Rubber market analysis system based on AI intelligent model prediction

Abstract

The invention relates to the technical field of market prediction, in particular to a rubber market analysis system based on AI intelligent model prediction, and the system comprises a bounce recognition module, a fluctuation structure extraction module, a trend filing and classification module, an AI input generation module and a trend prediction module. According to the method, the price jump meeting the minimum change unit is extracted, the micro-level fluctuation characteristics are captured, the recognition precision of the short-time transaction behavior is enhanced, the sequence is constructed according to the jump frequency and direction, the fluctuation structure is recognized in combination with the continuous direction consistency, the coherence expression of the trend segment is improved, and the classical price form is introduced as the training basis; according to the method, the identification capability of the AI model on the non-standard trend is enriched, the current trend and the historical structure are dynamically matched, the form reconstruction of a trend path is realized, the adaptability and the prediction precision to the complex market are enhanced, the structure sensing depth, the direction judgment definition and the prediction logic traceability of the AI model are improved, and the analysis and prediction accuracy of the rubber market is fully ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of market forecasting, and in particular to a rubber market analysis system based on AI intelligent model forecasting. Background Art

[0002] The market forecasting technology field includes methods and tools for trend analysis and future price forecasting of financial products, raw materials, futures and other market trends using historical data and real-time data. The core content includes data collection, data preprocessing, feature extraction and modeling analysis, and is mainly used for market monitoring and trend judgment of securities, futures, foreign exchange, commodities and other trading markets. It realizes the identification and prediction of complex market fluctuations by constructing statistical models and machine learning methods, emphasizing data-driven prediction mechanisms, and often relying on time series analysis, regression analysis, classification analysis and other methods for modeling and reasoning, aiming to provide assistance for investment decision-making, risk control and market analysis.

[0003] Among them, the rubber market analysis system based on AI intelligent model prediction refers to a technical solution that uses artificial intelligence models to predict rubber market price trends, covering the collection of historical transaction data on rubber futures and spot markets, training data preparation based on data cleaning and time series construction, and nonlinear trend extraction through artificial neural network modeling. It also uses rubber-related macroeconomic indicators to supplement sample features, and further combines regression prediction models to predict future rubber price trends. With rubber commodities as the core target, the market trend analysis is completed through model training and prediction result output.

[0004] In the existing rubber market analysis process, the data processing method focuses on statistical mean and rate-of-change aggregation indicators, lacks effective extraction of micro-jump features in high-frequency trading behaviors, and leads to a slow response to directional signals in short periods. In the process of structural identification, single-point features are the core, and the directional evolution trend in continuous time segments is not effectively associated, making it impossible for the model to form a continuous cognition of trend patterns. Sample generation mostly adopts a static window segmentation strategy, and does not incorporate the price structure evolution path, resulting in a single training data form and reducing the robustness of the model in non-standard structural identification. The prediction stage relies on instantaneous indicators of current data and lacks time dimension mapping with historical trend patterns, resulting in limited trend evolution identification, especially in trend shock and mutation scenarios, there are problems of prediction lag and directional drift. For example, in the stage of rapid rise and fall of rubber prices, the model lacks a continuous structural feature matching mechanism, making it difficult to accurately restore the trend change path, which ultimately leads to prediction misjudgment and deviation from the operation strategy. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and propose a rubber market analysis system based on AI intelligent model prediction.

[0006] To achieve the above object, the present invention adopts the following technical solutions: A rubber market analysis system based on AI intelligent model prediction includes: The jump recognition module obtains rubber futures transaction data, determines whether the price difference between two adjacent transactions meets the minimum price change unit and records jump events, divides the jump change band interval, and generates a jump frequency band recognition sequence; The fluctuation structure extraction module, based on the jump frequency band recognition sequence, determines whether there is continuity between consecutive bands, combines the fluctuation structures according to the direction consistency condition, marks the start and end times and direction characteristics, and generates a structure continuity recognition fragment group; The trend filing and classification module combines the structure continuity recognition fragment group, retrieves direction consistency fragments in historical data, groups them according to the structure form to construct a direction response sample set, and marks the subsequent trend direction corresponding to each group of samples, generating a historical trend classification and filing set; The AI input generation module extracts all technical pattern combination patterns according to the historical trend classification and filing set, organizes them into an AI model training data source, and obtains an AI model input data set; The trend prediction module uses the AI model input data set as the training input, uses the current real-time trading data of the rubber market as the inference input, deduces the potential trend direction of the rubber market, and generates a rubber market trend prediction result.

[0007] As a further solution of the present invention, the jump frequency band recognition sequence includes a jump count sequence, a time window distribution record, and a price jump direction. The structure continuity recognition fragment group includes a fluctuation start and end time, a continuous direction feature, and a jump frequency continuity. The historical trend classification and filing set includes a sample grouping label, a corresponding trend direction, and a technical indicator pattern. The AI model input data set includes a technical pattern combination pattern, a model input item data table, and a model label item data table. The rubber market trend prediction result includes a pattern matching degree index, a potential trend direction, and a prediction matching structure.

[0008] As a further solution of the present invention, the jump recognition module includes: The time-price acquisition sub-module obtains the transaction time and transaction price in the rubber futures transaction data, performs data alignment operations after arranging them in sequence, sorts each transaction record in the order of transaction time and pairs it with the previous one to obtain a continuous transaction record sequence; The minimum change judgment sub-module, based on the continuous transaction record sequence, calculates the price difference according to the prices of two adjacent transactions, determines whether the price difference is greater than or equal to the minimum price change unit of the rubber futures, marks the effective price jump records, and obtains an effective price jump mark sequence; The frequency hopping interval division sub-module counts the number of price jumps within a unit time window according to the effective price jump mark sequence, sets a time window and slides it for frequency accumulation, and uses the formula: ; Calculate the jump frequency band intensity value of each time window , and perform continuous interval segmentation on the frequency band intensity to obtain a jump frequency band identification sequence, where represents the th transaction price, represents the th transaction time (in seconds), indicates whether the jump is valid, with a valid jump taking the value of 1 and an invalid jump taking the value of 0, represents the number of records within the current window.

[0009] As a further solution of the present invention, the fluctuation structure extraction module includes: The direction marking sub-module extracts the starting and ending prices within each jump band based on the jump frequency band identification sequence, collects the starting transaction price and the ending transaction price corresponding to each band, calculates the difference between the two and determines the rising and falling directions, marks the change direction of each jump band, and establishes a band direction marking sequence; The direction consistency judgment sub-module, according to the band direction marking sequence, sequentially slides and compares whether the directions are the same based on the direction values of every three consecutive bands in the sequence. If there are three consecutive bands that are all rising or all falling, it is recorded as a direction-consistent segment group, and the formula is used: ; Calculate the direction consistency offset value , and screen the direction-consistent segment groups to obtain a direction continuous jump interval group, where represents the direction identifier of the th band (taking 1 for rising and -1 for falling), represents the jump frequency of the th jump band, represents the time span of the th band, is the total number of sliding combination segments, is the starting index of the current sliding window; The structure section generation sub-module extracts the timestamps corresponding to the first and last bands in each group according to the direction continuous jump interval group, integrates the direction values within the same group, records the start time, end time and overall direction attribute of the combined segment, and generates a structure continuation identification fragment group.

[0010] As a further solution of the present invention, the trend archiving and classification module includes: The similar segment retrieval submodule continuously identifies the time segments corresponding to each segment in the segment group based on the structure, extracts the interval sequence in the same direction from the historical jump frequency band identification sequence, selects the historical segments with the same direction and time span, extracts the jump frequency and direction attributes, and establishes a direction matching segment set; The structural sample construction submodule corresponds to the historical K-line data according to each segment in the direction matching segment set, calls the 5-day exponential moving average, the 20-day exponential moving average and the MACD bar chart data within the corresponding time range, and determines whether the long structure condition (5-day EMA is greater than 20-day EMA and the MACD bar chart expands for 3 consecutive cycles) or the short structure condition (5-day EMA is less than 20-day EMA and the MACD bar chart shrinks for 3 consecutive cycles) is met, and the samples that meet the conditions are classified into corresponding categories to obtain a direction response sample set; The trend archive generation submodule is based on the direction response sample set, collects the K-line trend direction in the subsequent time period of each group of samples, extracts the closing price change trend of continuous periods, marks the rising or falling categories according to the overall trend direction, records the sample structure characteristics and the corresponding trend direction, and establishes a historical trend classification archive set.

[0011] As a further solution of the present invention, the AI ​​input generation module includes: The morphological pattern extraction submodule collects the price trend sequence of each group of archived samples based on the historical trend classification archive set, identifies the K-line structure within the archived time period one by one, determines whether it has standard morphological features such as head and shoulders pattern, triangle consolidation pattern or flag continuation pattern, records the type and position of the pattern, and establishes a morphological combination mark list; The training data sorting submodule collects the price trend, indicator characteristics and morphological structure of the corresponding samples according to the morphological combination tag list, constructs the standard input content field, combines the morphological identification and archived trend category of each sample, generates the corresponding structural label field, and integrates the generation model training data source; The input data construction submodule performs numerical normalization and morphological label encoding processing on each field according to the model training data source, arranges and combines the input fields in a fixed order, outputs a matrix input format, and establishes an AI model input data set.

[0012] As a further solution of the present invention, the trend prediction module includes: The real-time input matching submodule obtains the real-time transaction data of the current rubber market, extracts the continuous price sequence, unifies the format, performs field alignment and sequence length standardization processing with the sample structure in the AI ​​model input data set, and generates a real-time reasoning input set; The morphological similarity recognition sub-module performs sequence comparison between the real-time inference input set and each sample in the AI model input dataset, calculates the temporal difference degree of each group of price patterns using the dynamic time warping method, combines the corresponding index sequence differences, and uses the formula: ; Calculate the morphological matching degree index , select the optimal historical combination according to the minimum matching degree, and obtain the current price structure matching degree sequence, where and are the price values of the current and the historical th points respectively, and are the corresponding technical index values of the current and the historical ones, and are the time step position index values of the current and the historical ones, is the number of points in the matching sequence; The trend direction derivation sub-module selects the historical sample with the highest matching degree value according to the current price structure matching degree sequence, extracts the corresponding trend direction label, counts the proportion of the same labels and combines the concentration degree of the main label to infer the overall trend direction of the current rubber market, and generates the rubber market trend prediction result.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In the present invention, by extracting price jumps that conform to the minimum change unit, capturing the micro-level fluctuation characteristics, enhancing the recognition accuracy of short-term trading behaviors, constructing sequences with jump frequencies and directions, combining the continuous direction consistency to identify the fluctuation structure, improving the coherent expression of trend segments, constructing sample labels based on the structure evolution law, forming multi-dimensional morphological clustering results, enhancing the recognition accuracy of trend characteristics, introducing classic price patterns as the training basis, enriching the recognition ability of the AI model for non-standard trends, realizing the morphological reconstruction of the trend path by dynamically matching the current trend with the historical structure, enhancing the adaptability to complex market conditions and prediction accuracy, and forming a chain-like feature conduction mechanism with multi-level abstract structures, improving the structure perception depth, direction judgment clarity and prediction logic traceability of the AI model, and fully ensuring the accuracy of rubber market analysis and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is the system flow chart of the present invention; Figure 2 is the flow chart of the jump recognition module of the present invention; Figure 3 is the flow chart of the fluctuation structure extraction module of the present invention; Figure 4 is the flow chart of the trend filing and classification module of the present invention; Figure 5 This is the flowchart of the AI input generation module of the present invention; Figure 6 This is the flowchart of the trend prediction module of the present invention. Detailed implementation manners

[0015] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0016] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by terms such as "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. In addition, in the description of the present invention, "a plurality of" means two or more, unless otherwise specifically defined.

[0017] Please refer to Figure 1 , a rubber market analysis system based on AI intelligent model prediction includes: The jump recognition module obtains the transaction time and transaction price in the rubber futures transaction data, and judges whether the price difference between two adjacent transactions meets the minimum price change unit (that is, the minimum price change amount allowed for a single quotation of the rubber futures contract stipulated by the exchange. For example, the rubber futures contract of the Shanghai Futures Exchange is 5 yuan / ton, and the price change needs to reach or exceed this unit to be recorded as an effective jump). If it meets the requirement, it is recorded as a jump event, and the number of jumps within a unit time is accumulated according to the time window, the jump change band interval is divided, and a jump frequency band recognition sequence is generated; Based on the continuous direction performance of each jump band in the jump frequency band recognition sequence, the fluctuation structure extraction module judges whether there is continuity between the unit time jump frequency (a statistical value representing the number of effective jumps within a unit time) and the change direction between continuous bands. If the direction consistency condition is continuously met (when three consecutive jump bands all show the same direction change (such as continuous increase or decrease), it is regarded as meeting the direction consistency determination standard), they are combined into the same group of fluctuation structures, and the start and end times and direction characteristics of the corresponding structures are marked to generate a structure continuity recognition fragment group; The trend archiving and classification module combines the time periods corresponding to all segments in the structure continuation recognition segment group, retrieves segments with consistent directions in historical data, conducts collection and identification, groups them by structural form to construct a set of direction response samples (verified using classical technical indicators: for a long structure to be confirmed, it is required that the 5-day EMA > 20-day EMA and the MACD histogram expands continuously for 3 periods; for a short structure to be confirmed, it is required that the 5-day EMA < 20-day EMA and the MACD histogram contracts continuously for 3 periods), and marks the subsequent trend directions corresponding to each group of samples to generate a historical trend classification and archiving set; The AI input generation module extracts all the technical pattern combination patterns that appear in the sample set (including classic price patterns recognized in the financial market such as head and shoulders patterns, triangle consolidation patterns (ascending / descending triangles), flag continuation patterns (bullish / bearish flags), etc.) from the historical trend classification and archiving set, organizes them into a data source for AI model training, constructs a data table for model input items and label items, and obtains an AI model input data set; The trend prediction module uses the AI model input data set as the training input, and the current real-time trading data of the rubber market as the inference input item, conducts corresponding relationship matching, identifies situations that conform to the historical combination structure (using the dynamic time warping (DTW) algorithm, by comparing the similarity of the current price pattern and the historical pattern in the time series, calculating the pattern matching degree index for deriving the potential trend direction), and derives the potential trend direction of the current rubber market to generate a rubber market trend prediction result.

[0018] The jump frequency band identification sequence includes the jump count sequence, time window distribution record, and price jump direction. The structure continuation recognition segment group includes the fluctuation start and end times, continuous direction characteristics, and jump frequency continuity. The historical trend classification and archiving set includes sample group labels, corresponding trend directions, and technical indicator patterns. The AI model input data set includes technical pattern combination patterns, model input item data tables, and model label item data tables. The rubber market trend prediction result includes the pattern matching degree index, potential trend direction, and predicted matching structure.

[0019] Please refer to Figure 2 , the jump recognition module includes: The time-price acquisition sub-module obtains the transaction time and transaction price in the rubber futures transaction data, performs data alignment operations after arranging them in sequence, sorts each transaction record in the order of transaction time and pairs it with the previous one to obtain a continuous transaction record sequence; To obtain the transaction time and transaction price in the rubber futures transaction data, it is necessary to first access the exchange market interface, collect the real-time transaction data stream, extract data records at a frequency of refreshing once per second. The record format includes the trading timestamp (such as 09:05:15 on April 17, 2024) and the corresponding price (such as 12,850 yuan / ton). After importing the original data stream into the local cache system, it is necessary to sort it in ascending order according to the time field to complete the time alignment operation, ensuring that any two adjacent data records can be paired in position. For example, four data records obtained during a certain trading period are [(09:05:01, 12850), (09:05:03, 12855), (09:05:06, 12850), (09:05:08, 12860)]. After sorting, the adjacent two data records are (09:05:01, 12850)-(09:05:03, 12855), (09:05:03, 12855)-(09:05:06, 12850), (09:05:06, 12850)-(09:05:08, 12860). Then, after pairing, each group forms a two-dimensional record sequence. After forming the sequence, the system calls the time field and price field of each pair of records to establish a complete continuous time series. During the processing, it is necessary to ensure that the format of the time field is uniformly HH:MM:SS to avoid matching exceptions caused by differences in record formats. During the process of generating the continuous transaction record sequence, a record integrity threshold needs to be set. If the continuous missing data in a certain period exceeds 3 seconds, the records in that period will be excluded and not included in the subsequent identification sequence construction. Assuming that 30 groups of complete paired data are detected in an actual period, a continuous transaction record sequence of 30 items will be finally formed.

[0020] Based on the continuous transaction record sequence, the minimum change judgment sub-module calculates the price difference according to the adjacent two transaction prices, judges whether the price difference is greater than or equal to the minimum price change unit of rubber futures, and marks the effective records of price jumps to obtain an effective price jump mark sequence; Based on the continuous transaction record sequence, each group of records contains a time pair and a price pair. Call the adjacent two price data, and perform difference calculations on the price pairs in turn. For example, if the two records are (12,850 yuan / ton, 12,855 yuan / ton), the difference is 5 yuan / ton. The difference calculation formula is , where represents the price of the latter record, represents the price of the previous record. If the difference is greater than or equal to 5 yuan / ton, that is, it meets the requirements of the minimum change unit set by the Shanghai Futures Exchange, mark this record as an effective jump event, otherwise set it as an invalid mark 0. To illustrate the execution details of this process, sample 10 price pairing data for trial calculation, as shown in Table 1: Table 1 Example Table of Jump Judgment

[0021] As shown in Table 1, the price difference is set as the jump effective threshold, and the set standard is 5 yuan / ton. Records that meet or exceed this threshold are set as jump flag 1, and those that do not are set as 0. The basis for setting this threshold is the quotation change unit of the rubber futures contract on the Shanghai Futures Exchange, and the unit value is formulated and publicly announced according to the contract rules standard. In this process, each price record comes from the continuous transaction record sequence, and all judgment results are written into the Boolean sequence to form the effective price jump flag sequence.

[0022] The frequency hopping interval division sub-module counts the number of price jumps within the unit time window according to the effective price jump flag sequence, sets the time window and slides it for frequency accumulation, using the formula: ; Calculate the jump frequency band intensity value of each time window , and perform continuous interval segmentation on the frequency band intensity to obtain the jump frequency band identification sequence, where, represents the th transaction price, represents the th transaction time (in seconds), represents whether this jump is effective. The value for an effective jump is 1, and the value for an ineffective jump is 0, represents the number of records in the current window; According to the effective price jump flag sequence, embed it into a sliding time window in minutes for accumulation. The initial set time window width is 60 seconds, and the step size is 10 seconds. Take 4 consecutive groups of records in the actual monitoring data. The price differences are 5, 5, 10, 0 (unit: yuan / ton) in sequence, and the time differences are 2s, 3s, 2s, 1s in sequence. The corresponding jump flags are 1, 1, 1, 0. Substitute them into the formula for intensity value calculation: ; If there are Paragraphs with values greater than 10 are identified as the jump high-frequency band. This value of 10 is the threshold for judging the intensity of the jump frequency band. The setting basis is to observe the cumulative frequency distribution interval of jumps in the typical period of the active trading days of Shanghai rubber. The occurrence times and average jump amplitudes of price jumps exceeding 5 yuan / ton in each jump section are statistically analyzed. After summarizing the number of jumps per minute through a time window, the average intensity value is calculated to be 9.3, and the maximum value is 13.7. Therefore, setting the threshold to be greater than 10 is more representative, which can effectively eliminate ordinary jump segments and avoid misidentifying normal small-intraday fluctuation segments. Moreover, this value will fluctuate with the increase in contract volatility or the change in the concentration of trading periods. In high-trading-density periods (such as 9:00 - 9:30 in the morning session), the upper limit of this value can reach above 12, and it is often lower than 9 after 14:30 in the afternoon session. Therefore, it is reasonable to conservatively define the critical intensity as 10. The system will form a time series of values under all sliding windows. If there are paragraphs with values greater than 10 in consecutive windows, they are identified as the jump high-frequency band, and finally a jump frequency band identification sequence is generated.

[0023] The operation logic of this formula is to comprehensively convert multiple effective price jump events within a unit time into a value reflecting the jump intensity. The composition of each item is the product of the jump amplitude and the jump validity mark, divided by the square root of the time interval between adjacent transaction records. Specifically, the jump amplitude reflects the absolute degree of price change between two adjacent transactions. If the paired record is a non-effective jump, its influence on the result is eliminated through to avoid the interference of invalid jumps. The square-rooted time difference is used for time-weight compression, amplifying the importance of intensive jump behaviors occurring within a short time. So that the same jump amplitude, if continuously occurring within a very short interval, will produce a higher intensity reflection, thereby enhancing the recognition degree of time-intensive jumps in the overall measurement. The results of all effective jump items are accumulated, representing the total intensity of all jump events within this time window. Finally, the absolute value processing is used to ensure the direction consistency and avoid the cancellation of the cumulative effect due to positive and negative direction changes. Therefore, this formula can present a higher intensity value when jump behaviors exist, which helps to quickly identify active fluctuation regions in the entire price sequence.

[0024] Please refer to Figure 3 , the fluctuation structure extraction module includes: Based on the jump frequency band identification sequence, the direction marking sub-module extracts the starting and ending prices within each jump band, collects the starting transaction price and ending transaction price corresponding to each band, calculates the difference between the two and determines the direction as rising or falling, marks the change direction of each jump band, and establishes a band direction marking sequence; Based on the hopping frequency band identification sequence, extract the starting and ending prices within each hopping band. After collecting the corresponding starting transaction price and ending transaction price for each band, the sequence needs to be sequentially partitioned. Each group of band data is disassembled segment by segment in the order of time sequence. Set the time span of each hopping band not to exceed 20 seconds, and collect the transaction price at the starting moment. and the transaction price at the ending moment , and then perform direction determination. That is, when , it is recorded as the upward direction, otherwise it is the downward direction. If it is 0, it is classified as no direction. Here, it should be noted that the transaction price should be from the actual rubber futures trading data. For example, at 10:01:00 on a trading day on April 15, 2024, the price is 12,435 yuan / ton, and at 10:01:15, the price is 12,460 yuan / ton, then the difference is 25 yuan, and the direction is marked as upward. Conversely, if it is 12,415 yuan / ton, the direction is downward; Subsequently, number the directions of all bands, use positive and negative 1 to mark the upward and downward directions respectively, and construct a direction array {1, -1, 1, 0, 1, -1}. For each direction identifier, it should be ensured that it is obtained based on the starting and ending prices within a clear time interval and can be repeatedly calculated and verified. In the direction marking step, it is not allowed to use the hopping frequency as the basis for direction, and the direction calibration should be independently completed relying on price changes. And the band division results need to be numbered in the system and correspond one by one with the starting and ending times. The direction array is used as the basis for the output direction judgment, and finally a band direction marking sequence is generated.

[0025] The direction consistency judgment sub-module, according to the band direction marking sequence, successively slides and compares whether the directions are the same based on the direction values of every three consecutive bands in the sequence. If there are three consecutive bands that are all upward or all downward, it is recorded as a direction-consistent segment group, using the formula: ; Calculate the direction consistency offset value , and screen the direction-consistent segment groups to obtain the direction continuous hopping interval groups. Among them, represents the direction identifier (take 1 for upward and -1 for downward) of the th band, represents the hopping frequency of the th hopping band, represents the time span of the th band, is the total number of sliding combinations, is the starting index of the current sliding window; According to the band direction marker sequence, the window is slid sequentially to check the direction continuity. The judgment window is set as a moving window with a length of 3. Starting from the first segment, the sequence is judged in groups of three. For example, if the sequence is {1, 1, 1, -1, -1, -1, 1}, the first group is {1, 1, 1} with consistent direction, and the second group is {1, 1, -1} with inconsistent direction, and so on. The judgment criterion is whether the product result of the three direction values is 1 or -1, that is, whether they are all positive or all negative. On this basis, to enhance the judgment accuracy, a direction consistency offset value is introduced, and the following data input conditions are set: Table 2 Jumping Band Sample Data Table

[0026] As shown in Table 2, assuming the direction is positive, the offset value is calculated.

[0027] Direction identification sequence:; Jumping frequency sequence: ; Time span: ; The formula is: ; The calculation process is as follows: , , ; ; Denominator ; Numerator ; The first part ; The second part ; Then: ; This value represents that the current direction offset is 0, indicating that the direction is completely consistent. Then, these three segments of the band form a direction-consistent group. If Set to offset overlimit, which is set based on the coupling relationship between directional consistency and the degree of fluctuation of beating frequency. When the band beating frequency fluctuates between 0.3 and 1.0 times per second and the direction remains continuous, the offset value is usually less than 0.2. Therefore, setting 0.3 as the critical value can accommodate the situation of slight changes in directional identification or local frequency disturbances in the short term, and avoid misjudging the directional offset due to instantaneous fluctuations. This value will increase with the increase of the absolute amplitude of the beating frequency. For example, when the beating frequency reaches more than 1.5 times per second and the time span between bands is shortened to less than 5 seconds, the offset value can increase to between 0.35 and 0.45. Therefore, in the trading range where the beating frequency range is mainly distributed between 0.5 and 1.2 times per second and the time span is between 8 and 15 seconds, setting the offset threshold to 0.3 can cover more than 80% of the characteristic intervals of the conventional directional consistent segment group, thereby providing a reliable numerical judgment basis for directional continuity. The current value has not exceeded the limit and meets the consistency standard, and finally a directional continuous beating interval group is obtained.

[0028] The operational logic of this formula is to measure the degree of directional consistency deviation between consecutive beating bands by constructing two parts, the first part of which is ; It is used to measure the relative relationship between the overall direction change amplitude and the beating frequency density, that is, by calculating the absolute value of the difference between the first and last direction marks to reflect the degree of direction change, and at the same time introduces the square sum of the ratio of three consecutive beating frequencies to the time span, and adds 1 to it and takes the square root as the normalization factor. Its purpose is to reduce the impact of small direction changes in the case of dense frequency, thereby suppressing the discontinuity caused by short-term violent beating. The second part ; It is used to separately reflect the degree of deviation between the direction of the middle band and the direction of the previous band, and reflects the intensity of the directional jump per unit time through time span normalization. The addition of these two parts forms a comprehensive characterization of the directional consistency deviation of three consecutive jumping bands in both overall and local dimensions, ensuring that the formula takes into account the interference of local disturbances on directional judgment while statistically analyzing the continuous trend, thereby improving the stability and rationality of structural judgment.

[0029] The structural segment generation submodule extracts the timestamps corresponding to the first and last segments in each group according to the direction continuous jump interval group, integrates the direction values ​​in the same group, records the start time, end time and overall direction attributes of the combined segment, and generates a structural continuation identification segment group; According to the group of continuously jumping intervals in a certain direction, extract the time tags of the starting band and the ending band of each group, and collect the direction identification values within the group. For example, if a certain group is from band 1 to band 3, the starting time is 10:00:00, the ending time is 10:00:27, and the direction identifications are all 1, then the integrated result is recorded as the structural segment {10:00:00, 10:00:27, rising}. This step requires summarizing and extracting all the band numbers within the group, and uniformly processing the time fields to avoid deviations caused by differences in time granularity of data from different sources. At the same time, it is necessary to verify whether the ending time is later than the starting time. If there is an error, exclude this segment. Subsequently, establish a structural segment record table and sort and output it in chronological order, and finally generate a group of structural continuation recognition segments.

[0030] Please refer to Figure 4 , the trend filing and classification module includes: Based on the time intervals corresponding to each segment in the group of structural continuation recognition segments, extract the interval sequences in the same direction from the historical jump frequency band recognition sequence, screen the historical segments with the same direction and the same time span, extract the jump frequency and direction attributes, and establish a set of direction-matching segments; Based on the time intervals corresponding to each segment in the group of structural continuation recognition segments, extract the starting time and ending time of the segment. For example, if a certain segment starts at 09:00 on July 1, 2024 and ends at 10:00 on July 1, 2024, then determine that the time interval of this segment is 1 hour. Subsequently, gradually retrieve the time intervals with the same direction attribute as this segment in the historical jump frequency band recognition sequence, and extract all the segments in the historical sequence that meet the same direction identification (such as taking 1 for the rising direction and -1 for the falling direction). For each segment, it is necessary to confirm whether its time length is close to 1 hour. If the time span of the actual historical segment is between 45 minutes and 75 minutes, it is regarded as a time-similar segment. Then extract its jump frequency within the unit time. For example, if the number of jumps within a certain segment is 120 times, then the frequency is 2 times / minute. Set the similarity threshold to ±0.5 times / minute, and judge whether the segment meets the frequency similarity requirement. If the frequency of a historical segment is 1.8 times / minute, then it is judged as a matching segment. Perform this screening operation for each historical segment, and record all the historical segments that simultaneously meet the requirements of the same direction, close time span, and matching jump frequency as the same-direction samples. Finally, establish a set of direction-matching segments.

[0031] The structural sample construction sub-module matches each segment in the direction matching segment set with the historical K-line data, and calls the 5-day exponential moving average, 20-day exponential moving average, and MACD histogram data within the corresponding time range to determine whether the long structure condition (5-day EMA is greater than 20-day EMA and the MACD histogram expands continuously for 3 periods) or the short structure condition (5-day EMA is less than 20-day EMA and the MACD histogram contracts continuously for 3 periods) is satisfied. The samples that meet the conditions are classified into the corresponding categories to obtain the direction response sample set; For each historical segment in the direction matching segment set, extract the K-line data sequence within its corresponding time period, count the closing prices within 5 trading days and 20 trading days, and use the exponentially weighted average method to calculate the 5-day EMA and 20-day EMA respectively to further compare the short-term and long-term trend directions. For example, if the closing prices from July 3 to July 7, 2023 are [13000, 13050, 13100, 13080, 13120], the 5-day EMA can be calculated by setting the weighting factor to 0.33. The initial EMA is set to 13000, and the subsequent EMA calculations are 13016.5, 13045.055, 13057.79, 13078.12 in turn. Then, the 5-day EMA value on the 5th day is obtained as 13078.12; calculate the 20-day EMA in the same way and obtain the corresponding MACD histogram value. By comparing, determine whether the height of the MACD histogram shows an expanding trend (such as the column values are 30, 35, 42 in turn) or a contracting trend (such as -30, -25, -20 in turn) within 3 consecutive periods. If 5-day EMA is greater than 20-day EMA and the MACD histogram expands, it is marked as a long structure. On the contrary, if 5-day EMA is less than 20-day EMA and the MACD histogram contracts, it is marked as a short structure. Table 3 shows some sample structure recognition situations, and finally, all samples that meet the structure recognition criteria are grouped into the direction response sample set; Table 3 Sample Structure Recognition Result Table

[0032] As shown in Table 3, multiple samples are classified and recognized according to the combination of EMA and MACD indicators, thus constructing the direction response sample set.

[0033] The trend filing and generating sub-module, based on the direction response sample set, collects the K-line trend directions in the subsequent time period of each group of samples, extracts the changing trend of the closing prices in consecutive periods, marks the rising or falling categories according to the overall trend direction, records the sample structure characteristics and the corresponding trend directions, and establishes the historical trend classification and filing set; According to each group of structure identification samples in the response sample set, read the K-line trend data that continues after the structure identification, and extract the closing price sequence of the 10 trading days after the end point of the sample structure. For example, the closing price of a sample is [13100, 13150, 13200, 13280, 13350, 13380, 13360, 13390, 13420, 13460], and the difference in price changes in consecutive cycles is calculated to be [50, 50, 80, 70, 30, -20, 30, 30, 40]. It is judged that the trend direction is continuously rising, so the subsequent trend direction of the sample is rising. And mark it as an upward trend for archiving. If the subsequent closing price of another sample is [12800, 12760, 12720, 12650, 12600, 12580, 12550, 12500, 12480, 12400], the price change sequence is [-40, -40, -70, -50, -20, -30, -50, -20, -80], which is judged as a continuous downward trend and archived as a downward trend category. Repeat the above judgment logic, process all directional response sample sets in sequence, and finally mark the structural characteristics of each group of samples and the corresponding subsequent trend direction, and summarize and generate a historical trend classification archive set.

[0034] See also Figure 5 , the AI ​​input generation module includes: The morphological pattern extraction submodule collects the price trend sequence of each group of archived samples based on the historical trend classification archive set, identifies the K-line structure within the archived time period one by one, determines whether it has standard morphological features such as head and shoulders pattern, triangle consolidation pattern or flag continuation pattern, records the type and position of the pattern, and establishes a morphological combination mark list; Based on the historical trend classification archive collection, the marked K-line trend segments in each archived sample are extracted. When performing morphological recognition on each segment, the corresponding K-line chart structure must be collected first. The structure contains the opening price, closing price, highest price and lowest price of each trading day. Suppose a sample segment is from July 1, 2024 to July 20, 2024, and its K-line data is collected as follows: the opening price array is {12250, 12300, 12400, 12550, 12600, 12680}, and the closing price array is {12250, 12300, 12400, 12550, 12600, 12680}. The highest price array is {12300, 12400, 12500, 12650, 12620, 12500}, the highest price array is {12350, 12480, 12560, 12700, 12730, 12600}, and the lowest price array is {12220, 12280, 12360, 12500, 12550, 12400}. Combine these data to determine whether they constitute a specific technical pattern; for example, when the middle high point is significantly higher than the left and right high points, If the middle low point is also higher than the low points on both sides, it is judged as a head and shoulders top pattern. If the data shows a continuous rise and then enters a convergence state, forming a structure in which the high points continue to decrease while the low points remain consistent, it is judged as a descending triangle pattern. During the execution process, the high points and low points are positioned and numerically calculated, where the high points are the local maximum points and the low points are the local minimum points. If the distance difference between the high points and the low points exceeds 50 yuan / ton, it is considered to constitute a significant turning point. If the difference between the consecutive high points is less than 20 yuan / ton, the pattern candidate is excluded. Further, To determine the validity of the formation pattern, it is necessary to observe whether the pattern structure is completed within 7 trading days within the cycle; for example, in the sample data, from the 3rd to the 9th day, the price rose from 12,400 yuan / ton to 12,650 yuan / ton, then adjusted back to 12,500 yuan / ton, and fell to 12,300 yuan / ton on the 10th day, forming a complete head and shoulders top structure. This structure will be recorded as "head and shoulders top", with its start time being July 3, 2024, and end time being July 10, and its pattern type will be marked; the data in the table are as follows: Table 4 Technical form identification sample table

[0035] As shown in Table 4, the price difference and cycle are the basic criteria for pattern recognition, where the high point difference is 130 yuan / ton, the low point difference is 80 yuan / ton, and the cycle is 8 days, which meets the judgment conditions. This type of structure recognition needs to traverse each group of archived sample segments, extract all possible price combinations that may constitute the pattern through a sliding window, execute the judgment rules, record all pattern types, start and end times, sample numbers, etc. that meet the structure, and finally establish a pattern combination mark list.

[0036] The training data collation sub-module collects the price trends, indicator features, and morphological structures of corresponding samples according to the morphological combination label list, constructs the standard input content fields, combines the morphological identifiers and archived trend categories of each sample, generates the corresponding structural label fields, and integrates them to generate the data source for model training; Combined with the morphological combination label list, the recognition results of each structural type are linked to their corresponding archived samples. First, the price trends and indicator data in the sample archive are extracted, specifically including the 5-day EMA, 20-day EMA, MACD histogram, KDJ indicator, etc. Combined with the previously marked morphological types, they jointly constitute the feature items of the model. Suppose the technical indicator data of a certain archived segment in the sample is as follows: the 5-day EMA sequence is {12340, 12380, 12450, 12510, 12540}, the 20-day EMA sequence is {12290, 12330, 12370, 12420, 12450}, and the MACD histogram value is {20, 25, 30, 28, 22}. From these data, it can be seen that from the 1st day to the 3rd day of this sample, the 5-day EMA is continuously greater than the 20-day EMA, and the MACD is continuously increasing, meeting the long structure standard. As a positive sample, it is entered. At the same time, its morphological recognition result is "ascending triangle", then its structural morphological type, technical indicator features, and archived trend category (rising) constitute a complete training sample entry; in each field, the feature item contains the standardized technical indicator sequence. For example, the standardized interval of the difference between the 5-day EMA and the 20-day EMA is set to [-1, 1], and linear normalization processing with the minimum value as the lower bound and the maximum value as the upper bound is adopted, that is , where is the current value, and are the minimum and maximum values of this field in the sample set respectively. In the above example, the maximum value of the 5-day EMA is 12540, and the minimum value is 12340, so the normalized value on the 3rd day is ; the corresponding label item uses an encoding method, marking "rising" as 1, "falling" as 0, and "oscillating" as -1. Finally, it is sorted and generated in field format to generate the data source for model training.

[0037] The input data construction sub-module performs numerical normalization and morphological label encoding processing on each field according to the data source for model training, arranges and combines the input fields in a fixed order, outputs a matrix input format, and establishes the input data set for the AI model; Based on each piece of sample data in the model training data source, construct an input matrix for its field items. First, perform unified length processing on the technical indicator sequences of each sample. Set the unified input sequence length to 10. If the sample has less than 10 data points, fill it up with the nearest value. If the sample has more than 10 data points, take the last 10 items as the input segment. At the same time, numerically process the morphological category using one-hot encoding. If a sample corresponds to the morphological pattern of "head and shoulders top", its morphological encoding is [1, 0, 0, 0] (corresponding morphological sequence is "head and shoulders top, ascending triangle, descending triangle, flag"). This encoding is concatenated with the technical indicators to form the final input vector. At the same time, its corresponding trend direction label is "falling", with a value of 0, which is output to the label column. Construct the input matrix as a two-dimensional array and the label set as a one-dimensional array. If the total number of samples is 300, the dimension of the input matrix is 300×Z (Z is the concatenation length of each input), and the dimension of the label set is 300×1. During the construction of the input data table, store the input matrix in CSV format for subsequent calls, and divide it into training set, validation set, and test set in a ratio of 6:2:2. At the same time, each sample retains its attribution number and archiving identifier for traceability. Finally, obtain the AI model input data set.

[0038] Please refer to Figure 6 , the trend prediction module includes: The real-time input matching sub-module obtains the real-time trading data of the current rubber market, extracts the continuous price sequence, and after unifying the format, performs field alignment and sequence length specification processing with the sample structure in the AI model input data set to generate a real-time inference input set; To obtain the real-time trading data of the current rubber market, first, it is necessary to clarify the scope of the collection object, including transaction price, transaction time, technical indicators, etc. Taking the one-minute K-line as the sampling benchmark, obtain the opening price, highest price, lowest price, closing price, and trading volume of each minute one by one, and perform real-time calculations in combination with indicators such as MACD and EMA. For example, collect the rubber futures data from 9:00 to 9:30 on April 10, 2025, with a total of 30 sampling points. In each minute's record, assume the opening price is 12835, the closing price is 12860, EMA5 is 12840, and EMA20 is 12820. Then this time period is an upward interval. After collection, data normalization processing is required to standardize each indicator to the interval [0, 1] to avoid model training deviation caused by different numerical scales. The normalization adopts the minimum-maximum normalization method. If the highest price in the current price sequence is 12950 and the lowest price is 12780, then 12860 is normalized to: ; After each indicator is converted accordingly, a normalized data array is constructed. At the same time, a structural comparison is made with the historical sample data fields in the AI model input dataset. The comparison fields must ensure the inclusion of "price sequence", "technical indicator sequence", and "label field". The order of these three types of fields is unified, and the data dimensions are consistent. Among them, the label field is used to correspond to the historical trend direction. For example, if the label of a certain historical sample is 1 (rising), then the current input sequence also needs to be set with a matching label. Finally, it is integrated into a two-dimensional input matrix. The column fields are such as price, EMA5, EMA20, MACD, etc., and the row fields are time nodes. Alignment processing is completed through the array structure. For example, if the historical time node K of a certain sample is 30, then the current sequence also needs to match the 30 nodes. If it is insufficient, it is filled with a moving window. If it exceeds, it is truncated to the 30 nodes. Finally, a real-time inference input set is constructed.

[0039] The morphological similarity recognition sub-module compares the sequences of each sample in the real-time inference input set with those in the AI model input dataset. The dynamic time warping method is used to calculate the temporal difference degree of each group of price morphologies. Combining the corresponding indicator sequence differences, the formula is used: ; Calculate the morphological matching degree index , and select the optimal historical combination according to the minimum matching degree to obtain the current price structure matching degree sequence. Among them, and are respectively the price values of the current and the historical th point, and are the corresponding technical indicator values of the current and the historical, and are the time step position index values of the current and the historical, is the number of points in the matching sequence; When comparing the morphologies according to the real-time inference input set and the samples in the AI model input dataset, it is necessary to traverse the historical sample sequences one by one and perform the dynamic time warping (DTW) matching process. First, it is clear that each comparison structure contains the current sequence and a historical sample sequence, and the length of each sequence is set to , that is, the price and technical indicator data of 30 time points. During the execution process, the morphological difference needs to be calculated for each corresponding point, and the formula is used for calculation.

[0040] Among them, is the closing price of the th time point of the current sequence, is the price of the corresponding time point of the historical sample, , are respectively the technical indicator values (such as MACD or EMA5) of the current and the historical corresponding points, , respectively represent the position index at that time point. For example, at the minute is , a total of 30 sets of complete matching items need to be constructed during the calculation process. Partial matching examples are given in the table as follows: Table 5 Data Table for Calculating Morphological Matching Degree

[0041] Referring to Table 5, the maximum difference between the current price and the historical price is 5 yuan, and the minimum is 2 yuan. If the corresponding index differences are 3.2, 2.1, and 1.8, and the time differences are all 0, the corresponding calculations are as follows: ; Compare this matching value with the set matching degree reference value . The setting basis is to statistically analyze 2000 groups of past samples. Among them, the matching values of more than 70% of the samples with consistent trends are lower than 6.0. Therefore, 6.0 is selected as the boundary value. This reference value approaches the critical state when both the P difference and the I difference are less than 15% and the time difference is less than 3 points. The matching degree value has a positive correlation with the price fluctuation range in the historical samples and the current index fluctuation rate. When the historical price fluctuation range exceeds ±150 points or the absolute value of the continuous change of MACD exceeds 0.5, the matching value will rapidly rise beyond this threshold. Therefore, it has an actual quantitative judgment basis, not obtained by empirical judgment. If it is less than the threshold, it is considered a match. The matching threshold is set with reference to the historical morphological deviation. According to the distribution of 95% of the matching values in the past 500 groups of historical samples, it is set to 6.0. The innovation of the formula lies in integrating the price and the index difference, and adjusting the weights through the time difference term normalization, taking into account both the structural displacement difference and the amplitude error. The obtained morphological matching degree index value will be used for subsequent trend judgment.

[0042] The operation logic of the above formula aims to comprehensively measure the morphological similarity between the current price structure and the historical samples in multiple dimensions. Among them, the parameter represents the price difference between the current and the historical at the th time point. The parameter represents the difference in technical indicators (such as MACD, EMA, etc.). Adding the two is to comprehensively consider the offsets in the two dimensions of price and technical indicators, reflecting the overall matching degree of the morphological structure; the denominator part adopts the form of to regularize the time position difference. Among them, the square term amplifies the time step difference, and the square root structure non-linearly compresses it to a reasonable range, ensuring that this term approaches 1 when the time alignment degree is good, and when the time difference expands, its penalty effect on the overall matching degree intensifies; the core of the entire expression is to construct a matching cost function through the price difference and the index difference, and perform normalization adjustment through the time alignment degree, so that the finally obtained average matching degree It can not only reflect the numerical differences but also embody the consistency of the time series structure, achieving a comprehensive judgment of the morphological structure matching.

[0043] The trend direction derivation sub-module screens the historical samples with the highest matching degree values according to the current price structure matching degree sequence, extracts the corresponding trend direction labels, counts the proportion of the same labels and combines the concentration degree of the main labels to infer the overall trend direction of the current rubber market and generate the prediction result of the rubber market trend; According to the current price structure matching degree sequence, the first 5 groups of historical samples with the lowest matching degree are screened out, their corresponding trend labels are recorded, the frequencies of each label are counted and the concentration degree of the main label is determined. For example, if the labels of 5 samples are up, up, up, down, up, then the proportion of up is 80%. The concentration threshold is set at 70%. The basis for setting this concentration threshold is that if the proportion of a certain trend label in the current sample combination exceeds 70%, then this trend has repeatability and representativeness in the historical samples. Looking back at the labeled data in 1000 groups of model samples, it is found that when the proportion of a certain trend direction exceeds 70%, the continuation rate of its trend within the subsequent 30 minutes reaches 83%. Therefore, 70% is set as the lower limit for dividing the trend stability signal. Its numerical change is mainly controlled by the number of matching groups K and the difference in sample labels of each group. If K increases and the label distribution tends to be uniform, the concentration degree decreases. If the samples tend to be consistent, the concentration degree increases. It is judged that the trend is consistent and the trend direction is up. Finally, the current market trend direction is derived. When counting, the labels are encoded, up is 1 and down is 0. The weighted voting method is used to output the final label. Let the current label set be and the weights are all 1, then the final value is , which is higher than the concentration threshold of 0.7. Therefore, the prediction result of the rubber market trend is generated as up. This numerical result shows that the current price pattern is highly consistent with the historical up pattern, and the prediction result has sample support basis, providing a decision-making direction for the output of the next stage AI model.

[0044] The above is only the preferred embodiment of the present invention and does not limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still belong to the protection scope of the technical solution of the present invention.

Claims

1. A rubber market analysis system based on AI intelligent model prediction, characterized in that: The system comprises: The jump identification module obtains rubber futures transaction data, determines whether the price difference between two adjacent transactions meets the minimum price change unit, records the jump event, divides the jump change band interval, and generates the jump frequency band identification sequence; The wave structure extraction module determines whether there is continuity between consecutive wave bands based on the hopping frequency band identification sequence, combines the wave structure according to the direction consistency condition, marks the start and end time and direction characteristics, and generates a structure continuity identification segment group; The trend archiving and classification module combines the structural continuation identification fragment group, retrieves the direction consistency fragments in the historical data, constructs the direction response sample set by grouping according to the structural form, and marks each group of samples corresponding to the subsequent trend direction, and generates the historical trend classification archive set; The AI ​​input generation module classifies and archives the historical trends, extracts all technical form combination patterns, and organizes them into AI model training data sources to obtain an AI model input data set; The trend prediction module uses the AI ​​model input data set as training input and the current real-time trading data of the rubber market as reasoning input to deduce the potential trend direction of the rubber market and generate a rubber market trend forecast result.

2. The rubber market analysis system based on AI intelligent model prediction according to claim 1 is characterized in that: The jump frequency band identification sequence includes the jump number sequence, time window distribution record, and price jump direction; the structure continuation identification segment group includes the fluctuation start and end time, continuous direction characteristics, and jump frequency continuity; the historical trend classification archive set includes sample grouping labels, corresponding trend directions, and technical indicator forms; the AI ​​model input data set includes a technical form combination pattern, a model input item data table, and a model label item data table; the rubber market trend forecast result includes a form matching index, a potential trend direction, and a predicted matching structure.

3. The rubber market analysis system based on AI intelligent model prediction according to claim 1 is characterized in that: The beat recognition module comprises: The time and price collection submodule obtains the transaction time and transaction price in the rubber futures transaction data, arranges them in order, and then performs data alignment operation. Each transaction record is sorted in order of transaction time and paired with the previous transaction to obtain a continuous transaction record sequence; The minimum change judgment submodule calculates the price difference based on the continuous transaction record sequence and the prices of two adjacent transactions, determines whether the price difference is greater than or equal to the minimum price change unit of rubber futures, marks the price jump valid record, and obtains the valid price jump mark sequence; The frequency hopping interval division submodule counts the number of price jumps within a unit time window according to the valid price jump mark sequence, sets the time window and slides to accumulate the frequency, using the formula: ； Calculate the hopping frequency band strength value for each time window , and divide the frequency band intensity into continuous intervals to obtain the hopping frequency band identification sequence, where Indicates The transaction price, Indicates Transaction time, Indicates whether the jump is valid. A valid jump has a value of 1, and an invalid jump has a value of 0. Indicates the number of records in the current window.

4. The rubber market analysis system based on AI intelligent model prediction according to claim 1 is characterized in that: The wave structure extraction module comprises: The direction marking submodule extracts the start and end prices in each hopping band based on the hopping band identification sequence, collects the starting transaction price and the ending transaction price corresponding to each band, calculates the difference between the two and determines the rising and falling directions, marks the change direction of each hopping band, and establishes a band direction marking sequence; The direction consistency judgment submodule compares the direction of each three consecutive bands in the sequence according to the band direction mark sequence. If there are three consecutive bands that are all rising or falling, they are recorded as a group of segments with consistent direction, using the formula: ； Calculate the direction consistency offset value , and filter the direction-consistent segment group to obtain the direction-continuous jump interval group, among which, Indicates The direction of the band, Indicates The beating frequency of the band, Indicates The time span of the band, is the total number of sliding combination segments, is the starting index of the current sliding window; The structural segment generation submodule extracts the timestamps corresponding to the first and last bands in each group according to the directional continuous jump interval group, integrates the direction values ​​in the same group, records the start time, end time and overall direction attributes of the combined segment, and generates a structural continuation identification segment group.

5. The rubber market analysis system based on AI intelligent model prediction according to claim 1 is characterized in that: The trend archiving classification module includes: The similar segment retrieval submodule continuously identifies the time segments corresponding to each segment in the segment group based on the structure, extracts the interval sequence in the same direction from the historical jump frequency band identification sequence, selects the historical segments with the same direction and time span, extracts the jump frequency and direction attributes, and establishes a direction matching segment set; The structural sample construction submodule determines whether each segment in the direction matching segment set corresponds to the historical K-line data, and whether the long structure condition or the short structure condition is met, and classifies the samples that meet the conditions into the corresponding categories to obtain the direction response sample set; The trend archive generation submodule is based on the direction response sample set, collects the K-line trend direction in the subsequent time period of each group of samples, extracts the closing price change trend of continuous periods, marks the rising or falling categories according to the overall trend direction, records the sample structure characteristics and the corresponding trend direction, and establishes a historical trend classification archive set.

6. The rubber market analysis system based on AI intelligent model prediction according to claim 1 is characterized in that: The AI ​​input generation module includes: The morphological pattern extraction submodule collects the price trend sequence of each group of archived samples based on the historical trend classification archive set, identifies the K-line structure within the archived time period one by one, determines whether it has standard morphological features such as head and shoulders pattern, triangle consolidation pattern or flag continuation pattern, records the type and position of the pattern, and establishes a morphological combination mark list; The training data sorting submodule collects the price trend, indicator characteristics and morphological structure of the corresponding samples according to the morphological combination tag list, constructs the standard input content field, combines the morphological identification and archived trend category of each sample, generates the corresponding structural label field, and integrates the generation model training data source; The input data construction submodule performs numerical normalization and morphological label encoding processing on each field according to the model training data source, arranges and combines the input fields in a fixed order, outputs a matrix input format, and establishes an AI model input data set.

7. The rubber market analysis system based on AI intelligent model prediction according to claim 1 is characterized in that: The trend prediction module includes: The real-time input matching submodule obtains the real-time transaction data of the current rubber market, extracts the continuous price sequence, unifies the format, performs field alignment and sequence length standardization processing with the sample structure in the AI ​​model input data set, and generates a real-time reasoning input set; The morphology similarity recognition submodule performs sequence comparison based on the real-time reasoning input set and each sample in the AI ​​model input data set, and uses the dynamic time warping method to calculate the time series difference of each group of price patterns, and combines the corresponding indicator sequence differences to adopt the formula: ； Calculate the morphological matching index , select the optimal historical combination according to the minimum matching degree, and obtain the current price structure matching degree sequence, where, and The current and historical The price value of a point, and is the current and historical corresponding technical indicator value, and is the index value of the current and historical time step positions, is the number of points in the matching sequence; The trend direction derivation submodule screens the historical samples with the highest matching value according to the current price structure matching degree sequence, extracts the corresponding trend direction labels, counts the proportion of the same labels and combines the concentration of the main labels to infer the overall trend direction of the current rubber market and generate the rubber market trend forecast results.

Images