On-chain capital information filtering and approach recognition method based on multi-dimensional filtering

Abstract

This invention relates to the field of on-chain fund identification, and more particularly to an on-chain fund information filtering and entry identification method based on multi-dimensional filtering. The method includes an information filtering method and an entry identification method. The filtering method includes steps such as multi-time granularity modeling of on-chain fund data, establishing a fund record set, extracting multi-dimensional fund behavior features, and multi-period trend filtering. The entry identification method obtains the entry identification signal based on the information obtained from the above filtering and through five dimensions: amount dimension, peak dimension, frequency dimension, trend dimension, and continuity dimension.

Description

Specification 1 A Method for Filtering and Identifying On-Chain Funds Information Based on Multi-Dimensional Filtering Technical Field This invention relates to the field of on-chain funds identification, and particularly to a method for filtering and identifying on-chain funds information based on multi-dimensional filtering. Background Art In the on-chain market of crypto assets, timely identification of on-chain funds entering a particular token is a core function of quantitative investment research and intelligent early warning systems. Its core purpose is to monitor the dynamics of on-chain funds, such as monitoring large transactions or abnormal transactions. The aim is to understand the dynamics of on-chain funds based on these special or abnormal operations, thereby providing a reference for determining whether to enter the market. In existing technologies, some reference information is provided to assist in judging entry signals. For example, when the net on-chain inflow of a token within a window exceeds a set threshold, the system triggers an entry signal and pushes a notification to the user. However, this information is too one-sided, and the signal judgment is too arbitrary. Summary of the Invention The purpose of this invention is to provide a method for filtering and identifying on-chain funds information based on multi-dimensional filtering to solve the technical problem of large signal judgment errors caused by incomplete and one-sided information in the background technology. To achieve the above objectives, this invention provides the following technical solution: a multi-dimensional filtering method for filtering on-chain funds information and identifying entry points, comprising: S1: multi-time granularity modeling of on-chain funds data; HK 20137911 A Specification 2; S2: based on the fund record set established in step S1, record purification is performed before feature extraction to obtain a purified fund record set; S3: multi-dimensional fund behavior feature extraction, based on the purified fund record set obtained in step S2, fund behavior features are extracted from three dimensions: amount, peak value, and frequency, and differentiated thresholds are set in conjunction with the token fund volume level; S4: multi-period trend filter; S5: on-chain behavior sequence continuity scoring. By adopting the above technical solution, based on multi-dimensional fund feature extraction, information can be more accurate and cover more dimensions of fund dynamic changes, avoiding the arbitrariness of single-dimensional judgment. At the same time, combined with multi-period trend filtering and sequence continuity scoring, erroneous signals caused by random fluctuations can be eliminated, significantly reducing the probability of misjudgment of entry signals and improving the reference value of the identification results. Furthermore, the method for multi-time-granularity modeling of the on-chain funding data is as follows: S1.1, the on-chain transaction data of the token to be analyzed is modeled as a collection of funding records. Each funding record contains the following information: token identifier, statistical period, net inflow amount within the period, transaction amount within the period, and transaction amount of the previous period with the same granularity. The statistical period covers multiple preset time scales from short to long periods and is arranged from shortest to longest period length to support subsequent cross-period comparisons and hierarchical filtering; S1.2, a price record sequence is constructed. Each price record contains the recording time, the highest price during the period, and the lowest price during the period, which are used for subsequent trend baseline calculations.Furthermore, the record purification includes three stages: time deduplication, internal transfer identification, and abnormal event exclusion. HK 20137911 A Specification 3 Further, the time deduplication method includes: for multiple fund records within the same token, the same time granularity, and the same statistical window, only the record with the larger net inflow amount is retained, and the rest are discarded; historical records within different statistical windows are retained to support subsequent continuous analysis. This rule is used to avoid repeatedly calculating the same fund behavior within the same statistical window, while not disrupting the cross-window time series. Further, the internal transfer identification method includes: for the same short-period window, first determining whether the sum of the total inflow amount and the total outflow amount in that window is greater than zero; if it is not greater than zero, then internal transfer judgment is not performed; if the sum is greater than zero, then the internal transfer ratio is calculated. Furthermore, the method for excluding abnormal events includes calculating the following statistical indicators for historical net inflow records within the same time granularity: Historical mean = average of all historical net inflow records; Standard deviation = dispersion of historical net inflow records; Upper bound of abnormality = historical mean + k × standard deviation, where k is usually configured as 2-3; When the net inflow amount of a record exceeds the upper bound of abnormality, it is regarded as an abnormal large-amount event and excluded to avoid false alarms caused by occasional large-amount transactions; To avoid the protection mechanism of falsely filtering out real market conditions; If the proportion of records exceeding the upper bound of abnormality in the most recent m consecutive statistical windows is ≥ the configured threshold, it is determined that a real large-amount inflow has started, and the exclusion of abnormal events is suspended, so that these continuous large-amount net inflows can enter the subsequent identification link. Furthermore, the multi-dimensional capital behavior feature extraction includes: HK 20137911 A Specification 4 S3.1 Amount dimension feature: Count of qualified transaction cycles; Within a preset time range, count the number of cycles with net inflow amounts not lower than the configured threshold. When this number reaches the minimum number of qualified periods, the amount dimension is considered passed; Judgment logic: Count the number of periods that meet the conditions → If the count ≥ the minimum number of qualified periods → The amount dimension is passed; S3.2 Peak dimension characteristics: The maximum net inflow within a long period window. Find the maximum net inflow amount within the preset long period range. When the maximum net inflow amount is not lower than the configured peak threshold, the peak dimension is considered passed.Judgment Logic: Find the maximum net inflow → If the maximum value ≥ peak threshold → Peak dimension passed; S3.3 Frequency Dimension Feature: Count the effective positive inflow cycle. Within the preset statistical range, count the number of cycles with positive net inflow. When the number of positive inflow cycles is not lower than the minimum frequency threshold, it is considered that the frequency dimension has passed. Judgment Logic: Count the positive inflow cycles → If the count ≥ minimum frequency threshold → Frequency dimension passed; S3.4 Differentiated Threshold Parameter Configuration Principle: The threshold parameters are divided into several continuous levels according to the token capital size. Each level is configured with a net capital inflow threshold, peak threshold, and frequency threshold. The design principle is: the higher the capital size level, the larger the absolute capital size required to trigger entry recognition; the lower the capital size level, the lower the threshold accordingly. Through differentiated threshold configuration, tokens of different sizes can obtain a trigger threshold that matches their capital size under a unified recognition framework. HK 20137911 A Specification 5 Further, the multi-period trend filter includes: S4.1 Time-series grouping divides price records into continuous period groups according to time sequence. Based on the latest price record time, historical price records are grouped according to a configurable fixed period length: price records less than one period from the latest record time are grouped into the most recent period, price records that are one period but less than two periods from the latest record time are grouped into the second most recent period, and so on. For records that fall exactly on the cycle boundary, they are uniformly grouped into the cycle closest to the current one to maintain consistency in grouping rules; S4.2 Cycle Mid-Price Calculation: For each valid cycle, calculate the mid-price of that cycle: Cycle Mid-Price = (Highest Price of the Cycle + Lowest Price of the Cycle) / 2. This mid-price represents the central level of price fluctuations in that cycle; S4.3 Trend Baseline Calculation: Calculate the average of the mid-prices of the most recent w cycles to obtain the trend baseline: Trend Baseline B = Sum of the mid-prices of the most recent w cycles / w. If the historical records are less than w cycles, the actual number of available cycles is used for calculation; If there are no available price records, the trend filter is not passed, and no entry signal is generated; S4.4 Trend Filter Judgment: Trend Filtering Conditions: Current Price ≥ Trend Baseline → Pass Trend Filtering. When the current price < Trend Baseline, even if other capital characteristics are met, no entry signal is triggered. This mechanism effectively eliminates false signals of short-term rebounds in a medium-term downtrend. HK 20137911 A Instruction Manual No. 6.Furthermore, the on-chain behavior sequence continuity scoring includes: S5.1 Cross-granularity deduplication: If two fund records of the same token come from different time granularities but their statistical time windows completely overlap, the record with finer granularity is retained first to avoid the same fund behavior being counted repeatedly; S5.2 Time normalization: Records of different time granularities are uniformly mapped to the start time of their respective statistical windows so that they can be sorted and compared on the same time axis; After preprocessing, the fund records are arranged in chronological order to form a time sequence. Then, all continuous positive net inflow segments are found along the time axis, and the length of each segment is calculated. The length of the longest segment is taken as the continuity score. A multi-dimensional filtering-based on-chain fund entry identification method can obtain the amount dimension, peak dimension, frequency dimension, trend dimension, continuity dimension, and scoring results of on-chain funds, thereby obtaining entry signals. Users can make a conclusion on whether to enter the market based on the entry signals. Furthermore, the judgment logic for the entry signal includes: identifying all time periods with continuous positive net inflows; taking the longest time period and using its cycle number as the continuity score; if the continuity score is ≥ the minimum number of consecutive cycles, then the token exhibits continuous fund accumulation behavior; boundary condition description: if there are no historical records available for sorting within the evaluation window, the continuity score is recorded as zero; HK 20137911 A Specification 7 if only a single valid cycle exists, the continuity score is recorded as one only when that cycle has a positive net inflow, otherwise it is recorded as zero; if adjacent records have partial time overlap, the aforementioned cross-granularity deduplication rule is used to eliminate the duplication effect before performing the continuity score; when all of the following five dimensions are satisfied, the token is deemed to meet the on-chain fund entry identification conditions, and the entry identification result is output: entry signal triggering conditions: amount dimension: number of qualified cycles ≥ minimum number of qualified cycles; peak dimension: maximum net inflow ≥ peak threshold; frequency dimension: number of positive inflow cycles ≥ minimum frequency threshold; trend dimension: current price ≥ trend baseline; Continuous dimension: The length of the longest continuous positive inflow segment is greater than or equal to the minimum number of continuous cycles; When all five dimensions are passed, an entry identification signal is generated.In summary, compared with existing technologies, this invention has the following advantages: 1. It purifies fund records by deduplication, internal transfer exclusion, and abnormal event exclusion, reducing false alarms caused by duplicate statistics, non-genuine transfers between accounts, and occasional large-scale disturbances; 2. It jointly filters fund behavior through three dimensions: amount, peak value, and frequency, so that entry identification no longer relies on a single net inflow threshold, but can simultaneously examine fund size, fund peak value, and frequency of positive inflows; 3. It proposes a multi-period trend filter, constructing a trend baseline based on the midpoint of price highs and lows over several recent fixed-time periods, and outputs identification results only when the current price is not lower than the baseline, thereby eliminating false signals of short-term rebounds in a medium-term downward trend; 4. It proposes an on-chain fund accumulation scoring method based on time continuity, performing cross-granularity deduplication, time normalization, and continuous positive net inflow segment statistics on net inflow records at multiple time granularities, identifying high-quality fund entry behavior with continuous accumulation characteristics; 5. By employing a tiered threshold configuration mechanism, the net inflow threshold, peak threshold, and frequency threshold are matched to the token's capital size level. This avoids the problem of the same set of thresholds being too lenient for large-capital tokens and too strict for small-capital tokens, and outputs more reliable entry identification results under a unified identification framework. Figure 1 is a diagram of the overall system architecture; Figure 2 is a flowchart of the three-dimensional capital behavior feature extraction and joint judgment process; Figure 3 is a flowchart of the multi-period trend filter algorithm; Figure 4 is a flowchart of the sequence continuity scoring algorithm. Detailed Implementation: The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example 1: A multi-dimensional filtering method for filtering on-chain funds information. This method mainly targets the on-chain fund inflows of a specific token in the crypto asset on-chain market. It is a core function of a quantitative investment research and intelligent early warning system, and belongs to the monitoring of this behavior (HK 20137911 A Specification 9) so that the monitor can react accordingly. The identification method of this application includes: S1, multi-time granularity modeling of on-chain funds data. S1.1, modeling the on-chain transaction data of the token to be analyzed as a collection of fund records. Each fund record contains the following information: token identifier, statistical period, net inflow amount in that period, transaction amount in that period, and transaction amount in the previous period of the same granularity. The statistical period covers multiple preset time scales from short to long periods and is arranged from shortest to longest period length to support subsequent cross-period comparisons and hierarchical filtering.S1.2 Construct a price record sequence, where each price record includes the recording time, the highest price during that period, and the lowest price during that period, for subsequent trend baseline calculation. Through this modeling method, the system can simultaneously describe the capital accumulation behavior and price change status of the same token across multiple time scales, providing a unified data foundation for subsequent capital record purification, multi-dimensional feature extraction, and trend filtering. Existing technologies are generally limited to a single or limited number of time scales, failing to balance the timeliness of short-term cycles with the trend-based nature of long-term cycles. This application introduces parallel modeling covering multiple time scales from short to long periods, enabling the system to comprehensively represent capital behavior across different time scales. S2: Based on the capital record set established in step S1, record purification is performed before feature extraction to obtain a purified capital record set. Record purification includes three steps: time deduplication, internal transfer identification, and anomaly event exclusion. S2.1: Time Deduplication For multiple fund records within the same token, time granularity, and statistical window (HK 20137911 A Instruction Manual 10), only the record with the larger net inflow amount is retained, and the rest are discarded. Historical records within different statistical windows are retained to support subsequent continuous analysis. This rule is used to avoid duplicate calculations of the same fund behavior within the same statistical window, while not disrupting the time series across windows. S2.2 Internal Transfer Identification and Exclusion For the same short-period window, first determine if the sum of the total inflow amount and the total outflow amount of the window is greater than zero; if not, internal transfer judgment is not performed. If the sum is greater than zero, the internal transfer ratio is calculated. Internal Transfer Ratio = |Total Inflow - Total Outflow| / (Total Inflow + Total Outflow). When the internal transfer ratio is not higher than the configurable ratio threshold (e.g., 0.2), and the transaction amount of the window does not show effective growth relative to the previous window with the same granularity, the record is considered an internal transfer and excluded. Transaction Amount Growth Rate Determination: If the transaction amount in the previous window > 0, then the growth rate is calculated as (current transaction - previous transaction) / previous transaction. If the growth rate does not exceed the configured threshold (e.g., 20%), it is considered invalid growth. If the transaction amount in the previous window ≤ 0, then the absolute transaction amount increment threshold is used. If the growth amount is less than this threshold, it is considered invalid growth. Internal Transfer Exclusion Condition: (Internal transfer ratio ≤ configured threshold) AND (no valid growth in transaction amount) → This record is excluded.S2.3: Abnormal Event Exclusion: Calculate the following statistical indicators for historical net inflow records within the same time granularity; HK 20137911 A Instruction Manual 11 Historical Mean = Average of all historical net inflow records; Standard Deviation = Dispersion (fluctuation range) of historical net inflow records; Upper Bound of Abnormality = Historical Mean + k × Standard Deviation (where k is usually configured as 2-3); When the net inflow amount of a record exceeds the upper bound of abnormality, it is considered an abnormal large-amount event and excluded to avoid false alarms caused by occasional large-amount transactions. Protection mechanism to avoid false filtering of genuine market activity: If, in the most recent m consecutive statistical windows, the percentage of records exceeding the upper bound of abnormality is ≥ the configured threshold (e.g., 50%), it is determined to be a genuine large-amount inflow, and abnormal event exclusion is suspended, allowing these continuous large-amount net inflows to enter the subsequent identification process. S3: Multi-dimensional Fund Behavior Feature Extraction. Based on the purified fund record set obtained in step S2, fund behavior characteristics are extracted from three dimensions: amount, peak value, and frequency. Differentiated thresholds are set in conjunction with the token fund size level. S3.1 Amount Dimension Feature: Count of Qualified Transaction Cycles; Within a preset time range, the number of cycles with net inflow amount not lower than the configured threshold is counted. When this number reaches the minimum number of qualified cycles (configurable, such as 3 cycles), it is considered that the amount dimension has passed. Judgment logic: Count the cycles that meet the conditions → If the count ≥ the minimum number of qualified cycles → the amount dimension has passed. S3.2 Peak Dimension Feature: Maximum Net Inflow within a Long Cycle Window; Within a preset long cycle range (such as the last 7 days), the maximum net inflow amount within this range is found. When the maximum net inflow amount is not lower than the configured peak value threshold, it is considered that the peak dimension has passed. Judgment logic: Find the maximum net inflow → If the maximum value ≥ the peak value threshold → the peak dimension has passed. S3.3 Frequency Dimension Feature: Count of Effective Positive Inflow Cycles; Within a preset statistical range, the number of cycles with positive net inflow (net inflow > 0) is counted. When the number of positive inflow cycles is not less than the minimum frequency threshold (e.g., 2 cycles), the frequency dimension is considered passed. Judgment logic: Count positive inflow cycles → If the count ≥ minimum frequency threshold → Frequency dimension passed. It is worth noting that: to avoid distortion of frequency statistics due to anomaly exclusion, this dimension prioritizes statistics based on records after time deduplication and before anomaly exclusion. S3.4 Differentiated threshold parameter configuration principle: Threshold parameters are divided into several continuous levels according to the token's capital size, with each level configured with a net capital inflow threshold, peak threshold, and frequency threshold. The design principle is: the higher the capital size level, the larger the absolute capital required to trigger entry identification; the lower the capital size level, the lower the threshold accordingly.By configuring differentiated thresholds, tokens of different sizes can obtain trigger thresholds that match their fund size under a unified identification framework. By jointly verifying fund behavior from three dimensions—amount, peak, and frequency—and introducing a differentiated threshold system stratified by fund size, the false alarm rate is effectively reduced while maintaining adaptability to tokens of different sizes. S4: Multi-Period Trend Filter: The multi-period trend filter constructs a trend baseline based on price data from several recent fixed-length periods, using this as the price threshold for entry signals. Its processing is as follows: HK 20137911 A Instruction Manual 13 S4.1 Time-Sequence Grouping: Price records are divided into consecutive period groups according to time sequence. Using the latest price record time as a reference, historical price records are grouped according to a configurable fixed period length: price records less than one period from the latest record time are grouped into the most recent period; price records one to two periods from the latest record time are grouped into the second most recent period, and so on. Records falling exactly on the period boundary are uniformly grouped into the period closest to the current time to maintain consistency in grouping rules. S4.2 Period Mid-Price Calculation: For each valid period, calculate the mid-price of that period: Period Mid-Price = (Highest Price of the Period + Lowest Price of the Period) / 2. This mid-price represents the central level of price fluctuations in that period. S4.3 Trend Baseline Calculation: Calculate the average of the mid-prices of the most recent w periods (configurable, e.g., 3 periods) to obtain the trend baseline: Trend Baseline B = (Sum of Mid-Prices of the Most Recent w Periods) / w. If the historical records are less than w periods, the actual number of available periods is used. If there are no available price records, the trend filter is not passed, and no entry signal is generated. S4.4 Trend Filtering Judgment: Trend Filtering Conditions: Current Price ≥ Trend Baseline → Pass Trend Filtering. When the current price < Trend Baseline, even if other capital characteristics are met, no entry signal is triggered. This mechanism effectively eliminates false short-term rebound signals in a medium-term downtrend. HK 20137911 A Instruction Manual 14 S5: On-Chain Behavior Sequence Continuity Scoring: Building upon the multidimensional features in step S3 and the trend filtering in step S4, a sequence continuity scoring is further introduced to detect the sustainability of net capital inflows. S5.1 Cross-Granularity Deduplication: If two capital records for the same token come from different time granularities but their statistical time windows completely overlap, the record with the finer granularity is retained first to avoid duplicate calculations of the same capital behavior. S5.2 Time Normalization: Records at different time granularities are uniformly mapped to the start time of their respective statistical windows for sorting and comparison on the same time axis. After preprocessing, capital records are arranged chronologically to form a time-series sequence.Then, all consecutive positive net inflow segments (time periods with multiple consecutive positive inflows) are identified along the time axis. The length (number of periods) of each segment is calculated, and the length of the longest segment is taken as the continuity score. A multi-dimensional filtering-based on-chain fund entry identification method, based on the four steps S1-S4 above, can obtain signals in the amount, peak, frequency, trend, and continuity dimensions of on-chain funds. Users can determine whether to enter the market based on these signals. This embodiment provides a logic for determining entry signals: Identify all consecutive positive net inflow time periods; Take the longest time period and use its number of periods as the continuity score; If the continuity score ≥ the minimum number of consecutive periods (e.g., 2), then the token exhibits continuous fund accumulation behavior. Boundary Conditions: If no historical records are available for sorting within the evaluation window, the continuity score is zero (HK 20137911 A Specification 15). If only a single valid period exists, the continuity score is one only if that period has a positive net inflow; otherwise, it is zero. If adjacent records have partial time overlap, the aforementioned cross-granularity deduplication rule is used to eliminate the duplication effect before performing the continuity score. When all five dimensions below are met, the token is deemed to meet the on-chain fund entry identification conditions, and the entry identification result is output: Entry Signal Trigger Conditions (Five-Dimensional AND Logic): 1. Amount Dimension: Number of qualified periods ≥ Minimum number of qualified periods; 2. Peak Dimension: Maximum net inflow ≥ Peak threshold; 3. Frequency Dimension: Number of positive inflow periods ≥ Minimum frequency threshold; 4. Trend Dimension: Current price ≥ Trend baseline; 5. Continuity Dimension: Length of the longest continuous positive inflow segment ≥ Minimum number of continuous periods. Only when all five dimensions are met will an entry identification signal be generated. By statistically analyzing the maximum length of consecutive positive net inflow segments over time, the quality of capital accumulation is quantified in terms of time continuity. Only when several consecutive time periods show positive net inflows can the identification and verification be passed, and a more reliable entry identification result can be output. In another implementation, trend filtering and continuity verification can also be achieved using equivalent multi-period statistical benchmarks. That is, while maintaining the main multi-dimensional filtering link, trend and continuity judgments are constructed using statistical results from other periods, which can also achieve the effect of identifying genuine capital entry behavior. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “described,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items.It should be understood that although the present invention may use terms such as first, second, third, etc., to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination." Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents. HK 20137911 A Claim 1: A method for filtering on-chain funds information based on multi-dimensional filtering, comprising: S1: Multi-time granular modeling of on-chain funds data; S2: Based on the fund record set established in step S1, performing record purification before feature extraction to obtain a purified fund record set; S3: Multi-dimensional fund behavior feature extraction, based on the purified fund record set obtained in step S2, extracting fund behavior features from three dimensions: amount, peak value, and frequency, and setting differentiated thresholds in conjunction with the token fund volume level; S4: Multi-period trend filter; S5: On-chain behavior sequence continuity scoring. 2. The on-chain fund information filtering method based on multi-dimensional filtering according to claim 1, characterized in that: the method for multi-time granularity modeling of the on-chain fund data is as follows: S1.1, modeling the on-chain transaction data of the token to be analyzed as a collection of fund records, each fund record containing the following information: token identifier, statistical period, net inflow amount within the period, transaction amount within the period, and transaction amount of the previous period with the same granularity; the statistical period covers multiple preset time scales from short to long periods, and is arranged from short to long period length to support subsequent cross-period comparison and hierarchical filtering; S1.2 constructing a price record sequence, each price record containing the recording time, the highest price of the period, and the lowest price of the period, for subsequent trend baseline calculation.3. The on-chain fund information filtering method based on multi-dimensional filtering according to claim 1, characterized in that: the record purification includes three stages: time deduplication, internal transfer identification, and abnormal event exclusion; the time deduplication method includes: for multiple fund records within the same statistical window of the same token, the same time granularity, and the same HK 20137911 A claim, only the one with the larger net inflow amount is retained, and the rest of the records are removed; historical records within different statistical windows are retained to support subsequent continuous analysis. This rule is used to avoid repeatedly calculating the same fund behavior within the same statistical window, while not disrupting the time series across windows. 4. The on-chain fund information filtering method based on multi-dimensional filtering according to claim 3, characterized in that: the internal transfer identification method includes, for the same short-period window, first determining whether the sum of the total inflow amount and the total outflow amount of the window is greater than zero; if it is not greater than zero, then internal transfer judgment is not performed; if the sum of the two is greater than zero, then the internal transfer ratio is calculated. 5. A method for filtering on-chain funds information based on multi-dimensional filtering according to claim 3, characterized in that the method for excluding abnormal events includes calculating the following statistical indicators for historical net inflow records within the same time granularity: Historical mean = average of all historical net inflow records; Standard deviation = dispersion of historical net inflow records; Upper bound of abnormality = historical mean + k × standard deviation, where K is usually configured as 2-3; When the net inflow amount of a certain record exceeds the upper bound of abnormality, it is regarded as an abnormal large-amount event and excluded to avoid false alarms caused by occasional large-amount transactions; To avoid the activation of a protection mechanism for real market conditions being mistakenly filtered; If the percentage of records exceeding the upper bound of abnormality in the most recent m consecutive statistical windows is greater than the configured threshold, it is determined that a real large-amount inflow has been activated, and the exclusion of abnormal events is suspended, so that these continuous large-amount net inflows can enter the subsequent identification link.2 HK 20137911 A Claim 6. A method for filtering on-chain funds information based on multi-dimensional filtering according to claim 1, characterized in that the multi-dimensional fund behavior feature extraction includes: S3.1 Amount dimension feature: Count of qualified transaction cycles; Within a preset time range, count the number of cycles with net inflow amount not less than the configured threshold. When this number reaches the minimum number of qualified cycles, it is considered that the amount dimension passes; Judgment logic: Count the cycles that meet the condition → If the count > the minimum number of qualified cycles → Amount dimension passes; S3.2 Peak dimension feature: Maximum net inflow within a long cycle window. Within a preset long cycle range, find the maximum net inflow amount within that range. When the maximum net inflow amount is not less than the configured peak threshold, it is considered that the peak dimension passes; Judgment logic: Find the maximum net inflow → If the maximum value > the peak threshold → Peak dimension passes; S3.3 Frequency dimension feature: Valid positive inflow cycle count: Within a preset statistical range, the number of cycles with positive net inflow is counted. When the number of positive inflow cycles is not lower than the minimum frequency threshold, it is considered that the frequency dimension has passed. The judgment logic is: count positive inflow cycles → if the count > minimum frequency threshold → frequency dimension has passed; S3.4 Differentiated threshold parameter configuration principle: The threshold parameter is divided into several continuous levels according to the token capital size. Each level is configured with a net capital inflow threshold, peak threshold, and frequency threshold. The design principle is: the higher the capital size level, the larger the absolute capital size required to trigger entry recognition; the lower the capital size level, the lower the threshold is. Through differentiated threshold configuration, tokens of different sizes can obtain a trigger threshold that matches their capital size under a unified recognition framework.7. A multi-dimensional filtering chain for filtering capital information according to claim 1, characterized in that the multi-period trend filter includes: S4.1 Time-series grouping: Price records are divided into consecutive period groups according to time sequence. Taking the latest price record time as a reference, historical price records are grouped according to a configurable fixed period length: price records less than one period length from the latest record time are assigned to the most recent period; price records one period length but less than two period lengths from the latest record time are assigned to the second most recent period, and so on. For records that fall exactly on the period boundary, they are uniformly assigned to the period closer to the current time to maintain consistency in grouping rules; S4.2 Period mid-price calculation: For each valid period, the mid-price of the period is calculated: Period mid-price = (highest price of the period + lowest price of the period) I 2. This mid-price represents the central level of price fluctuation in the period; S4.3 Trend baseline calculation: The average of the mid-price of the most recent w periods is used to obtain the trend baseline: Trend baseline B = sum of the mid-price of the most recent w periods I w. If the historical records are less than w If there are no available price records, the trend filter is not passed and no entry signal is generated; S4.4 Trend Filter Judgment 4 HK 20137911 A Claim Trend Filter Condition: Current Price > Trend Baseline ➔ When the current price < trend baseline, even if other capital characteristics are met, no entry signal is triggered. This mechanism effectively eliminates false short-term rebound signals in a medium-term downtrend. 8. A method for filtering on-chain funds information based on multi-dimensional filtering according to claim 1, characterized in that the on-chain behavior sequence continuity scoring includes: S5.1 Cross-granularity deduplication: If two fund records of the same token come from different time granularities but their statistical time windows completely overlap, the record with finer granularity is retained first to avoid the same fund behavior being counted repeatedly; S5.2 Time normalization: Records of different time granularities are uniformly mapped to the start time of their respective statistical windows so that they can be sorted and compared on the same time axis; After preprocessing, the fund records are arranged in chronological order to form a time sequence, and then all continuous positive net inflow segments are found along the time axis. The length of each segment is calculated, and the length of the longest segment is taken as the continuity score. 9. A method for identifying on-chain funds entry based on multi-dimensional filtering, characterized in that it includes the on-chain fund information filtering method based on multi-dimensional filtering as described in claim 1, which obtains the amount dimension, peak dimension, frequency dimension, trend dimension, continuity dimension and scoring result of on-chain funds through the filtering method, thereby obtaining an entry signal, and the user makes a conclusion on whether to enter the market based on the entry signal.10. A method for identifying on-chain funds entering the market based on multi-dimensional filtering according to claim 10, characterized in that the judgment logic of the entry signal includes: finding all time periods of continuous positive net inflow; taking the longest time period and using its period number as the continuity score; if the continuity score > the minimum number of continuous periods, then the token has continuous fund accumulation behavior; boundary condition description: if there are no historical records available for sorting within the evaluation window, the continuity score is recorded as zero; if there is only a single valid period, the continuity score is recorded as one only when the period is a positive net inflow, otherwise it is recorded as zero; if there is partial time overlap between adjacent records, the repeated influence is eliminated by the aforementioned cross-granularity deduplication rule before the continuity score is executed; when all of the following five dimensions are satisfied, the token is deemed to meet the on-chain funds entry identification conditions, and the entry identification result is output: entry signal triggering conditions: amount dimension: number of qualified periods > minimum number of qualified periods; peak dimension: maximum net inflow > peak value; frequency dimension: number of positive inflow periods > minimum frequency threshold; Trend Dimension: Current Price > Trend Baseline; Continuity Dimension: Longest Continuous Positive Inflow Segment Length > Minimum Continuous Period Count; An entry signal is generated when all five dimensions are met. 6 HK 20137911 A Instruction Manual Appendix 1 Figure 1 HK 20137911 A Instruction Manual Appendix 2 Figure 2 HK 20137911 A Instruction Manual Appendix 3 Figure 3 HK 20137911 A Instruction Manual Appendix 4 Figure 4 HK 20137911 A.

Images