granularity demand forecast report. A feasibility check is performed by comparing the resource allocation recommended in the draft sports demand strategy with the company's historical resource allocation to determine if there are any abnormal configurations that exceed the normal range (the normal range refers to the statistical interval of resource allocation under similar strategies or similar demand levels in the company's historical resource allocation, usually within one standard deviation above or below the historical average). At the same time, policy environment data is used to verify whether there are any access barriers in the recommended region. If any infeasible items are found, they are corrected. All correction operations during the verification process are logged to ensure traceability. The draft sports demand strategy that passes the completeness and feasibility checks is confirmed as the official version, and the sports market demand strategy is output.

[0048] The sports market demand strategy is encapsulated into data messages through the enterprise decision support interface and transmitted to enterprise managers via the HTTPS protocol to complete the sports market demand forecast. Specifically, the sports market demand strategy is formatted according to the company's internal data structure and converted into a standard JSON string. This JSON string includes fields such as strategy number, generation time, strategy content body, confidence level, and attachment links. The formatted JSON string is then encapsulated as the payload in an HTTP request body, with the request header set to application / json type and the necessary authentication token configured. A secure communication connection is initiated through the company's decision support interface, and the encapsulated data message is sent to the receiving endpoint specified by the company administrator using the HTTPS protocol. After transmission, the status code returned by the server is received. Once the transmission process is complete, the sports market demand forecasting process is marked as finished, thus completing the sports market demand forecasting.

[0049] This embodiment also provides a computer device applicable to the sports market demand forecasting method based on deep learning, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the sports market demand forecasting method based on deep learning as proposed in the above embodiment.

[0050] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing 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 communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0051] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the deep learning-based sports market demand prediction method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0052] In summary, this invention achieves multi-level, accurate prediction and decision support for sports market demand through the synergistic effect of graph attention spatiotemporal networks and spatiotemporal convolution. The graph attention spatiotemporal network performs relationship mining and graph structure learning on preprocessed sports demand data, dynamically capturing complex relationships between market entities and providing a structured foundation for spatiotemporal pattern analysis. Spatiotemporal convolution extracts spatiotemporal patterns from the dynamic knowledge graph, generating a macro-level demand potential graph to quantify global market demand trends. Combined with a pre-trained sports market simulation model, potential region identification and consumer decision simulation are performed to generate executable strategies, comprehensively improving the granularity, reliability, and practicality of the prediction results.

[0053] It should be noted that the above 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based method for predicting sports market demand, characterized in that: include, Collect sports demand data and preprocess it; the sports demand data includes marketing strategy data, historical sales data, market trend data, and policy environment data; By using a graph attention spatiotemporal network to perform relation mining and graph structure learning on the preprocessed sports demand data, a dynamic knowledge graph is output. Spatiotemporal convolution is then used to extract spatiotemporal patterns from the dynamic knowledge graph to generate a macro-demand potential energy graph. A pre-trained sports market simulation model is used to identify potential areas in the macro demand potential map, and marketing strategy data is loaded to simulate the consumer decision-making process, outputting local demand forecast data. The macro demand potential map and local demand forecast data are weighted, fused, and error-corrected to form a multi-granularity demand forecast report. The confidence level of the multi-granularity demand forecast report is then assessed, and a sports demand forecast score is output. Develop sports market demand strategies based on sports demand forecasting scores and multi-granularity demand forecasting reports, and transmit these strategies to enterprise managers through an enterprise decision support interface.

2. The deep learning-based sports market demand forecasting method according to claim 1, characterized in that: The process involves using a graph attention spatiotemporal network to perform relationship mining and graph structure learning on the preprocessed sports demand data, outputting a dynamic knowledge graph. The specific steps are as follows: Different sports entities in the preprocessed sports demand data are defined as market entity nodes. Business relationship edges are established according to the business logic between sports entities, and integrated to form an initial structure diagram. The dynamic influence weights between market entity nodes are obtained by using a graph attention spatiotemporal network, and the dynamic influence weights are added to the initial structure graph to generate a weighted structure graph. Feature extraction is performed on the weighted structure graph in the spatiotemporal dimension, and the spatiotemporal feature matrix is ​​output. Based on the spatiotemporal feature matrix, the market entity nodes of the weighted structure graph are updated, and a dynamic knowledge graph is output.

3. The deep learning-based sports market demand forecasting method according to claim 2, characterized in that: The method employs spatiotemporal convolution to extract spatiotemporal patterns from dynamic knowledge graphs, generating a macroscopic demand potential energy map. The specific steps are as follows: Perform spatiotemporal structuring transformation on dynamic knowledge graphs and output spatiotemporal feature tensors; Spatiotemporal feature tensors are subjected to pattern recognition and feature diffusion analysis through spatiotemporal convolution, and spatiotemporal feature maps are output. Regional feature aggregation and geographic grid mapping are performed on the spatiotemporal feature map to generate a macroscopic demand potential energy map.

4. The deep learning-based sports market demand forecasting method according to claim 3, characterized in that: The specific steps for identifying potential regions using a pre-trained sports market simulation model on the macro demand potential map are as follows: Regional grid demand intensity analysis is performed on the macro demand potential energy map using a grid clustering algorithm to obtain regional demand intensity data. Based on regional demand intensity data, a pre-trained sports market simulation model is used to assess the regional potential of sports market demand and output a list of regional potential assessments. The list of regions with potential is sorted in reverse order and intensity is selected to output a list of potential regions.

5. The deep learning-based sports market demand forecasting method according to claim 4, characterized in that: The process of loading marketing strategy data to simulate the consumer decision-making process and outputting partial demand forecast data involves the following steps: The marketing strategy data is linked and matched with the regional potential assessment list to form a strategy-region configuration set; Perform consumer decision simulation on the strategy-region configuration set, generate a simulated decision dataset, summarize the demand on the simulated decision dataset, and output the initial demand forecast results; The initial demand forecast results are aggregated and calibrated across regions to form local demand forecast data for each potential region.

6. The deep learning-based sports market demand forecasting method according to claim 5, characterized in that: The specific steps for weighted fusion and error correction of the macro-level demand potential map and local demand forecast data to generate a multi-granularity demand forecast report are as follows: Align and match the macro demand potential map with local demand forecast data to generate a demand fusion dataset; Based on a preset weighting strategy, the demand fusion dataset is weighted and statistically analyzed to generate sports market demand data. The sports market demand data is integrated and formatted into a multi-granularity demand forecast report.

7. The deep learning-based sports market demand forecasting method according to claim 6, characterized in that: The specific steps for performing a confidence assessment on the multi-granularity demand forecast report and outputting a sports demand forecast score are as follows: The predicted values ​​in the multi-granularity demand forecast report are compared with historical sales data point by point to generate a residual sequence. Autocorrelation analysis and white noise test are performed on the residual sequence to generate a time series stability index; Based on market trend data and policy environment data, this study uses event correlation analysis to assess the impact of market trend data and policy environment data on sports market demand forecasts and generates event impact coefficients. The time series stability index and event impact coefficient are scored and fused to output a sports demand prediction score.

8. The deep learning-based sports market demand forecasting method according to claim 7, characterized in that: The process involves developing a sports market demand strategy based on sports demand forecasting scores and multi-granularity demand forecasting reports, and then transmitting this strategy to enterprise managers through an enterprise decision support interface. The specific steps are as follows: The sports demand forecasting score and multi-granularity demand forecasting report are matched with a pre-set set of strategy templates to output a draft sports demand strategy. The draft sports demand strategy is validated for completeness and feasibility, and a sports market demand strategy is output. The sports market demand strategy is encapsulated into data messages through the enterprise decision support interface and transmitted to enterprise managers via the HTTPS protocol to complete the sports market demand forecast.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based sports market demand forecasting method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based sports market demand forecasting method according to any one of claims 1 to 8.

Images