Market signal processing method and system

Abstract

The invention relates to the technical field of data processing, in particular to a market signal processing method and system. The method comprises the following steps: acquiring a market signal, wherein the market signal at least comprises e-commerce platform sales volume data of a target product and social media sound volume data representing the popularity of the target product; performing wavelet transform on the market signal to obtain a wavelet coefficient; calculating a noise threshold, and comparing the energy intensity of the wavelet coefficient with the noise threshold to filter noise in the market signal; and based on the market signal after noise filtering, constructing a market state vector capable of representing the market state at each moment, and inputting the market state vector at each moment in a preset historical period into a pre-trained classifier to obtain the probability that the current market is in each market state. The method and the device have the effect of improving the accuracy of market state recognition.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a market signal processing method and system. Background Technology

[0002] With the development of data-driven marketing and intelligent decision-making technologies, these technologies are rapidly evolving and profoundly changing corporate marketing models and decision-making processes. In the process of carefully formulating marketing strategies, companies are increasingly reliant on the real-time collection and in-depth analysis of market signals. Market signals act as a weathervane in the business world, accurately reflecting dynamic market changes and guiding companies in their direction.

[0003] Market signals encompass a wide range of data, including but not limited to multi-source time-series data such as sales changes on e-commerce platforms, public opinion trends on social media, search trends, and competitor price adjustments. Sales changes on e-commerce platforms directly reflect consumers' willingness and demand for products, providing a direct indication of market demand. Public opinion trends on social media reflect the level of discussion and emotional inclination among consumers regarding brands and products, helping businesses understand consumer attitudes and expectations. Search trends reveal the degree of consumer attention to specific products or services, indicating potential market demand. Competitor price adjustments affect the competitive landscape, requiring close monitoring by businesses to formulate appropriate pricing strategies. These multi-source time-series data intertwine and influence each other, collectively forming a complex and ever-changing market signal system. They can accurately reflect changes in market demand, consumer behavior trends, and the evolution of the competitive landscape to a certain extent, and are therefore applied in marketing decision-making systems, market monitoring systems, and automated marketing platforms, becoming an important basis for businesses to formulate marketing strategies.

[0004] In related technical fields, market signal processing methods primarily focus on trend identification based on statistical analysis or simple time series models. For example, a common approach involves using moving averages, exponential smoothing, or year-on-year / month-on-month calculations on historical data to determine whether the market is experiencing growth or decline. Moving averages, by averaging data over a certain period, eliminate the impact of short-term fluctuations, thus more clearly showing the long-term trend of the data; exponential smoothing assigns different weights to data from different periods, paying more attention to the impact of recent data and reflecting market changes more promptly; year-on-year / month-on-month calculations, by comparing data with the same period or previous period, intuitively show the magnitude of market growth or decline. Another approach involves setting fixed thresholds. When indicators such as sales volume, click-through rate, or social media buzz exceed or fall below preset thresholds, it triggers alerts or prompts decision-makers to intervene. This approach is simple and direct, promptly alerting companies to market changes when indicators show anomalies. In addition, some systems incorporate regression models for periodic analysis or trend fitting of market changes. Regression models predict future market trends by establishing relationships between variables.

[0005] However, in practical applications, market signals themselves exhibit significant non-stationarity and high noise characteristics. On the one hand, market data is easily affected by short-term promotional activities, holiday effects, platform algorithm adjustments, data collection delays or anomalies, leading to a large amount of random fluctuations or occasional anomalies mixed in with the signals. On the other hand, differences in sampling frequency, statistical caliber, and response lag between different data sources make it difficult to directly align market signals across multiple dimensions in terms of time and amplitude. Under these circumstances, existing technical solutions based on simple smoothing or threshold judgment often fail to effectively distinguish between valid signals that truly reflect structural changes in the market and random noise.

[0006] Furthermore, existing technologies typically focus on monitoring "numerical changes" while lacking the ability to identify the market's current stage or evolutionary state. For example, when sales of a certain product category experience a short-term decline, existing systems often only provide a "decline warning," but cannot determine whether the change is caused by seasonal fluctuations, a demand inflection point, intensified competition, or a market entering a recession. Due to the lack of identification of market evolution phases, decision-making systems struggle to distinguish between short-term tactical adjustments and medium- to long-term strategic adjustments, easily leading to delayed decision-making or directional errors.

[0007] In summary, market signal analysis methods in related technologies struggle to distinguish between real market change signals and random noise in a timely and accurate manner under multi-source, high-noise market signal environments, and further identify the market's evolutionary stage, leading to erroneous and outdated market decision-making. Summary of the Invention

[0008] To improve the accuracy of market signal identification and market phase judgment, this application provides a market signal processing method and system.

[0009] In a first aspect, this application provides a market signal processing method, which adopts the following technical solution: A market signal processing method includes: acquiring market signals, wherein the market signals include at least e-commerce platform sales data of the target product and social media volume data representing the popularity of the target product; Wavelet transform is performed on the market signal to obtain wavelet coefficients; a noise threshold is calculated, and the energy intensity of the wavelet coefficients is compared with the noise threshold to filter noise in the market signal; Based on the noise-filtered market signal, construct a market state vector that can represent the market state at each time. Input the market state vector at each time within a preset historical period into a pre-trained classifier to obtain the probability of the current market being in each market state. Among them, the market state vector includes at least the demand growth characteristics, which are positively correlated with the current e-commerce platform sales volume and also positively correlated with the growth rate of positive media volume in social media volume data.

[0010] Acquiring both traditional e-commerce platform sales data and social media buzz data, which characterize market activity, effectively compensates for the lag in relying solely on sales data. For example, in the early stages of a product's surge, social media discussion often precedes sales growth. Through this fusion of perception, the system can detect market recovery or crisis signals earlier. Wavelet transform is then used to decompose the signal into different scales. Wavelet transform possesses time-frequency localization properties, effectively separating low-frequency components representing long-term trends from high-frequency components representing random disturbances (such as crawler anomalies or instantaneous data spikes). This allows the system to filter out invalid background noise while retaining structured information in non-stationary market signals, avoiding misjudgments caused by signal distortion. Finally, by constructing a market state vector and inputting it into a pre-trained classifier, accurate and timely identification of market states is achieved in multi-source, high-noise market signal environments. Furthermore, demand growth features are introduced into the market state vector, calculating the growth rate of both source data, significantly enhancing the system's sensitivity to changes in market demand for target products. In real-world scenarios, a single increase in sales volume may only be a short-term result of promotional activities. However, if it is accompanied by a simultaneous positive increase in positive media coverage, it signifies a genuine increase in market demand. Therefore, introducing demand growth characteristics can improve the accuracy of subsequent market status identification. Furthermore, it filters out the volume of negative information from the original text data corresponding to social media buzz. This eliminates the possibility of a target product market boom caused by negative events, making the demand growth characteristics more accurately reflect market growth momentum.

[0011] Optionally, positive media volume refers to the volume of data containing positive evaluations or purchase motivations for the target product, or objective descriptions or routine inquiries.

[0012] Positive reviews, purchase motivations, objective descriptions, and routine inquiries all reflect the media's attention to the target product and can, to some extent, translate into sales. Therefore, incorporating all of these factors into positive media coverage can improve the accuracy of positive media coverage.

[0013] Optionally, the noise threshold is calculated, including: for any given time, constructing a historical benchmark sliding window based on that time; determining the noise threshold based on the historical statistical characteristics of the noise obtained after decomposing the market signal in the historical benchmark sliding window and the number of original observation points collected in the historical benchmark sliding window at that time; the noise threshold is positively correlated with the number of original observation points in the historical benchmark sliding window.

[0014] By analyzing the statistical characteristics of high-frequency components within the sliding window, the system can dynamically adjust the noise filtering level based on the recent background noise level in the market. For example, during promotional seasons with drastic market fluctuations, noise levels generally rise, and the system automatically raises the threshold to prevent false alarms; while during stable periods, the threshold is lowered to capture weak signals. Simultaneously, the introduction of the original number of observation points for statistical calibration resolves the statistical distortion problem caused by missing or filled data.

[0015] Optionally, the mean of noise can be extracted from the reference signal of a pre-selected specific time period, and the sum of the mean and the standard deviation of the preset multiple can be used as the noise threshold for each time period.

[0016] By pre-selecting a stable baseline signal that excludes major promotions (such as Singles' Day), the system can extract the basic background noise level of the product's industry. By combining the mean with multiples of the standard deviation, and utilizing the three-standard-deviation criterion in probability theory, it can effectively intercept most random interferences with relatively low computational overhead.

[0017] Optionally, market signals may also include: competitor prices of similar products to the target product; the market state vector may also include: competitive volatility characteristics representing the volatility of competitor prices and cross-channel correlation characteristics representing the correlation between changes in sales volume on e-commerce platforms and positive media coverage.

[0018] Competitive fluctuations can reflect the pressure of price wars in the market and help identify whether the market has entered a turning point where demand for the target product changes rapidly due to competitors' actions.

[0019] Optionally, the calculation steps for demand growth characteristics include: for any given moment, obtaining the growth rate of e-commerce platform sales and the growth rate of positive media buzz at that moment, and using the weighted sum of the two as the demand growth characteristics.

[0020] Demand growth characteristics are obtained by weighted summing of the growth rate of e-commerce platform sales and the growth rate of positive media voice, so that the demand growth characteristics can comprehensively represent the growth of e-commerce platform sales and positive media voice.

[0021] Optionally, the pre-trained classifier is a phase classifier based on the Transformer architecture.

[0022] The introduction of the Transformer architecture enables the system to handle long-term time dependencies. The evolution of market signals often has a causal lag; for example, a marketing campaign two weeks ago (positive media buzz) may be the main reason for this week's sales inflection point. The Transformer's self-attention mechanism can automatically identify the correlation between these discontinuous time points across historical periods of up to 30 days.

[0023] Optionally, the market status includes at least two of the following: a demand inflection point characterized by a sudden change in demand for the target product; a plateau period characterized by stable demand for the target product; and a recession period characterized by a continuous contraction of the target product market.

[0024] Optionally, if the probability of responding to a market state is greater than a preset threshold and persists for more than a preset duration, the market is determined to be in that market state.

[0025] Secondly, this application provides a market signal processing system, which adopts the following technical solution: A market signal processing system includes a processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the market signal processing method described above.

[0026] The market signal processing method described above generates a computer program, which is stored in a memory so that it can be loaded and executed by a processor. Thus, a system can be built based on the memory and processor for convenient use.

[0027] This application has the following technical effects: This paper utilizes wavelet transform to denoise multi-dimensional market signals in the time-frequency domain, and combines it with dynamic thresholding technology based on data density to effectively separate random noise from true trends. Secondly, it constructs a multi-dimensional feature that integrates sales volume, social media buzz, competitor prices, and cross-channel correlation. Finally, it applies the Transformer model to accurately identify the evolutionary stage of the market, thereby solving the problem that traditional statistical methods cannot accurately judge complex market states in high-noise environments. Attached Figure Description

[0028] Figure 1This is a flowchart of a market signal processing method according to an embodiment of this application.

[0029] Figure 2 This is a schematic flowchart of a market signal processing method according to an embodiment of this application. Detailed Implementation

[0030] This application discloses a market signal processing method. By accessing market signals from multiple dimensions, it utilizes wavelet transform technology to perform deep decomposition of the signals in the time-frequency domain. Subsequently, it dynamically calculates a noise threshold based on historical noise distribution, filtering out invalid fluctuations and retaining significant signal feature components. Finally, it combines the market state vector with... The phase classifier in the architecture outputs the probability distribution of market states. This improves the real-time performance of signal processing and ensures the accuracy of phase determination through adversarial verification and time-frequency analysis, thus effectively solving the problem of decision failure caused by sudden changes in market trends.

[0031] Reference Figure 1 A market signal processing method includes steps S1-S3.

[0032] S1: Obtain market signals, which include at least e-commerce sales data for the target product and social media buzz data that represents the popularity of the target product.

[0033] In this embodiment, market signals from multiple dimensions are acquired; for any given moment, this includes data from multiple dimensions, which can be represented as: .in, Indicates the first Data sources at time The data values. Market signals from multiple dimensions include, but are not limited to, the e-commerce sales volume of the target product, the social media buzz of the target product, the competitor prices of the target product, and the supply chain inventory of the target product. Here, the target product is the product that needs to be observed or investigated.

[0034] E-commerce platform sales figures for a target product provide direct insights into market trends. For example, an increase in e-commerce platform sales for a target product indicates that its overall market share is expanding.

[0035] Social media buzz surrounding a target product refers to the frequency of discussions, number of posts, and popularity of comments about the product or category on platforms such as Xiaohongshu, Weibo, and Douyin. This is a precursor to identifying a market recovery period or a public opinion crisis.

[0036] The prices of competing products for the target product, which are products that directly compete with the target product, can also be understood as products from different brands. Real-time monitoring of competitors' final prices, promotional discounts, and subsidies across various channels reflects competitive pressure.

[0037] Supply chain inventory of the target product refers to the real-time inventory level and in-transit volume of finished goods warehouses, distribution centers, and stores.

[0038] It is understandable that raw market signals often exhibit inconsistent dimensions and sampling frequencies. In this embodiment, the preprocessing step includes normalizing the signals across different dimensions, for example, by using... Standardization maps data to Intervals are used to eliminate differences in units. Meanwhile, for missing data, linear interpolation or historical averages are used to fill in the gaps, ensuring the continuity of the time series.

[0039] S2: Perform wavelet transform on the market signal to obtain wavelet coefficients; calculate the noise threshold, and filter the noise in the market signal by comparing the energy intensity of the wavelet coefficients with the noise threshold.

[0040] In order to separate the true trend signal from random noise, this embodiment uses wavelet transform to decompose the preprocessed signal to obtain wavelet coefficients.

[0041] Specifically, in this embodiment, the wavelet transform is not just a single-level decomposition, but a multi-scale discrete wavelet transform (DWT) is employed. The Daubechies (db4) wavelet basis is selected to perform a three-level decomposition of the market signal. The first level consists of high-frequency components. Capture extremely short-term random disturbances (such as abnormal crawling behavior); second layer Capture sudden, small-scale promotional disruptions; third-layer low-frequency components This represents the structured trend of the market. Wavelet transform can decompose market signals into different frequency scales. The high-frequency component usually corresponds to the instantaneous noise or random fluctuations in the market, while the low-frequency component better reflects the long-term trend of the market. Compared with the traditional moving average method, wavelet transform can more sensitively capture inflection point information in non-stationary signals.

[0042] Subsequently, noise in the market signal is filtered out by calculating the noise threshold and comparing the noise threshold with the energy intensity of the wavelet coefficients.

[0043] In one embodiment, in order to reduce the computational overhead of the system and improve the efficiency of the initial cleaning, the system can use a fixed threshold filtering method to purify market signals.

[0044] Specifically, a reference signal is first pre-selected from historical signals. The length of the reference signal is typically set to 14 to 30 working days. Periods with major promotional events, such as the "Double Eleven" promotional period, need to be removed from the reference signal. The mean noise in this reference signal is calculated, which is the mean noise obtained after wavelet decomposition. The sum of the mean noise and three times the standard deviation of the noise is used as the noise threshold.

[0045] In another embodiment, noise can be filtered using a dynamic threshold.

[0046] Regarding the calculation of the dynamic noise threshold, in one embodiment, a historical benchmark sliding window can be constructed based on each time point; the noise threshold is determined according to the historical statistical characteristics of the noise obtained after decomposing the market signal in the historical benchmark sliding window and the number of original observation points collected in the historical benchmark sliding window at that time; the noise threshold is positively correlated with the number of original observation points in the historical benchmark sliding window.

[0047] Specifically, for any given moment, a historical benchmark sliding window is first established based on that moment. In this embodiment, the duration of this window is set to seven days prior to the current moment. Then, the mean and standard deviation of the historical noise within the benchmark sliding window are calculated. The steps for obtaining the mean of the historical noise include: In market signals, the true trends of sales growth and price adjustments typically exhibit low-frequency smooth changes, while random interference such as daily data spikes and invalid clicks exhibit high-frequency sharp fluctuations. Therefore, the system defines the decomposed high-frequency components as a noise signal sequence. The mean of each high-frequency component in the noise signal sequence is used as the mean of the historical noise. The standard deviation of each high-frequency component in the noise signal sequence is the standard deviation of the historical noise. Simultaneously, the number of original observation points actually obtained within the historical benchmark sliding window is counted, excluding data points filled in step S1 using linear interpolation or historical averages.

[0048] For any given moment, the formula for calculating the noise threshold can be expressed as: ; In this formula, Indicates time The noise threshold; This represents the mean of historical noise. The standard deviation of historical noise; express The number of data points accumulated over time; The confidence interval adjustment coefficient is preferably used in this embodiment. ,correspond The confidence interval. It can adaptively adjust based on the volatility of the industry in which the target product operates. For example, in the rapidly iterating FMCG industry, The value is 1.65 ( (Confidence interval) to improve sensitivity to new signals; while in the field of bulk industrial products, The value is 2.58. (Confidence interval) to ensure that only significant market reversal signals are captured.

[0049] The formula incorporates historical noise mean and standard deviation, enabling the threshold to dynamically adjust with changes in the market environment. During periods of market stability, when noise levels are low, the threshold automatically decreases, allowing the system to sensitively detect extremely weak change signals. During periods of drastic market signal fluctuations, the threshold automatically increases, effectively shielding against non-featured random interference.

[0050] By utilizing the asymptotic properties of extreme value theory, the number of effective sampling points within the sliding window is statistically calibrated. When there are many original sampling points within the window, the risk of statistical error increases, and this factor drives a slight increase in the threshold to maintain stringency. When data loss leads to a decrease in the number of original sampling points, the threshold shrinks accordingly to compensate for the loss of statistical confidence. This ensures that the false alarm rate of the system can be strictly controlled under any data density. Within the confidence interval.

[0051] Finally, for any data point, if the energy intensity of the wavelet coefficients responding to that moment is higher than the noise threshold, the system determines that the fluctuation is statistically significant and does not belong to random noise. In this case, the corresponding wavelet coefficients are considered as valid market signals. Conversely, if the energy intensity of the wavelet coefficients responding to that moment is less than or equal to the noise threshold, the system determines that the data point is random interference caused by background noise, and in this case, the data point is discarded.

[0052] S3: Construct market state vectors that can represent the market state at each time point based on the market signal after noise filtering. Input the market state vectors at each time point within a preset historical period into a pre-trained classifier to obtain the probability of the current market being in each market state.

[0053] After the market signals are filtered and evolved, a pure signal is obtained. A market state vector is constructed based on the pure signal to describe the current operating state of the market.

[0054] In one embodiment, the market state vector may include demand growth characteristics, competition fluctuation characteristics, seasonal residual characteristics, and cross-channel correlation characteristics. Specifically, the market state vector at any given time can be represented as: ;in, Representative moment The corresponding market state vector; Representative moment The demand growth characteristics are calculated by weighted summation of the growth rates of sales volume on e-commerce platforms and positive media volume in social media data; Representative moment The competitive fluctuation characteristics are characterized by the degree of dispersion in competitor prices, which represents market competitive pressure. Representative moment The seasonal residual characteristics represent abnormal market fluctuations after removing regular seasonal patterns; This represents the cross-channel relevance characteristics, specifically the correlation coefficient between social media buzz and e-commerce sales.

[0055] In other embodiments, the market state vector may include only one, two, or three of the above features, and the specific details can be adjusted according to market conditions during implementation. For example, the market state vector may include only the demand growth feature.

[0056] For any given moment's demand growth characteristics, the system first removes negative media volume from social media buzz data to obtain positive media volume. Specifically, a pre-trained sentiment analysis model scans the original text stream corresponding to the social media buzz data. A deep learning model (such as a BERT-based sequence labeling model) is used to classify the volume data into three categories: positive text data, neutral text data, and negative text data. Positive volume is defined as signals containing positive evaluations or purchase motivations; neutral volume is defined as objective descriptions or routine inquiries. The system then weights and sums the volumes corresponding to the positive and neutral text data to obtain the positive media volume. Subsequently, the growth rate of current e-commerce platform sales and the growth rate of positive media volume in social media buzz are calculated. These two weighted sums are then used to obtain the demand growth characteristics.

[0057] Specifically, the formula for calculating the characteristics of demand growth can be expressed as: In the formula, Indicates time The characteristics of demand growth; Indicates time Sales on e-commerce platforms; Indicates time Sales on e-commerce platforms; Indicates time Positive media coverage; Indicates time Positive media coverage; Indicates the first weighting coefficient; This represents the second weighting coefficient; This represents a preset zero-prevention constant, which can be 0.001. It is mainly used to prevent calculation errors caused by a zero denominator.

[0058] and The configuration that best suits the target product can be determined through historical backtesting experiments. Specifically, select a known historical data period (e.g., data from the entire previous year) that includes distinct periods of demand surges or market declines. Experiment with different ratios between the two weighting coefficients, such as a first weighting coefficient of 70% and a second weighting coefficient of 30%, or both at 50%. Input the demand growth characteristic value calculated for each ratio into the classifier and see which ratio best matches the actual market state at that time. If the experiment finds that the system can predict the actual market state most quickly and accurately when the first weighting coefficient is set to 40% and the second weighting coefficient is 60%, then this set of values ​​is determined as the first and second weighting coefficients for the target product.

[0059] The demand growth characteristics can overcome the limitations of a single data source in measuring market trends by weighting and compounding the growth rate of sales on e-commerce platforms with the growth rate of positive media voice.

[0060] The competitive volatility characteristic is mainly used to characterize the dispersion of price changes of competing products.

[0061] In one embodiment, the competitive volatility characteristic can be the standard deviation of competitor prices within an observation window constructed at the current moment.

[0062] In another embodiment, instead of directly using the standard deviation, the interquartile range of competitor prices can be calculated. Specifically, the difference between the 75th and 25th percentiles of the competitor price series within the observation window is obtained, and this difference is used as a competitive volatility characteristic. The standard deviation is easily affected by a few exceptionally low or high prices, while the interquartile range is more robust to outliers and can more accurately reflect the price fluctuation range of the mainstream competitive market.

[0063] For the seasonal residual feature, an STL (Trend-Seasonal-Residual) time series decomposition is performed on the e-commerce platform sales data in the pure signal, and the residual term is taken as the seasonal residual feature. This reflects the degree of deviation between the current observed value and the historical benchmark (such as the same month last year or the same period last week). Through time series decomposition technology, the original signal is decomposed into a trend term, a seasonal term, and a residual term. The residual term refers to the irregular fluctuation component remaining after removing the trend and seasonal terms. It includes sudden and unpredictable signal changes beyond long-term trends and known cyclical patterns. Using the residual term as the seasonal residual feature masks fluctuations in e-commerce platform sales under normal circumstances. For example, a surge in e-commerce platform sales during promotional activities is often misjudged as a market recovery. By decomposing the seasonal term, the system can determine whether the surge conforms to historical cyclical patterns. If the surge magnitude is consistent with the seasonal term, the residual term remains low, meaning the seasonal residual feature is small, and the market can be considered unchanged.

[0064] For cross-channel correlation characteristics, at any given time, the Pearson correlation coefficient between the e-commerce platform sales sequence and the positive media buzz sequence within the observation window constructed at that time will be used. In a healthy growth phase, an increase in positive media buzz should lead to a synchronous increase in e-commerce platform sales after a certain time shift. If positive media buzz increases significantly, but e-commerce platform sales decline, it reflects a break in the decision-making chain. In real-world scenarios, this indicates that the brand may encounter a public opinion crisis or conversion bottleneck, thus determining that the market has entered a "decline period" or "plateau period."

[0065] After obtaining the market state vectors at each moment, the feature vector sequence within a preset historical period (e.g., the past seven days) is input into a pre-trained phase classifier, which outputs the probability distribution of the current market state. In this embodiment, the market states mainly include inflection points, plateau periods, recession periods, and recovery periods.

[0066] Inflection points, in particular, represent critical moments when market trends undergo a fundamental reversal. They are typically characterized by a decline in growth momentum or a surge in sudden competitive pressure, i.e., significant demand growth coupled with increased competitive volatility.

[0067] The plateau period refers to a stage where the market enters a relatively stable phase with reduced volatility. At this time, market demand reaches saturation or enters a state of competition for existing market share. Demand growth approaches zero, while seasonal residuals remain within a stable range. A disconnect emerges between e-commerce platform sales and positive media coverage, resulting in weak cross-channel correlation.

[0068] The recession phase refers to a stage where market demand continues to shrink and brand influence declines. At this time, demand growth characteristics remain negative, and cross-channel relevance characteristics may also decrease. The decrease in cross-channel relevance characteristics indicates that even if social media investment is increased, it will not be able to boost e-commerce sales.

[0069] The recovery period refers to the stage after the market has experienced a trough, when demand begins to pick up again and consumer activity increases. At this time, the seasonal residual characteristics show positive anomalies (i.e., growth that exceeds the normal seasonal pattern), and the cross-channel correlation characteristics are enhanced, indicating that positive media attention translates into sales on e-commerce platforms and a decrease in supply chain inventory.

[0070] In this embodiment, the phase classifier adopts... The architecture, utilizing its self-attention mechanism, can automatically identify the correlation between historical phases and current signals across historical periods of up to 30 days. For example, the system can identify that positive media buzz two weeks ago was the primary cause of the current sales inflection point. The classifier's output layer employs... The function maps the hidden layer state to the probability distribution of the four market states mentioned above. In this embodiment, when the probability of determining a certain market state is higher than a preset threshold and the duration exceeds a preset time, the current market is determined to be in that market state, and then the warning or decision mechanism corresponding to that market state is triggered.

[0071] Combination Figure 2 This section briefly describes the overall process of the method in this application to further illustrate the application. The program first acquires and preprocesses market signals to avoid dimensional issues. Then, noise filtering is applied to the market signals to reduce noise. Next, a market state vector is extracted from the filtered market signal and input into a pre-trained classifier. The probability of each market state is output, and a warning is issued based on the probability. For example, if the system detects that the probability of beverage A's market state being at an inflection point is... If this trend continues for more than 24 hours, the market is considered to have entered a demand inflection point. Based on this, the system can output suggestions: forecasts. If competitors launch a price war within hours, it is recommended to secure annual contracts with key customers in advance or start small-scale packaging testing. This recommendation is set manually.

[0072] This application also discloses a market signal processing system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a market signal processing method according to this application.

[0073] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0074] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A market signal processing method, characterized in that, include: Obtain market signals, which include at least e-commerce sales data for the target product and social media buzz data that indicates the popularity of the target product; Wavelet coefficients are obtained by performing wavelet transform on market signals; Calculate the noise threshold and filter noise in the market signal by comparing the energy intensity of the wavelet coefficients with the noise threshold; Based on the noise-filtered market signal, construct a market state vector that can represent the market state at each time. Input the market state vector at each time within a preset historical period into a pre-trained classifier to obtain the probability of the current market being in each market state. Among them, the market state vector includes at least the demand growth characteristics, which are positively correlated with the current e-commerce platform sales volume and also positively correlated with the growth rate of positive media volume in social media volume data.

2. The market signal processing method according to claim 1, characterized in that, Positive media volume refers to the volume of media coverage containing data that provides positive evaluations of the target product, purchase motivations, objective descriptions, or routine inquiries.

3. The market signal processing method according to claim 1, characterized in that, The noise threshold is calculated by: constructing a historical benchmark sliding window based on any given time; determining the noise threshold based on the historical statistical characteristics of the noise obtained after decomposing the market signal in the historical benchmark sliding window and the number of original observation points collected in the historical benchmark sliding window at that time; the noise threshold is positively correlated with the number of original observation points in the historical benchmark sliding window.

4. The market signal processing method according to claim 1, characterized in that, The noise threshold calculation includes: extracting the mean noise from a reference signal within a pre-selected specific time period, and using the sum of the mean and the standard deviation of a preset multiple as the noise threshold for each time period.

5. A market signal processing method according to claim 1, characterized in that, Market signals also include: competitor prices of similar products to the target product; the market state vector also includes: competitive volatility characteristics representing the volatility of competitor prices and cross-channel correlation characteristics representing the correlation between changes in sales volume on e-commerce platforms and positive media coverage.

6. A market signal processing method according to claim 2, characterized in that, The calculation steps for demand growth characteristics include: for any given moment, obtaining the growth rate of e-commerce platform sales and the growth rate of positive media buzz at that moment, and using the weighted sum of the two as the demand growth characteristics.

7. The market signal processing method according to claim 1, characterized in that, The pre-trained classifier is a phase classifier based on the Transformer architecture.

8. A market signal processing method according to claim 1, characterized in that, The market status includes at least two of the following: a demand inflection point that indicates a sudden change in demand for the target product; a plateau period that indicates stable demand for the target product; and a recession period that indicates a continuous contraction in the market for the target product.

9. A market signal processing method according to claim 1, characterized in that, It also includes determining that the current market is in a market state if the probability of responding to a market state is greater than a preset threshold and persists for more than a preset duration.

10. A market signal processing system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a market signal processing method according to any one of claims 1-9.

Images