Market quotation prediction method and system based on block chain data and multi-modal fusion

Abstract

The invention discloses a market quotation prediction method and system based on block chain data and multi-modal fusion, and relates to the technical field of data analysis, by fusing block chain native data, an objective physical state reference is established from a market microstructure level, and the depth and uniqueness of analysis are enhanced; and by capturing and analyzing the original sequential logic triggered by the emotion instead of simple emotion scores, the key insight in the market psychological evolution rhythm is reserved. According to the method, the joint anomaly is judged only when the physical state deviates from the standard and the emotion triggering sequence is abnormal at the same time, the recognition specificity of a real inflection point signal is remarkably improved, noise interference caused by fluctuation of a single factor is effectively filtered, and therefore the reliability of a prediction result is improved.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a market forecasting method and system based on blockchain data and multimodal fusion. Background Technology

[0002] With the evolution of big data and artificial intelligence technologies, market forecasting methods have shifted from traditional econometric models to data-driven intelligent analysis. Current mainstream technical approaches mainly fall into two categories: one is time-series forecasting based on large-scale historical market data, using algorithms such as deep learning to uncover non-linear patterns in price sequences; the other is multi-source information fusion analysis, aiming to integrate technical indicators, fundamental data, and market sentiment to provide a more comprehensive market view. While these methods have improved the diversity of forecasts to some extent, they still face core challenges such as limited data dimensions and superficial analytical logic, making it difficult to capture deep inflection point signals driven by changes in market microstructure and the evolution of collective sentiment. For example: CN119741131A discloses a market analysis method and system based on big data. This technical solution represents a paradigm that heavily relies on historical price data for prediction and anomaly detection. Its shortcomings are: First, the data source is limited to traditional market transaction data (such as historical prices), completely ignoring the on-chain physical state data natively recorded by emerging market infrastructures such as blockchain, which reflects changes in the underlying liquidity distribution and holder structure of assets. This results in a lack of measurement of the market's physical foundation. Second, the anomaly detection module operates independently, only identifying statistical outliers in price sequences. It does not construct a dynamic benchmark interval for quantifying the healthy normal state of the market, nor can it determine whether the outbreak of market sentiment (such as news and public opinion) matches the current underlying physical state of the market. Therefore, it is difficult to distinguish between ordinary emotional fluctuations and state-emotion mismatch anomalies that may lead to trend reversals.

[0003] CN121329659A discloses a method and system for analyzing financial futures market trends based on multi-agent collaboration. While this technical solution represents an advanced direction in multi-source information fusion, it has significant limitations in sentiment and multimodal data processing. Firstly, the processing of textual information such as news and social media ultimately aims to output a comprehensive market sentiment index or score. This approach obscures and loses the original sequential information of how different external events trigger market sentiment. The temporal logic of sentiment triggers is often key to predicting the pace of market reactions. Secondly, although its sentiment analysis and risk warning modules are weighted and fused, they are essentially static superpositions at the feature level. No mechanism is designed to verify the rationality of the sentiment; that is, the sentiment sequence is not compared and cross-validated in real-time with another independent and objective on-chain physical state benchmark that reflects the market's internal state. Therefore, it cannot accurately identify high-risk joint anomalies such as "the market foundation is unstable, but sentiment continues to react blindly in the old order."

[0004] In summary, existing market analysis technologies, whether focusing on price sequence modeling or multi-source information fusion, all suffer from common problems such as missing data dimensions (failure to effectively incorporate and structure on-chain data), one-sided analytical logic (failure to construct a cross-validation mechanism between state benchmarks and sentiment trigger sequences), and superficial signal recognition (failure to define deep anomalies from the combined perspective of state mismatch and sequence reversal). Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a market forecasting method based on blockchain data and multimodal fusion, characterized by comprising: collecting blockchain ledger data corresponding to a target asset; constructing an on-chain physical state dataset based on the blockchain ledger data; and determining a state baseline interval based on a historical time window; collecting off-chain multimodal data associated with the target asset; constructing an off-chain sentiment state dataset based on the off-chain multimodal data; recording the triggering order of the off-chain sentiment state dataset in the time dimension to form an off-chain sentiment triggering sequence; calculating the offset of the on-chain physical state relative to the state baseline interval; comparing the off-chain sentiment triggering sequence with a historical sequence template to generate a sequence offset identifier; fusing the offset and the sequence offset identifier to generate a joint abnormal state identifier; using the joint abnormal state identifier as input, classifying and labeling the market state within a target time period, and outputting a market inflection point prediction result, wherein the market inflection point prediction result is used to indicate whether the market trend continues or reverses.

[0008] In a preferred embodiment of the present invention, the on-chain physical state dataset includes at least valuation state parameters and capital structure state parameters; the off-chain sentiment state dataset includes at least sentiment intensity parameters and external event triggering parameters.

[0009] As a preferred embodiment of the present invention, the construction of the on-chain physical state dataset includes: collecting raw on-chain data of the target asset within a historical time window from a publicly available blockchain explorer API; the raw on-chain data includes at least the number of newly active addresses and transaction flow records organized by time series, wherein the transaction flow records include transaction timestamps, transaction amounts, and transaction directions; aggregating and statistically analyzing the raw on-chain data according to a preset time granularity: for the number of newly active addresses, counting the number of new addresses by time granularity; for transaction flow, summarizing the large transaction amounts transferred into and out of exchanges by time granularity; calculating two on-chain physical state parameters based on the aggregated time series data: using the number of newly active addresses after logarithmic transformation as the valuation state parameter; dividing the difference between the amount transferred into and out of exchanges within each time granularity by the total transaction amount of the corresponding time granularity as the capital structure state parameter; the construction of the state benchmark interval includes: calculating the respective first-order intervals of the valuation state parameter and the capital structure state parameter within a historical time window using a rolling window. With the Percentiles; the interval formed by two percentiles is defined as the on-chain physical state reference interval for the corresponding parameters.

[0010] As a preferred embodiment of the present invention, the construction of the off-chain sentiment state dataset includes: collecting off-chain multimodal data associated with the target asset from a preset news and social platform API, including at least news headline text and corresponding publication time; using a pre-trained word vector model to convert each news headline text into a corresponding word vector representation, forming a text vector sequence; performing density-based clustering analysis on the text vector sequence on the time axis, setting a minimum cluster size parameter to filter noise; identifying each cluster that meets the minimum cluster size as an external event type, and assigning a unique event code to each event type as an external event triggering parameter; calculating the similarity score between the text vector within each time window and its corresponding cluster center, the similarity score being obtained by normalizing the cosine similarity to the [0,1] interval; quantifying the weighted average of the similarity scores of all text vectors within the time window into a sentiment intensity parameter at the corresponding moment; associating the corresponding sentiment intensity parameter with the moment when the external event triggering parameter is first identified, forming a sentiment state unit; arranging all sentiment state units in chronological order to form an off-chain sentiment triggering sequence.

[0011] As a preferred embodiment of the present invention, the calculation of the offset includes: obtaining the on-chain physical state parameter value aligned with the timestamp of each emotional state unit in the off-chain emotional triggering sequence; using a normalized position score algorithm to calculate the relative position of each on-chain physical state parameter value within the corresponding on-chain physical state reference interval; and aggregating all relative position scores to obtain a comprehensive state offset.

[0012] In a preferred embodiment of the present invention, the generation of the sequence offset identifier includes: extracting the order of external event trigger parameter types from the off-chain emotion trigger sequence, forming a sequence of length [missing information]. The event sequence is calculated; the statistical distribution distance between the event sequence and the historical event subsequences in all equal-length historical sequence templates is calculated; if the minimum distance is greater than a specific upper quantile of the historical distance, a trigger order offset identifier is generated.

[0013] As a preferred embodiment of the present invention, the generation of the joint anomaly status identifier includes: when the status offset is in the tail region of its own historical value distribution and the trigger sequence offset identifier is generated, a joint anomaly is determined to have occurred, and a joint anomaly status identifier is generated.

[0014] As a preferred embodiment of the present invention, the step of classifying and labeling the market state within the target time period and outputting the market inflection point prediction result includes: dividing the target time period into a joint abnormal state period and a normal period based on the timestamp of the occurrence of the joint abnormal state identifier within the target time period; for each joint abnormal state period, extracting a market price sequence of a fixed time length after the end of the abnormal period; calculating the linear regression slope of the market price sequence, if it is positive, marking it as rising, and if it is negative, marking it as falling, as the actual trend direction label corresponding to the joint abnormal state period; for the current joint abnormal state period to be predicted... Extract the state offset time series and calculate the feature vector, which includes at least the mean, variance, skewness, and kurtosis. In historical data, calculate the Euclidean distance between the current feature vector and the feature vector of each historical joint abnormal state period, and select the K historical events with the smallest distance as a set of similar events. Statistically analyze the distribution of the actual trend direction labels in the set of similar events. If the proportion of rising labels exceeds a preset threshold, predict that the current state will continue its upward trend. If the proportion of falling labels exceeds a preset threshold, predict that the trend will reverse and turn downward. If neither exceeds the preset threshold, output the trend that continued before the anomaly occurred.

[0015] To address the aforementioned technical problems, this invention provides the following technical solution: a market forecasting system based on blockchain data and multimodal fusion, characterized in that it includes: The on-chain state construction module collects blockchain ledger data corresponding to the target asset, constructs an on-chain physical state dataset based on the blockchain ledger data, and determines the state baseline interval based on historical time windows. The off-chain sentiment sequence module collects off-chain multimodal data associated with the target asset, constructs an off-chain sentiment state dataset based on the off-chain multimodal data, and records the triggering order of the off-chain sentiment state dataset in the time dimension to form an off-chain sentiment triggering sequence. The state order offset module calculates the offset of the on-chain physical state relative to the state baseline interval, and compares the off-chain emotion trigger sequence with the historical order template to generate an order offset identifier; it then merges the offset and the order offset identifier to generate a joint abnormal state identifier. The joint anomaly prediction module takes the joint anomaly status identifier as input, classifies and labels the market status within the target time period, and outputs the market inflection point prediction result. The market inflection point prediction result is used to indicate whether the market trend continues or reverses.

[0016] The beneficial effects of this invention are as follows: By integrating native blockchain data, this invention establishes an objective physical state benchmark at the level of market microstructure, enhancing the depth and uniqueness of the analysis; and by capturing and analyzing the original temporal logic of emotion triggers, rather than simple emotional scores, it retains key insights into the rhythm of market psychological evolution.

[0017] This invention requires that a deviation from the baseline in physical state and an abnormal sequence of emotional triggers occur simultaneously to determine a joint anomaly. This significantly improves the specificity of identifying true inflection point signals, effectively filters out noise interference caused by fluctuations in a single factor, and thus enhances the reliability of prediction results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of the market forecasting method based on blockchain data and multimodal fusion, as shown in this invention.

[0019] Figure 2 This is a structural diagram of the market forecasting system based on blockchain data and multimodal fusion, as shown in this invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates a market forecasting method based on blockchain data and multimodal fusion, including: S1: Collect blockchain ledger data corresponding to the target asset, construct an on-chain physical state dataset based on the blockchain ledger data, and determine the state baseline interval based on historical time windows.

[0024] The on-chain physical state dataset includes at least valuation state parameters and capital structure state parameters.

[0025] S1.1: Collect raw on-chain data of the target asset within a historical time window from at least three independent, publicly available blockchain explorer APIs.

[0026] The raw on-chain data includes at least the number of newly active addresses organized in time series (by calling the getDailyActiveAddressCount interface of the blockchain explorer to obtain the number of unique addresses that first appear in a transaction each day, and recording them as a timestamp-value pair sequence) and transaction flow records, where the transaction flow records include the transaction timestamp, transaction amount, and transaction direction.

[0027] Specifically, the transaction log records are filtered by calling the blockchain explorer interface to identify transactions whose single transaction amount exceeds a preset threshold. Each record contains the following fields: transaction hash, timestamp, transaction amount, and transaction direction identifier.

[0028] S1.2: Process the raw data on the chain according to the preset time granularity. Aggregate statistics (e.g., 1 hour or 1 day): For the number of newly active addresses, count the number of new addresses by time granularity; for transaction volume, summarize the large transaction amounts transferred to and from exchanges by time granularity.

[0029] Specifically, the aggregation process for the number of newly active addresses is as follows: the timeline is divided into continuous segments. The intervals, with the starting time point for each interval being: ,in is the starting timestamp of the historical time window, and k is a non-negative integer representing the sequence number of the time interval.

[0030] For each time interval The number of addresses that first engage in a transaction within this interval is counted and denoted as . For raw data at the daily granularity... For cases with a duration of less than one day, a time-weighted allocation algorithm is used: if the number of new addresses added on a certain day is... Then that day was Divided into Each time interval was obtained Each address is incremented by an interval.

[0031] The aggregation process for large transaction flows is as follows: for each time interval Calculate the total transaction amount in both directions: Total amount transferred to the exchange. Total amount transferred out of the exchange The total transaction amount is obtained by adding the two together. Based on this, an aggregation sequence is generated: .in, This represents the total number of time intervals.

[0032] S1.3: Based on the aggregated time series data, calculate two on-chain physical state parameters: use the number of newly added active addresses after logarithmic transformation as the valuation state parameter; use the difference between the amount transferred into and out of exchanges within each time granularity, divided by the total transaction amount of the corresponding time granularity, as the capital structure state parameter.

[0033] A better approach is to calculate the physical state parameters of the two chains using the following deterministic algorithm: (1) Valuation status parameters: Number of newly active addresses for each time interval A logarithmic transformation is performed to compress the impact of extreme values. This parameter reflects the marginal change in network adoption rate. Based on Metcalfe's Law, the network value is proportional to the square of the number of users, and the first-order difference after logarithmic transformation approximates the growth rate.

[0034] (2) Capital structure status parameters: Calculate the net capital outflow ratio for each time interval: when hour, ; when hour, ; The parameter ranges from [-1, 1]. A positive value indicates a net inflow of funds into the exchange (potential increase in selling pressure), while a negative value indicates a net outflow of funds from the exchange (potential increase in the willingness to hold coins).

[0035] It should be noted that the VSP parameter transforms the power-law distributed data into an approximately normal distribution through logarithmic transformation, facilitating subsequent statistical analysis. The CSP parameter eliminates the scale effect of absolute amounts through ratio calculation, making assets with different market capitalizations comparable. The two parameters provide complementary market fundamental views from the two dimensions of user growth and capital flow.

[0036] S1.4: The construction of the state benchmark interval includes: calculating the respective first-order intervals of the valuation state parameters and the capital structure state parameters within a historical time window using a rolling window (e.g., 90 days). (e.g., 25) and the (e.g., 75) Percentile; the interval formed by two percentiles is defined as the on-chain physical state reference interval for the corresponding parameter.

[0037] It should be noted that the construction of the state benchmark interval aims to establish a dynamically evolving, normalized reference framework for on-chain physical state parameters, used to quantitatively assess the deviation of current parameter values ​​from historical norms. Its core lies in using a rolling window mechanism to enable the benchmark interval to adaptively reflect changes in market conditions, avoiding the lag or insufficient adaptability issues caused by using fixed thresholds. The benchmark interval is constructed independently for valuation state parameters and capital structure state parameters, forming two parallel but logically consistent time series intervals. Specifically: The first step in the construction process is to determine the length of the rolling window. This length is set in units of time granularity; for example, setting it to ninety days means ninety consecutive time granularity units. If the time granularity is one hour, the rolling window length corresponds to 2,160 hours; if the time granularity is one day, it corresponds to ninety days. This window length needs to be sufficient to cover the short- to medium-term fluctuation cycles of the market to capture the normal fluctuation range of the parameters, while not being so long as to dilute the characteristics of the recent market state.

[0038] Furthermore, for each specific time point, for each physical state parameter on the chain, historical parameter values ​​within the length of the scrolling window are extracted by tracing back from that time point, forming a subset of parameter values. For example, at time point... Extract the estimated state parameters within the time range. All values ​​within this subset constitute a subset of the estimated state parameters, where, The window length is defined; similarly, all values ​​of the capital structure state parameters within the same time range are extracted to form a subset of the capital structure state parameters.

[0039] Subsequently, statistical quantiles were calculated for each parameter subset. Specifically, the 25th percentile and 75th percentile of each parameter subset were calculated. The 25th percentile represents the critical point where the parameter is at a low level within the historical window, meaning only 25% of the data points are below this value; the 75th percentile represents the critical point where the parameter is at a high level, meaning only 25% of the data points are above this value. The interval between these two percentiles, covering the core 50% distribution range of the parameter within the historical window, is defined as the baseline interval for the parameter's on-chain physical state at the current time point. For the valuation state parameter, the baseline interval is represented as a closed interval from the 25th percentile to the 75th percentile. For the capital structure state parameter, the baseline interval is represented as a closed interval from the 25th percentile to the 75th percentile. These two intervals respectively characterize the normal fluctuation boundaries of valuation activity and capital flow in recent history.

[0040] In the initial phase, when the available historical data points are less than the preset rolling window length, an extended window strategy is adopted. This involves using all existing data for percentile calculations, while simultaneously labeling the confidence level of the results, with the confidence weight proportional to the amount of accumulated data. As data accumulates, once the data volume reaches the full window length, the system switches to standard rolling window calculation mode.

[0041] It should be noted that the rolling window design allows the baseline interval to dynamically adapt to changes in market conditions, avoiding the obsolescence problem caused by fixed thresholds. Furthermore, the use of the interquartile range (IQR) instead of the full range to construct the interval provides robustness against outliers. The linear interpolation algorithm ensures the accuracy and repeatability of percentile calculations.

[0042] S2: Collect off-chain multimodal data associated with the target asset, construct an off-chain sentiment state dataset based on the off-chain multimodal data, and record the triggering order of the off-chain sentiment state dataset in the time dimension to form an off-chain sentiment triggering sequence.

[0043] The off-chain emotion state dataset includes at least emotion intensity parameters and external event triggering parameters.

[0044] S2.1: Collect off-chain multimodal data associated with the target asset from the preset news and social platform APIs, including at least the news headline text and the corresponding publication time; use a pre-trained word vector model to convert each news headline text into a corresponding word vector representation, forming a text vector sequence.

[0045] Specifically, a distributed crawler system is deployed, connecting to three types of pre-defined data source APIs: news aggregation platforms (including Cryptopanic and CoinTelegraph's RSS subscription interfaces), social media platforms (including Twitter's FilteredStream API and Reddit's Pushshift API), and professional forum platforms (including Bitcointalk's API interface). A dynamic set of keywords is set for the target asset, such as the target asset's official name, common abbreviations, trading codes, and common aliases.

[0046] Text data containing any keyword is collected in real time through the filtering mechanism of each platform's API. Each collection record contains the following metadata fields: text identifier, original text content, publication timestamp, source platform identifier, and dissemination popularity index (calculated in the following ways: for news, the number of cited media; for Twitter, the number of retweets and likes; for Reddit, the number of upvotes; for forums, the number of replies, etc.).

[0047] Furthermore, forming the text vector sequence includes: A pre-trained multilingual BERT model was used as the base word vector generator.

[0048] Perform the following deterministic processing flow for each text: (1) Text cleaning: Remove URL links, @ mentions, and special emojis, while retaining punctuation marks; (2) Tokenization and encoding: The text is converted into a sequence of tokens using BERT's WordPiece. (3) Vector generation: Input the token sequence into the BERT model and extract the 768-dimensional vector corresponding to the [CLS] tag in the last hidden state as the text representation vector. ; (4) Perform L2 normalization on v to obtain a unit vector. .

[0049] Finally, a text vector sequence is formed. ,in For timestamps, For normalized text vectors, For the corresponding dissemination popularity index, This refers to the total number of valid texts collected, cleaned, and successfully vectorized from all off-chain multimodal data sources within a preset historical time window.

[0050] It should be noted that the multi-source parallel acquisition in this invention ensures the comprehensiveness and real-time nature of sentiment data, and the dynamic keyword set avoids omitting relevant discussions. The use of a multilingual BERT model enables the capture of semantic information across languages, and L2 normalization eliminates the impact of vector length differences on similarity calculation, providing standardized input for subsequent clustering analysis.

[0051] S2.2: On the time axis, perform density-based clustering analysis on the text vector sequence and set the minimum cluster size parameter to filter noise.

[0052] Ideally, set the time window length. To maintain consistency with step S1 and facilitate time series alignment, the entire timeline is divided into continuous time windows. .

[0053]

[0054] For each time window Extract all text vectors within the window to form the window dataset:

[0055] For each window dataset DBSCAN clustering is performed using the following algorithm: (1) Calculation The cosine distance matrix C between all pairs of text vectors (where cosine distance = 1 - cosine similarity) is used, and the upper quartile of C is taken as the neighborhood radius.

[0056] (2) Dynamically set the minimum sample size : in The number of text items in the window.

[0057] (3) Each text vector is considered a data point. If the neighborhood of point p contains at least If there are 1 point, then p is the core point; clusters are constructed by the density direct access relationship of the core points.

[0058] Optionally, points that do not meet the core point criteria and do not belong to the neighborhood of any core point can be marked as noise points and excluded from subsequent analysis.

[0059] It should be noted that if the number of texts in the window is less than a certain threshold (for example, less than 2), the window is skipped, no clustering is performed, and no sentiment state unit is generated.

[0060] Record the following attributes for each cluster: cluster center vector (calculated as the arithmetic mean of all text vectors within the cluster and then L2 normalized), cluster size (number of texts within the cluster), first appearance timestamp (timestamp of the earliest text within the cluster), and total propagation popularity (sum of propagation popularity indices of all texts within the cluster).

[0061] Furthermore, an event tracing graph G=(V,E) is constructed, where nodes represent a cluster and edges represent cross-window relationships. For adjacent time windows, the following calculations are performed. Each cluster in The cosine similarity of the center vectors of each cluster in G is calculated. If the cosine similarity of the center vectors is greater than a similarity threshold (set to 0.85), a directed edge is added to graph G to represent an event continuation relationship. Connected components in the graph are identified using depth-first search, with each connected component corresponding to a continuous event.

[0062] It should be noted that in this invention, the cosine similarity metric effectively captures semantic similarity and avoids the curse of dimensionality problem of Euclidean distance in high-dimensional space. The event tracking graph realizes event association across time windows and can accurately identify the start, duration and end of events, providing a foundation for event sequence construction.

[0063] S2.3: Identify each cluster that meets the minimum cluster size as an external event type, and assign a unique event code to each event type as an external event triggering parameter.

[0064] For each event connected component identified in step S2.2, the weighted average of all cluster center vectors is extracted as the event topic vector, with the weight being the sum of the propagation heat of each cluster.

[0065] Use a hierarchical clustering algorithm to cluster all event topic vectors: a. Calculate the cosine similarity matrix between event topic vectors; b. Use the average linkage method for agglomerative hierarchical clustering; c. Cut the tree diagram with a distance threshold (0.75) to form event type groups.

[0066] Each event type group is assigned a unique event type code, in the format ETC-XXX, where XXX is a three-digit serial number.

[0067] Furthermore, for each specific event, it is grouped according to its event type and inherits the corresponding event type code as the base value for the external event triggering parameter.

[0068] To distinguish different instances of the same type, a timestamp prefix can be added to form a complete external event trigger parameter value. The above operations can establish a mapping table between event instances and event types.

[0069] S2.4: For each text vector within a time window, calculate the similarity score with its cluster center. The similarity score is obtained by normalizing the cosine similarity to the [0,1] interval. The weighted average of the similarity scores of all text vectors within the time window is quantified into the emotional intensity parameter at the corresponding moment, where the weight is the text's popularity index.

[0070] Furthermore, for each external event, the emotional intensity parameter value corresponding to its first occurrence time window is extracted as the initial emotional intensity of the event.

[0071] S2.5: Based on the moment when the external event trigger parameter is first identified, associate it with the corresponding emotion intensity parameter to form an emotion state unit; arrange all emotion state units in chronological order to form an off-chain emotion trigger sequence.

[0072] The emotional state unit is a quadruple, including external event trigger parameters (fully encoded), the precise timestamp of the first identification of the event (taken from S2.3), the emotional intensity parameters of the first time window of the event, and the average emotional intensity parameters during the duration of the event.

[0073] All emotional state units are sorted in ascending order by their first timestamp, and a sequence index is added to each emotional state unit to indicate its position in the sequence.

[0074] S3: Calculate the offset of the on-chain physical state relative to the state baseline interval, and compare the off-chain sentiment trigger sequence with the historical sequence template to generate a sequence offset identifier; merge the offset and the sequence offset identifier to generate a joint abnormal state identifier.

[0075] It should be noted that this invention quantifies the deviations of on-chain physical states and off-chain sentiment states from their historical norms, and fuses anomalous signals from both dimensions based on strict statistical criteria to generate a high-confidence joint anomalous state identifier. Furthermore, this invention does not simply superimpose two independent anomalous signals, but establishes a statistical correlation mechanism between extreme deviations in on-chain fundamentals and structural disorder in the sequence of off-chain sentiment events. This method can effectively filter out noise fluctuations in a single dimension and accurately capture key moments when market fundamentals and collective sentiment perception deviate in tandem. Specifically: S3.1: The calculation of the offset includes: Obtain the on-chain physical state parameter value aligned with the timestamp of each emotion state unit in the off-chain emotion trigger sequence. If the aligned timestamp does not have corresponding on-chain data, it is obtained by interpolation from adjacent time points using linear interpolation.

[0076] Specifically, for each emotional state unit in the off-chain emotional trigger sequence, the precise timestamp of its first identification is obtained. Based on the data collection time granularity of the on-chain physical state parameters, a down-rounding alignment operation is performed on this timestamp, aligning it to the start point of the most recent complete data collection cycle. For example, if on-chain data is collected once per hour, the timestamp of the emotional event is aligned to the hour. Through this alignment operation, the valuation state parameter value, capital structure state parameter value, and the upper and lower bounds of the dynamic benchmark interval calculated from the historical data of the rolling window are retrieved from the preset on-chain physical state parameter time series, corresponding to the aligned timestamp.

[0077] Furthermore, a normalized position score algorithm is used to calculate the relative position of each on-chain physical state parameter value within the corresponding on-chain physical state reference interval.

[0078] Specifically, if the parameter value falls within the benchmark interval, the position score is zero, indicating it is within the historical normal range. If the parameter value is below the lower bound of the benchmark interval, the position score is negative, equal to the difference between the parameter value and the lower bound, divided by the difference between the parameter's minimum value and the lower bound in the same rolling window of historical data, thus standardizing the degree of deviation below normal. Similarly, if the parameter value is above the upper bound of the benchmark interval, the position score is positive, equal to the difference between the parameter value and the upper bound, divided by the difference between the parameter's maximum value and the upper bound in the same historical window. This calculation process is applied to both the valuation status parameter and the capital structure status parameter, yielding two independent normalized position scores.

[0079] Finally, all relative position scores are aggregated to obtain a comprehensive state offset. Specifically, the aggregation strategy is to take the maximum of the absolute values ​​of the two values. This strategy ensures that when any parameter experiences an extreme shift, the comprehensive state offset will produce a significant response, thereby avoiding false negatives caused by parameter cancellation. Ultimately, each off-chain sentiment state unit will be associated with a comprehensive state offset, which is a standardized scalar value. The larger its absolute value, the more drastic the deviation of the on-chain physical state from its historical norm at the corresponding moment.

[0080] S3.2: The generation of the sequence offset identifier includes: (1) Extract the order of external event trigger parameter types from the off-chain emotion trigger sequence, forming a sequence of length [length missing]. The sequence of events.

[0081] First, extract the event type codes corresponding to all emotional state units from the off-chain emotion trigger sequence to form an original event type sequence.

[0082] Subsequently, a fixed-length sliding window is used to traverse the sequence, thereby extracting a series of continuous, fixed-length event type subsequences. The length of this sliding window is preset to five to seven events, which is sufficient to capture meaningful local sequential patterns while avoiding dilution of the pattern features of recent events due to excessively long sequences.

[0083] (2) Calculate the statistical distribution distance between the event sequence and the historical event subsequences in all historical sequence templates of equal length.

[0084] In this embodiment of the invention, to assess the anomaly of the current event sequence, a historical sequence template library needs to be constructed as a comparison benchmark. This template library consists of all equal-length event type subsequences extracted from historical data. The core of the comparison is to calculate the statistical distribution distance between the current subsequence and all historical templates.

[0085] The specific method is as follows: the event type subsequence is transformed into a probability distribution vector, where each dimension of the vector corresponds to a possible event type, and its value is the frequency of that type in the current subsequence (after Laplace smoothing). The Jensen-Shannon divergence is used as a measure of distribution distance. This divergence is derived from the Kullback-Leibler divergence, possesses symmetry and boundedness, and can quantify the degree of difference between two probability distributions.

[0086] Calculate the Jensen-Shannon divergence between the probability distribution vector of the current subsequence and the probability distribution vector of each historical template, and record the minimum value.

[0087] (3) Calculate the distance between the current subsequence and each historical template, take the minimum value, and take the 95th percentile of the historical distance distribution (e.g., the minimum distance between each historical subsequence and a historical template other than itself) as the threshold. If the minimum value is greater than the threshold, the current subsequence is abnormal. That is, if the minimum distance is greater than a certain upper quantile of the historical distance (e.g., 95%), a trigger order offset flag is generated. If at least one subsequence is detected as being marked as an order offset within the sliding window of the current analysis, a global trigger order offset flag is generated for the entire analysis period.

[0088] It should be noted that this invention transforms the event sequence into a probability distribution and uses Jensen-Shannon divergence for comparison. This is a soft matching method based on statistical distribution. Compared with strict sequence exact matching, it is more tolerant of natural fluctuations in the number of repetitions of events or slight order, while focusing on fundamental changes in the composition structure of event types.

[0089] S3.3: Generating a joint anomaly status identifier includes: When the state offset is in the tail region of its own historical value distribution (e.g., greater than the 95th percentile) and a trigger sequence offset flag is generated, a joint anomaly is determined to have occurred, and a joint anomaly state flag is generated.

[0090] S4: Using the joint abnormal state identifier as input, classify and label the market state within the target time period, and output the market turning point prediction result. The market turning point prediction result is used to indicate whether the market trend continues or reverses.

[0091] It should be noted that, unlike existing predictive models that directly rely on price time series or single indicators, this invention does not attempt to directly model the complex nonlinear mapping between anomalies and prices. Instead, it performs a similarity search and comparison of the currently detected joint anomaly state within a database of similar historical anomalies. By statistically analyzing the subsequent market trends of similar historical events, it infers the most likely trend direction caused by the current anomaly state in probabilistic form. This method effectively and robustly decodes the abstract intermediate signal of the joint anomaly state into a concrete operational indication of trend continuation or trend reversal, thus completing a full closed loop from multimodal data perception to final investment decision support. Specifically: S4.1: Based on the timestamps of the joint abnormal state identifiers appearing within the target time period, the target time period is divided into a joint abnormal state period and a normal period.

[0092] The partitioning process follows these rules: First, the start and end time ranges covered by all joint anomaly status identifiers are merged to form several non-overlapping joint anomaly status periods. The starting point of each anomaly status period is the generation timestamp of the first confirmed joint anomaly status identifier within that period, and the ending point is the generation timestamp of the last joint anomaly status identifier within that period, extended by a preset buffer duration. This buffer duration is used to cover the potential lasting impact of the anomaly signal and is typically set to twice the granularity of on-chain data collection.

[0093] All consecutive intervals within the target time period that are not covered by any joint abnormal state period are classified as normal periods. If there is an ongoing abnormal state period before the start of the target time period, that period is extended to the start of the target time period.

[0094] This division deconstructs the target time period into alternating sequences of joint abnormal and normal states, providing a basis for subsequent differentiated predictions.

[0095] S4.2: For each joint abnormal state period, extract the market price sequence of a fixed time length after the end of the abnormal state period; calculate the linear regression slope of the market price sequence. If it is positive, mark it as rising; if it is negative, mark it as falling, as the actual trend direction label corresponding to the joint abnormal state period.

[0096] Specifically, for a historical joint abnormal state period, the market price sequence of a fixed time length is extracted before the official start.

[0097] This fixed timeframe should be sufficient to characterize short-term trends, for example, set to twice the duration of the anomaly period itself, but not exceeding 5 trading days. Using this historical price series, calculate its linear regression slope using ordinary least squares. If the calculated linear regression slope is positive and passes the statistical significance test, then the reference trend direction label for this anomaly period is marked as the leader of the upward trend period.

[0098] This label does not predict the trend after the abnormal period, but objectively records the trend state of the market before the abnormality occurs, providing a benchmark for subsequent judgment on whether the abnormality strengthens or reverses the previous trend.

[0099] S4.3: For the current joint abnormal state period to be predicted, extract the state offset time series and calculate the feature vector. The feature vector includes at least the mean, variance, skewness and kurtosis.

[0100] In addition, the duration of the abnormal period can be used as a supplementary feature. After the calculation is completed, all features need to be standardized, that is, the mean of the features corresponding to all historical abnormal periods is subtracted and divided by its standard deviation, in order to eliminate the influence of differences in the dimensions and scales of different features on the distance metric, and finally form a standardized multidimensional feature vector.

[0101] It should be noted that by selecting four statistical moments—mean, variance, skewness, and kurtosis—the dynamic distribution characteristics of abnormal state offsets can be described completely and without information redundancy from four dimensions: central tendency, dispersion, distribution symmetry, and distribution shape. Furthermore, the standardization process ensures that the Euclidean distance-based similarity metric has equal weight across all feature dimensions, improving the fairness and accuracy of similarity retrieval.

[0102] S4.4: In historical data, calculate the Euclidean distance between the current feature vector and the feature vector of each historical joint abnormal state period, and select the K historical events with the smallest distance as the set of similar events.

[0103] Euclidean distance is defined as the square root of the sum of the squares of the differences between the corresponding dimensions of two vectors. The smaller the value, the closer the two anomalous periods are in the multidimensional feature space, that is, the more similar their patterns are.

[0104] The value of parameter K is determined through historical backtesting optimization. It is usually an integer between 5 and 20, and its selection needs to strike a balance between the representativeness of the sample and statistical stability.

[0105] S4.5: Statistically analyze the distribution of actual trend direction labels in the set of similar events. If the proportion of rising labels exceeds a preset threshold (e.g., 60%), it is predicted that the current state will continue the upward trend; if the proportion of falling labels exceeds a preset threshold, it is predicted that the trend will reverse and turn downward; if neither exceeds the preset threshold, the prediction is that the trend will continue, that is, the market will continue the dominant trend direction before the abnormal state occurred. This direction needs to be determined independently by analyzing the short-term price trend before the start of the abnormal period.

[0106] The final output is a structured prediction of market turning points. In addition to directional judgments such as trend continuation and strengthening, trend reversal and upward, or trend continuation, it should also include the prediction confidence level calculated based on the proportion, as well as the number of similar historical events K on which it is based, to provide decision-makers with a quantitative assessment of the reliability of the prediction results.

[0107] This invention also includes a market forecasting system based on blockchain data and multimodal fusion, comprising: The on-chain state construction module collects blockchain ledger data corresponding to the target asset, constructs an on-chain physical state dataset based on the blockchain ledger data, and determines the state baseline interval based on historical time windows. The off-chain sentiment sequence module collects off-chain multimodal data associated with the target asset, constructs an off-chain sentiment state dataset based on the off-chain multimodal data, and records the triggering order of the off-chain sentiment state dataset in the time dimension to form an off-chain sentiment trigger sequence. The state order offset module calculates the offset of the on-chain physical state relative to the state baseline interval, and compares the off-chain sentiment trigger sequence with the historical order template to generate an order offset identifier; it then merges the offset and the order offset identifier to generate a joint abnormal state identifier. The joint anomaly prediction module takes joint anomaly status identifiers as input, classifies and labels the market status within the target time period, and outputs the market inflection point prediction results. The market inflection point prediction results are used to indicate whether the market trend continues or reverses.

[0108] The system also includes one or more processors and memory.

[0109] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the market forecasting method based on blockchain data and multimodal fusion described in the foregoing embodiments, especially... Figure 1 The flowchart of the method is shown.

[0110] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of the market forecasting method based on blockchain data and multimodal fusion of the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0111] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0112] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0113] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if necessary, the program can be implemented in assembly or machine language.

[0114] In any case, the language can be either compiled or interpreted.

[0115] Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit.

[0116] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0117] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0118] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0119] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0120] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A market forecasting method based on blockchain data and multimodal fusion, characterized in that: include: Collect blockchain ledger data corresponding to the target asset, construct an on-chain physical state dataset based on the blockchain ledger data, and determine the state baseline interval based on historical time windows; Collect off-chain multimodal data associated with the target asset, construct an off-chain sentiment state dataset based on the off-chain multimodal data, and record the triggering order of the off-chain sentiment state dataset in the time dimension to form an off-chain sentiment triggering sequence; Calculate the offset of the on-chain physical state relative to the state baseline interval, and compare the off-chain emotion trigger sequence with the historical sequence template to generate a sequence offset identifier; fuse the offset and the sequence offset identifier to generate a joint abnormal state identifier; Using the joint abnormal state identifier as input, the market state within the target time period is classified and labeled, and the market inflection point prediction result is output. The market inflection point prediction result is used to indicate whether the market trend continues or reverses.

2. The market forecasting method based on blockchain data and multimodal fusion as described in claim 1, characterized in that: The on-chain physical state dataset includes at least valuation state parameters and capital structure state parameters; the off-chain sentiment state dataset includes at least sentiment intensity parameters and external event triggering parameters.

3. The market forecasting method based on blockchain data and multimodal fusion as described in claim 2, characterized in that: The construction of the on-chain physical state dataset includes: Collect raw on-chain data of the target asset within a historical time window from publicly available blockchain explorer APIs. The raw on-chain data includes at least the number of newly active addresses and transaction records organized in time series, wherein the transaction records include transaction timestamps, transaction amounts, and transaction directions. The raw on-chain data is aggregated and statistically analyzed according to a preset time granularity: for the number of newly added active addresses, the number of new addresses is counted according to the time granularity; for transaction flow, the large transaction amounts transferred into and out of exchanges are summarized according to the time granularity. Based on the aggregated time series data, two on-chain physical state parameters are calculated: the number of newly added active addresses is transformed logarithmically and used as the valuation state parameter; the difference between the amount transferred into and out of the exchange within each time granularity is divided by the total transaction amount of the corresponding time granularity and used as the capital structure state parameter. The construction of the state reference interval includes: For the valuation status parameters and capital structure status parameters, their respective first-order values ​​are calculated using a rolling window within the historical time window. With the Percentiles; the interval formed by two percentiles is defined as the on-chain physical state reference interval for the corresponding parameters.

4. The market forecasting method based on blockchain data and multimodal fusion as described in claim 3, characterized in that: The construction of the off-chain sentiment state dataset includes: Collect off-chain multimodal data associated with the target asset from the preset news and social platform APIs, including at least the news title text and the corresponding publication time; use a pre-trained word vector model to convert each news title text into a corresponding word vector representation to form a text vector sequence; On the time axis, density-based clustering analysis is performed on the text vector sequence, and a minimum cluster size parameter is set to filter noise; Each cluster that meets the minimum cluster size is identified as an external event type, and a unique event code is assigned to each event type as an external event triggering parameter; For each text vector within a time window, a similarity score with its cluster center is calculated. The similarity score is obtained by normalizing the cosine similarity to the [0,1] interval. The weighted average of the similarity scores of all text vectors within the time window is quantized into the emotional intensity parameter at the corresponding moment. Based on the moment when the external event trigger parameter is first identified, the corresponding emotion intensity parameter is associated to form an emotion state unit; All emotional state units are arranged in chronological order to form an off-chain emotional trigger sequence.

5. The market forecasting method based on blockchain data and multimodal fusion as described in claim 4, characterized in that: The calculation of the offset includes: Obtain on-chain physical state parameter values ​​aligned with the timestamp of each emotion state unit in the off-chain emotion triggering sequence; The normalized position score algorithm is used to calculate the relative position of each on-chain physical state parameter value within the corresponding on-chain physical state reference interval. Aggregate all relative position scores to obtain the overall state offset.

6. The market forecasting method based on blockchain data and multimodal fusion as described in claim 5, characterized in that: The generation of the sequence offset identifier includes: The order in which external event trigger parameter types are extracted from the off-chain emotion trigger sequence constitutes a sequence of length [length missing]. The sequence of events; Calculate the statistical distribution distance between the event sequence and the historical event subsequences in all equal-length historical sequence templates; If the minimum distance is greater than a specific upper quantile of the historical distance, a trigger order offset identifier is generated.

7. The market forecasting method based on blockchain data and multimodal fusion as described in claim 6, characterized in that: The generation of the joint abnormal state identifier includes: When the state offset is in the tail region of its own historical value distribution and the trigger sequence offset identifier is generated, a joint anomaly is determined to have occurred, and a joint anomaly state identifier is generated.

8. The market forecasting method based on blockchain data and multimodal fusion as described in claim 7, characterized in that: The process of classifying and labeling the market status within the target time period and outputting the market inflection point prediction results includes: Based on the timestamps of the joint abnormal state identifiers appearing within the target time period, the target time period is divided into a joint abnormal state period and a normal period. For each of the aforementioned joint abnormal state periods, a market price sequence of a fixed time length is extracted after the end of the abnormal period; the linear regression slope of the market price sequence is calculated, and if it is positive, it is marked as rising, and if it is negative, it is marked as falling, which serves as the actual trend direction label corresponding to the aforementioned joint abnormal state period. For the current joint abnormal state period to be predicted, extract the state offset time series and calculate the feature vector, which includes at least the mean, variance, skewness and kurtosis. In historical data, calculate the Euclidean distance between the current feature vector and the feature vector of each historical joint abnormal state period, and select the K historical events with the smallest distance as the set of similar events; The distribution of actual trend direction labels in the set of similar events is statistically analyzed. If the proportion of rising labels exceeds a preset threshold, the current state is predicted to continue the upward trend. If the proportion of falling labels exceeds a preset threshold, the trend is predicted to reverse to a downward trend. If neither exceeds the preset threshold, the trend before the anomaly occurs is output as continuing.

9. A market forecasting system based on blockchain data and multimodal fusion, based on the market forecasting method based on blockchain data and multimodal fusion as described in any one of claims 1 to 8, characterized in that: Also includes: The on-chain state construction module collects blockchain ledger data corresponding to the target asset, constructs an on-chain physical state dataset based on the blockchain ledger data, and determines the state baseline interval based on historical time windows. The off-chain sentiment sequence module collects off-chain multimodal data associated with the target asset, constructs an off-chain sentiment state dataset based on the off-chain multimodal data, and records the triggering order of the off-chain sentiment state dataset in the time dimension to form an off-chain sentiment triggering sequence. The state order offset module calculates the offset of the on-chain physical state relative to the state baseline interval, and compares the off-chain emotion trigger sequence with the historical order template to generate an order offset identifier; it then merges the offset and the order offset identifier to generate a joint abnormal state identifier. The joint anomaly prediction module takes the joint anomaly status identifier as input, classifies and labels the market status within the target time period, and outputs the market inflection point prediction result. The market inflection point prediction result is used to indicate whether the market trend continues or reverses.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the market forecasting method based on blockchain data and multimodal fusion as described in any one of claims 1 to 8.

Images