Futures intelligent decision agent method based on NLP and multi-modal data fusion

Abstract

The invention relates to the technical field of intelligent futures decision, discloses an intelligent futures decision agent method based on NLP and multi-modal data fusion, and aims to solve the problems of representation decoupling, causal chain deficiency, decision response lag and the like caused by separation and shallow fusion of multi-modal data processing in the prior art. According to the method, a multi-modal data acquisition and preprocessing module, a dynamic heterogeneous causal atlas construction module, a causal conduction space-time diagram neural network reasoning module and a futures decision signal generation and interpretation module are adopted, so that a system for uniformly representing dynamic endogenesis of macroscopic events, industrial logic and microcosmic prices is constructed, and cross-modal is realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent decision-making technology for futures trading, and more specifically, relates to an intelligent decision-making agent method for futures trading based on the fusion of NLP and multimodal data. Background Technology

[0002] In the modern financial system, futures, as an important financial derivative, are subject to price fluctuations influenced by a complex interplay of factors, including macroeconomic conditions, industrial policies, and market sentiment. Therefore, accurate prediction and effective decision-making regarding the future trends of the futures market are crucial for risk management and asset allocation for investors and related businesses. With the rapid development of artificial intelligence and big data technologies, utilizing algorithmic models to deeply analyze massive amounts of market data and construct intelligent decision support systems has become a core research direction and cutting-edge practice in the fintech field. The fundamental goal is to leverage the powerful computing capabilities and objective analytical paradigms of machines to overcome the inherent limitations of human traders in terms of the breadth and speed of information processing and emotional control, thereby improving the scientific rigor and timeliness of decision-making.

[0003] In the early stages of this technological evolution, research and applications in this field primarily focused on the quantitative modeling of structured historical data. For example, Chinese patent CN110047001B discloses a typical artificial intelligence analysis method for futures data. The design of this technical solution involves collecting numerical time-series data such as prices, trading volumes, and open interest of historical futures contracts to construct complex mathematical or statistical models, such as the Autoregressive Integrated Moving Average (ARIMA) model, the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model, and its deep learning variants, aiming to uncover potential periodic patterns and volatility characteristics. Based on this, a risk control model is established to assess the current market state and generate trading signals. In specific market environments, this type of method effectively shifts trading decisions from subjective judgment to objective data-driven approaches, providing a solid technical foundation for avoiding irrational emotional decisions caused by drastic market fluctuations. It represents an early exploration and significant achievement in intelligent futures trading.

[0004] However, with the increasing complexity of market participants and the exponential growth in information dissemination speed, the decision-making paradigm relying solely on historical transaction data—a single modality of information—is showing its inherent limitations. Market price fluctuations are not isolated numbers games, but rather a comprehensive reflection of various events and information in the real world. Against this backdrop, technological development in this field has entered a new stage, attempting to introduce a wider range of information sources. A multi-agent collaborative control algorithm for the securities and futures industry, disclosed in Chinese Patent No. CN117788164B, embodies this trend. This scheme introduces a Large Language Model (LLM) and a multi-agent collaborative mechanism, assigning different information sources (such as financial news, research reports, and social media sentiment) to specialized agent modules for analysis. Through game theory or collaboration among agents, a comprehensive decision is ultimately formed. This advancement acknowledges the value of unstructured textual information for market prediction and, by simulating the collaborative working mode of human expert teams, improves the breadth and depth of decision-making to a certain extent.

[0005] While the aforementioned technologies have made significant progress at their respective stages of development, a deeper technological contradiction emerges when we turn our attention to the higher-dimensional needs of real-time market decision-making. This contradiction is not simply a matter of incomplete information coverage or insufficient model accuracy, but rather stems from an inherent flaw in existing technological paradigms when processing heterogeneous data: a general lack of "representation decoupling" and a "missing causal chain." Specifically, whether it's early quantitative models or subsequent multi-agent systems, they essentially treat structured price series data and unstructured textual semantic data as two relatively independent parallel information streams. The former captures the "pattern" of prices through time-series models, while the latter extracts semantic "sentiment" or "theme" through natural language processing techniques. Decision fusion often occurs at the end of the analysis process, such as through simple weighting, voting, or rule engines. This back-end, shallow fusion approach ignores the inherent, deep causal transmission relationships between different modalities of information. For example, the impact of news about a conflict in a specific oil-producing country does not directly and linearly affect all futures contracts. Instead, it propagates along a clear industry chain knowledge graph path: "Oil extraction disruption → expected decline in global supply → rise in crude oil futures prices → increased costs for downstream chemical products (such as PTA) → corresponding changes in PTA futures prices." Existing technological frameworks lack the ability to model such cross-modal and cross-domain entity relationships and causal logic, preventing them from truly "understanding" the market. Models may discover statistical correlations between crude oil news and PTA prices through correlation analysis, but they cannot make forward-looking, structured inferences and predictions based on the inherent industry logic when news emerges. This gap in representation makes the system's response to sudden, structural market shocks often delayed and fragile, hindering truly meaningful, deeply understanding-based predictive decision-making.

[0006] Therefore, designing a novel technical framework that transcends the current model of parallel processing and shallow fusion of heterogeneous data, and constructing a model that can endogenously unify the macro-level event narratives and industry logic carried by natural language with the micro-level price dynamics reflected by structured data, and on this basis, accurately capture and quantify cross-modal causal transmission paths, thereby endowing intelligent agent systems with deep market insights and forward-looking decision-making capabilities, has become a key challenge and an urgent technical problem for those skilled in the art. Summary of the Invention

[0007] The technical problem this invention aims to solve is to overcome the shortcomings of existing technologies that process structured price data and unstructured text data as separate, parallel information streams and perform shallow fusion at the end of the analysis process, resulting in decoupling of representations and missing causal chains. Existing technical frameworks cannot model the inherent, cross-domain entity relationships and causal transmission logic between different modalities of information, leading to delayed and unforeseen decision-making responses to sudden, structural market shocks.

[0008] To achieve the above-mentioned objectives, this invention provides a futures intelligent decision-making agent method based on NLP and multimodal data fusion. This method aims to construct a system that can endogenously unify the macro-event narratives and industry logic carried by natural language and the micro-price dynamics reflected by structured data. On this basis, it can accurately capture and quantify the cross-modal causal transmission path, thereby endowing the intelligent agent system with deep market insight and predictive decision-making capabilities.

[0009] To achieve the above objectives, this invention provides a futures intelligent decision-making agent method based on NLP and multimodal data fusion, comprising the following steps: A preprocessing module is used to acquire and process multimodal data from multiple heterogeneous data sources in parallel. The multimodal data includes at least: structured futures time series data, unstructured text data, and static knowledge graph data describing the relationships between entities in commodity futures and their related industrial chains. A causal graph construction module receives the multimodal data and maps the separated structured futures time series data, unstructured text data and static knowledge graph data into a dynamic, evolving, heterogeneous causal graph containing nodes and edges. The nodes are used to represent commodity entities, driving event entities and concept entities in the market, and the edges are used to represent the causal relationships between the entities. Through a causal neural network inference module, iterative message passing and node state updates are performed on the dynamic heterogeneous causal graph to simulate the transmission process of information impact in the graph, and the final hidden state vector of each commodity entity node is calculated. This final hidden state vector is a unified representation after deep fusion of multimodal information. A futures decision signal generation module generates a trading decision signal for the commodity futures based on the final hidden state vector of each commodity entity node.

[0010] Furthermore, the steps of acquiring and processing multimodal data through the preprocessing module specifically include: A structured data processing unit connects to the data distribution system of the futures exchange to acquire structured futures time series data of futures contracts in real time. The data fields of the structured futures time series data include opening price, highest price, lowest price, closing price, trading volume, open interest, and basis. The data is then cleaned to remove invalid data points and outliers. Subsequently, the data is normalized using the Z-score standardization method. Finally, the data is cut into fixed-length data panes according to a preset time step. Through an unstructured data processing unit, text information such as financial news, policy documents, and industry research reports is selectively crawled using web crawler programs or news information API interfaces. The raw data is then parsed to remove irrelevant content, and a natural language processing toolkit is applied to perform preprocessing on the plain text information, including sentence segmentation, lexicalization, and stop word removal. A knowledge graph data processing unit loads and verifies a pre-built static knowledge graph stored in the form of resource description framework triples. This static knowledge graph provides the basic industrial logic framework for the subsequent construction of dynamic knowledge graphs.

[0011] Furthermore, the step of constructing a dynamic heterogeneous causal graph through the causal graph construction module specifically includes: Through an extraction unit, using a pre-trained language model based on the Transformer architecture, parallel named entity recognition and event extraction tasks are performed on the pre-processed unstructured text data to identify futures commodity entities, regional entities, organizational entities, as well as trigger words and related arguments describing market-driven events in the text, and assign unique identifiers to all identified entities and events. An initial feature vector is generated for the nodes in the graph through a node initialization unit. The nodes are divided into three categories: product nodes, event nodes, and concept nodes. A static edge and a dynamic edge generation unit are used to establish static edges and dynamic edges between the initialized nodes to form the topological structure of the dynamic heterogeneous causal graph.

[0012] Furthermore, the step of generating initial feature vectors for nodes in the graph through the node initialization unit is as follows: For a product node, the latest data pane corresponding to it, generated by the structured data processing unit, is input into a one-dimensional convolutional neural network. The one-dimensional convolutional neural network includes at least one convolutional layer, an activation function layer, and a pooling layer. By performing convolution and pooling operations on the data pane, a fixed-length vector is finally output as the initial state embedding of the product node. The initial state embedding encodes the recent price pattern and transaction dynamics of the product. For each event node, the corresponding original sentence text is input into the pre-trained language model to extract semantic base vectors. At the same time, the extracted event type is converted into a one-hot encoded vector, and the extracted quantization index is numerically normalized to form a quantization index vector. Finally, the semantic base vector, the one-hot encoded vector, and the normalized quantization index vector are concatenated to form a comprehensive initial feature vector for the event node. For concept nodes, which come from the static knowledge graph, their feature vectors are initialized by finding the vectors of the corresponding words through a word vector model trained on a large-scale general corpus, or by setting them as a learnable embedding vector.

[0013] Furthermore, the static and dynamic edge generation unit (the steps for establishing static and dynamic edges are as follows:) The static edges are directly established based on the entity relationships in the static knowledge graph loaded by the knowledge graph data processing unit, and are connected between the corresponding product nodes or concept nodes. A learnable weight parameter that is optimized during model training is set for each static edge. The dynamic edge is generated in real time based on the output of the extraction unit. When a new event node is created, a directed dynamic edge is established between the event node and all product nodes that are identified as affected objects in the event text that generated the event node. The weights of the dynamic edges are dynamically generated through a dedicated weight calculation network, which is a multilayer perceptron. Its input is the concatenated initial feature vector of the event node and the initial state embedding of the corresponding product node. Its output is a scalar value between 0 and 1 constrained by the Sigmoid activation function, which is used to quantify the direct influence of the event corresponding to the event node on the product corresponding to the product node.

[0014] Furthermore, the reasoning steps performed by the causal neural network reasoning module are essentially achieved by using a gated graph sequence neural network to update the hidden state of all nodes in the graph through iterative message passing with a preset number of steps K.

[0015] Furthermore, the node hidden state update process in each propagation step k of the gated graph sequence neural network specifically includes: Message aggregation calculation: For any node v in the graph, the hidden states of all its neighboring nodes u in step k-1 are aggregated using an attention-based aggregation function; where the importance of the message passed from node u to node v is calculated by an attention network, which calculates the attention weight α based on the hidden states of nodes v and u in step k-1. vuThe aggregated message is the weighted sum of the hidden states of all neighboring nodes according to their corresponding attention weights; Node state update: Taking the aggregated message as input and the hidden state of node v in step k-1 as the initial hidden state, the data is fed into a gated recurrent unit (GRU) for calculation. The output of the GRU is the new hidden state h of node v in step k. v (k) .

[0016] Furthermore, the step of generating trading decision signals through the futures decision signal generation module (400) is as follows: A decision signal generation unit generates the final hidden state vector h of each product node output by the causal neural network inference module. v (k) Each input is an independent feedforward decision network; the decision network is a multilayer perceptron, whose output layer has multiple neurons corresponding to the "buy", "hold", and "sell" operations respectively, and applies the Softmax activation function to output the probability of each operation; the operation with the highest probability value is selected as the final transaction decision signal for the commodity.

[0017] Furthermore, the futures decision signal generation module further includes a decision interpretation unit, the function of which is: During the reasoning process of the causal neural network reasoning module, all calculated attention weights α are recorded and stored; When the decision signal generation unit outputs a "buy" or "sell" decision signal, it traces back from the commodity node corresponding to the decision signal in the dynamic heterogeneous causal graph. Using the recorded attention weights, find the top N causal transmission paths that contribute the most to the final decision. The causal transmission path is a sequence of nodes and edges that starts from a source event node and goes through multiple transmission steps to finally reach the target product node. The identified high-weight causal transmission paths are transformed into a structured natural language description to clearly explain the source events and industry transmission logic that lead to the decision-making recommendations.

[0018] This invention also proposes a futures intelligent decision-making agent system, comprising: A preprocessing module is configured to acquire and process multimodal data from multiple heterogeneous data sources in parallel. The multimodal data includes at least: structured futures time series data, unstructured text data, and static knowledge graph data describing the relationships between entities in commodity futures and their related industrial chains. A causal graph construction module is connected to the preprocessing module and is configured to receive the multimodal data, and to uniformly map and construct the separated data into a dynamically evolving, heterogeneous causal graph containing nodes and edges. The nodes are used to represent commodity entities, driving event entities and concept entities in the market, and the edges are used to represent the causal relationships between the entities. A causal neural network inference module is connected to the causal graph construction module and is configured to perform iterative message passing and node state updates on the dynamic heterogeneous causal graph to simulate the transmission process of information impact in the graph and calculate the final hidden state vector of each commodity entity node. A futures decision signal generation module is connected to the causal neural network inference module and is configured to generate a trading decision signal for the commodity futures based on the final hidden state vector of each commodity entity node.

[0019] Terminology Explanation: NLP stands for Natural Language Processing.

[0020] Beneficial Effects: Through the collaborative work of the modules described above, the method described in this invention fundamentally changes the way financial decision-making models process multimodal data. It no longer simplifies textual information into one or a few external factors, but instead seamlessly unifies the semantics of events, the structure of industries, and the dynamics of prices into a single representation space by constructing a dynamic, heterogeneous, and internally logical causal graph. Through graph neural network inference performed on this spatiotemporal graph, the system can explicitly model and quantify the transmission process of information shocks along the causal chain, thereby making truly forward-looking decisions based on deep market understanding. This significantly improves the accuracy, timeliness, and interpretability of decisions, effectively solving the key problems existing in the background technology. Attached Figure Description

[0021] Figure 1 The functional module block diagram of the futures intelligent decision-making agent system provided by the present invention.

[0022] Figure 2 This is a flowchart illustrating the intelligent decision-making agent method for futures trading provided by the present invention.

[0023] Figure 3 for Figure 1 Internal structure diagram of the preprocessing module.

[0024] Figure 4 This is a schematic diagram of the structure of the dynamic heterogeneous causal graph in this invention.

[0025] Figure 5 This is a schematic diagram illustrating the principle of the causal transmission spatiotemporal graph neural network inference process in this invention.

[0026] The reference numerals in the attached figures are as follows: 100, Preprocessing module; 110, Structured data processing unit; 120, Unstructured data processing unit; 130, Knowledge graph data processing unit; 200, Causal graph construction module; 210, Extraction unit; 220, Node initialization unit; 230, Static and dynamic edge generation unit; 300, Causal neural network inference module; 400, Futures decision signal generation module; 410, Decision signal generation unit; 420, Decision interpretation unit. Detailed Implementation

[0027] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention will be further described below with reference to specific figures and embodiments. It should be noted that these embodiments are only used to illustrate the technical solutions of this invention and not to limit it. Those skilled in the art can make various modifications or equivalent substitutions to this invention. Any simple modifications, equivalent changes, and alterations made based on the essential spirit of this invention still fall within the scope of the technical solutions of this invention.

[0028] Please see Figures 1 to 5 This invention provides a futures intelligent decision-making agent method based on NLP and multimodal data fusion. This method is implemented through an integrated, multi-module collaborative system architecture, which can be physically deployed on one or more high-performance server hardware. In one specific embodiment, the server hardware is configured with two Intel Xeon Gold 6248R CPUs, 512 gigabytes of DDR4 error-correcting memory, four NVIDIA A100 GPUs with 80 gigabytes of dedicated video memory, and a storage array consisting of 20 terabytes of NVMe SSDs. At the software level, the system runs on Ubuntu 20.04 Long Term Support operating system, uses PyTorch 1.10 as the deep learning framework, relies on the PyTorch Geometric 2.0 library for graph neural network computations, implements natural language processing tasks using the Hugging Face Transformers 4.15 library, and manages the storage and retrieval of knowledge graph data using the Neo4j 4.4 Enterprise Edition graph database.

[0029] like Figure 1 As shown, the system architecture of this invention is logically divided into four core functional modules: a preprocessing module 100, a causal graph construction module 200, a causal neural network inference module 300, and a futures decision signal generation module 400. These modules are connected in series to form a complete data processing, analysis, inference, and decision-making process, as follows: Figure 2 As shown.

[0030] As a further explanation of this embodiment, firstly, the function of the preprocessing module 100 is to provide standardized and normalized data input for subsequent in-depth analysis. For example... Figure 3 As shown, this module contains three parallel processing units.

[0031] As a further explanation of this embodiment, specifically, the structured data processing unit 110 directly connects to the data distribution system of the futures exchange, such as the data interface of the Chicago Mercantile Exchange (CME) or the Shanghai Futures Exchange (SHFE), through a standard financial information exchange protocol (FIX Protocol version 4.4) or the WebSocket application programming interface provided by the exchange. This unit acquires structured futures time series data of a specific futures contract in real time at millisecond-level frequency. The raw data fields of the acquired structured futures time series data precisely correspond to: opening price, high price, low price, closing price, volume, open interest, and basis. The acquired raw data stream first undergoes a cleaning process. This process filters out data points outside the exchange's legal trading hours using timestamps and filters out significant price gaps caused by data transmission errors using a sliding window outlier detection algorithm based on the 3-sigma principle. Subsequently, the cleaned data is normalized. Specifically, the Z-score normalization method is adopted. For each independent feature sequence (such as a closing price sequence), its mean μ and standard deviation σ are calculated within a sliding window of 240 time units. Each data point x within the window is then transformed using the formula z=(x-μ) / σ. This aims to eliminate the adverse effects of the price dimension differences between different futures contracts (e.g., crude oil prices are tens of dollars per barrel, while rebar prices are thousands of RMB per ton) on the subsequent model training weights. Finally, the normalized seven-dimensional time series data is cut into fixed-length data panes according to a preset time step (e.g., 60 consecutive minute-level data points). Each data pane forms a two-dimensional tensor of shape [60, 7], which serves as an independent structured data sample, is labeled with the corresponding timestamp, and is transmitted to subsequent modules.

[0032] As a further illustration of this embodiment, the unstructured data processing unit 120 is responsible for continuously crawling text information related to the futures market from the Internet. This unit deploys multiple highly customized web crawlers and integrates paid API interfaces from several professional news and information providers (such as Bloomberg and Reuters). The crawling targets cover mainstream global financial media websites, policy documents issued by government departments of various countries, research reports published by major industry associations and research institutions, and information published by verified expert accounts with industry influence on specific social media platforms (such as Twitter). The crawled raw data consists of web pages or JSON objects containing HTML tags, advertising scripts, and other irrelevant elements. The unit first applies a parser based on the lxml library to precisely remove all non-text content, retaining only the plain text information and its core metadata, including the title, publication time accurate to the second, and the URL of the information source. Subsequently, a natural language processing pipeline based on the spaCy toolkit is applied to preprocess the plain text. The pipeline performs the following operations sequentially: sentence segmentation, breaking down long paragraphs into independent sentence units; tokenization, decomposing sentences into sequences of lexical units that conform to the grammatical rules of a specific language (such as English or Chinese); and stop word removal, a step that uses an expanded custom stop word list containing a large number of meaningless words specific to the financial field (such as "it is reported" and "the author believes") to minimize information noise. The processed text data and corresponding metadata are structured and stored in a PostgreSQL database, forming a continuously growing corpus of text to be analyzed.

[0033] As a further illustration of this embodiment, the preprocessing module 100 further includes a knowledge graph data processing unit 130. This unit is responsible for loading and maintaining a static knowledge graph pre-constructed by domain experts, describing the intrinsic relationships between entities in global commodities and their related industrial chains. This knowledge graph uses Resource Description Framework (RDF) triples (subject-entity, predicate-relationship, object-entity) for logical expression and is physically stored in the aforementioned Neo4j graph database. The entity nodes in the graph cover all key futures commodities, their direct upstream and downstream products, important production enterprises, and key production factors. Relationship edges precisely define the logical connections between these entities. For example, the triple <crude oil, upstream product, polyethylene terephthalate (PTA)> explicitly expresses that crude oil is an upstream chemical raw material for PTA, and <Brazil, main producing area, soybean> defines the relationship between region and agricultural product. This unit loads the entire knowledge graph upon system startup and runs a series of validation rules based on SHACL (Shapes Constraint Language) to ensure the integrity of the graph's topological structure, the absence of dangling nodes, and the global uniqueness of all entity names. This static knowledge graph provides a foundational, time-invariant industrial logic framework for the subsequent construction of dynamic knowledge graphs.

[0034] As a further explanation of this embodiment, the causal graph construction module 200 then receives the data processed by the aforementioned module 100. Its core task is to map the separated, multimodal information into a dynamically evolving heterogeneous graph structure that endogenously represents market entities, driving events, and the causal relationships between them, such as... Figure 4 As shown.

[0035] Specifically, module 200 includes an extraction unit 210. At its core is a pre-trained bidirectional encoder model based on the Transformer architecture, such as a BERT-based model fine-tuned on a proprietary corpus containing over five million financial documents. For each piece of preprocessed text information (such as a sentence) input from the unstructured data processing unit 120, the extraction unit 210 performs two parallel sequence labeling tasks. The first task is Named Entity Recognition (NER), which aims to accurately identify predefined entity categories from the text, including commodity futures entities (COMM), such as "soybeans" and "iron ore"; geographical entities (GEO), such as "Brazil" and "Middle East"; and organizational entities (ORG), such as "OPEC" and "Federal Reserve". The second task is Event Extraction (EE), which aims to identify trigger words describing key market-driving events and their associated arguments. For example, given the text "OPEC+ announced at its Riyadh headquarters that it will further reduce crude oil production by 2 million barrels per day to stabilize oil prices," the event extraction task will identify the event type as "supply reduction," the event subject as "OPEC+" (ORG), the affected object as "crude oil" (COMM), the quantitative indicator as "2 million barrels per day," and the location as "Riyadh headquarters" (GEO). All successfully identified entities and events are assigned a globally unique identifier (UUID) and stored in a temporary cache along with their source text, timestamp, and other metadata.

[0036] Subsequently, a node initialization unit 220 is responsible for generating the initial feature vector for each type of node in the graph, i.e., its hidden state at time zero. Node types are strictly divided into three categories: Commodity Nodes, Event Nodes, and Concept Nodes, such as... Figure 4 Indication.

[0037] As a further explanation of this embodiment, for a product node, such as "crude oil," the initialization process is as follows: The latest time pane data corresponding to this product, generated by the structured data processing unit 110, is retrieved, i.e., a tensor with dimensions [60, 7]. This tensor is input into a specially designed one-dimensional convolutional neural network (1D-CNN). The architecture of this 1D-CNN is as follows: the input layer is followed by a first convolutional layer containing 64 filters, with a kernel size of 3, a stride of 1, and padding of 1, followed by a Rectified Linear Unit (ReLU) activation function and a max-pooling layer with a kernel size of 2; the second convolutional layer contains 128 filters, with the same kernel size, stride, padding, activation function, and pooling layer configuration as the first layer; the third convolutional layer contains 256 filters, configured as above. Finally, through a global max-pooling operation, the output feature map of the third convolutional layer is compressed into a 256-dimensional fixed-length vector. This vector represents the initial state embedding of the commodity node at the current moment, highly condensed and encoding the recent price fluctuations, trading activity, and open interest trends of the commodity. For an event node, the initialization process is more complex, aiming to integrate semantic, typological, and quantitative information. First, the original sentence text triggering the event is input into the finely tuned BERT model used by extraction unit 210, and the output vector corresponding to the classification label ([CLS]) in its final hidden layer is extracted, resulting in a 768-dimensional semantic base vector. Simultaneously, the event type identified in the event extraction task (e.g., "supply decrease") is queried from a pre-defined type dictionary and converted into an N-dimensional one-hot encoded vector (N being the total number of event types, e.g., N=20). Furthermore, the extracted quantitative indicators (e.g., "2 million barrels") are numericalized using a pre-defined dimensional table and min-max normalized to the [0, 1] interval. Finally, the 768-dimensional semantic vector, the 20-dimensional type vector, and the 1-dimensional normalized quantization index vector are concatenated to form a comprehensive initial feature vector for the event node with a total dimension of 789.

[0038] As a further explanation of this embodiment, for concept nodes, which originate from the static knowledge graph data processing unit 130, such as "inflation" or a specific chemical process, their feature vectors are initialized by finding the 300-dimensional vector of the corresponding word using a word vector model (e.g., the GloVe.840B.300d model) pre-trained on a large-scale general corpus. If a concept in the graph has no corresponding entry in the word vector model, its feature vector is set to a learnable, randomly initialized embedding vector.

[0039] Next, a static edge and dynamic edge generation unit 230 is responsible for establishing connections between the initialized nodes, i.e., the edges of the graph. The types of edges are divided into static edges and dynamic edges.

[0040] Static edges are directly derived from the static knowledge graph loaded by the knowledge graph data processing unit 130. For example, if the knowledge graph contains the triple <crude oil, is an upstream product, PTA>, then during the graph construction process, the system will establish a directed static edge from "crude oil" to "PTA" between the "crude oil" commodity node and the "PTA" commodity node. The initial weights of all static edges are set to a learnable parameter sampled from a uniform distribution U(-0.1, 0.1), which will be optimized through backpropagation during the subsequent end-to-end training of the model.

[0041] Dynamic edges are generated in real time based on the output of extraction unit 210. When a new event node is created, such as the aforementioned "OPEC+ production cut" event, the static and dynamic edge generation units 230 establish directed dynamic edges between the event node and all product nodes identified as "affected objects" in the event text that generated the event node, with the direction pointing from the event node to the product node. The weights of the dynamic edges are not fixed values, but are dynamically generated through a dedicated, lightweight weight calculation network. This network is a two-layer multilayer perceptron (MLP), whose inputs are the concatenated initial feature vector of the event node (789 dimensions) and the initial state embedding of the corresponding product node (256 dimensions), for a total input dimension of 1045. The hidden layer of this MLP has a dimension of 128, the activation function is ReLU, and the output layer is a single neuron, with the Sigmoid activation function constraining the output value between 0 and 1. This output scalar value is the weight of the dynamic edge, which dynamically and contextually quantifies the direct impact strength of the event corresponding to the event node on the product corresponding to the product node.

[0042] As a further illustration of this embodiment, a dynamic heterogeneous causal graph integrating price dynamics, event semantics, and industry logic has now been fully constructed. The topology (presence or absence of nodes and edges) and node characteristics of this graph are updated in real time as new structured and unstructured data streams enter, forming a digital twin snapshot of the market.

[0043] Next, the causal neural network inference module 300 receives this dynamically constructed graph and performs core causal inference calculations. This module aims to simulate how information shocks (typically introduced by high-weight event nodes) are transmitted, attenuated, amplified, and evolved through the graph's topology (defined by static and dynamic edges), ultimately comprehensively affecting the state of each commodity node.

[0044] The core of this module is a gated graph sequence neural network (GGS-NN), whose information transmission process is as follows: Figure 5 As shown, this network propagates information among all nodes in the graph through an iterative message-passing mechanism. This process is defined as multi-step propagation, iteratively executing a preset number of steps K. The hyperparameter K corresponds physically to the length of the maximum causal propagation chain to be modeled, and its value is determined experimentally on a validation set. For example, setting it to 3 means that the system can capture the indirect effects triggered by an event, after propagation through at most two intermediate entities.

[0045] At each propagation step k (from 1 to K), each node v in the graph updates its hidden state h based on the information of its neighboring nodes. v (k) The update process is as follows: First, for node v, it needs to compute an aggregate message av in step k. (k) This message is obtained by aggregating the hidden states h of all its neighboring nodes u (u∈N(v)) in step k-1. u (k-1) The aggregation function employs an attention-based weighted summation method to adaptively evaluate the importance of information from different neighbors. Specifically, the importance of the message passed from node u to node v is represented by the attention weight α. vu This is calculated by an attention network. The calculation process of this attention network follows the mechanism of Graph Attention Network (GAT): first, an attention coefficient is calculated. Where LeakyReLU is the activation function, W is a learnable linear transformation matrix, and a is a learnable attention vector. Let || denote the transpose of the parameter vector. Then, the final attention weights are obtained by performing Softmax normalization on the attention coefficients of all neighboring nodes of node v. Where softmax is the normalization function, for Normalize along the dimension of node u. Aggregate message a. v (k) This refers to the hidden state of all neighboring nodes u after the W transformation, based on their corresponding attention weights α. vu Weighted sum: .

[0046] Secondly, the new hidden state h of node v v (k)It is computed using a gated recurrent unit (GRU). This design borrows the idea of ​​recurrent neural networks for processing sequential information, treating multi-step propagation on a graph as a time-evolutionary process. The input to the GRU unit is the aggregated message a computed in the current step k. v (k) Its internal hidden state is determined by the hidden state h of node v in the previous step k-1. v (k-1) Initialization is performed. The GRU's update and reset gates adaptively determine how much old information (from the node's previous state) to retain and how much new information (aggregated messages from neighboring nodes) to absorb, thus achieving smooth and efficient information fusion. The node's initial hidden state h... v (0) It is set as the initial feature vector generated in node initialization unit 220.

[0047] The iterative process of message passing and node state updates will repeat K=3 steps. After 3 steps of propagation, the final hidden state h of each product node will be... v (3) It no longer merely encodes its own historical price information. More importantly, it has deeply integrated information from multi-hop neighbors through the graph's messaging mechanism, including semantic information of directly related emergencies, as well as the influence of other goods and events indirectly transmitted through upstream and downstream relationships in the industry chain (static edges). This final hidden state vector with a dimension of 256 is the endogenous unified representation that this invention seeks to achieve by deeply integrating multimodal information.

[0048] Finally, the futures decision signal generation module 400 generates specific and actionable trading decision suggestions based on the final node representation output by the aforementioned causal neural network reasoning module 300, and provides a basis for decision-making.

[0049] This module includes a decision signal generation unit 410. For each commodity node v that needs to make a decision (e.g., "crude oil" or "PTA"), its final hidden state vector h v (K)The input is fed into a separate, feedforward decision network. This network is a multilayer perceptron (MLP) with two hidden layers. Specifically, the input layer has a dimension of 256, the first hidden layer has 128 neurons, and the second hidden layer has 64 neurons, both using the ReLU activation function. The output layer has three neurons, using the Softmax activation function. The output values ​​of these three neurons correspond to the probabilities of taking a "Buy," "Hold," or "Sell" action on the commodity within the next preset trading period (e.g., the next 4 hours). The system selects the action with the highest probability as the final decision signal output; for example, the output might be {Commodity: "PTA", Decision: "Buy", Confidence: 0.85}.

[0050] In a preferred embodiment of the present invention, the module further includes a decision interpretation unit 420 to enhance the transparency and credibility of the entire decision-making process. In each message passing iteration of the causal neural network inference module 300, all calculated attention weights α are recorded and stored. When the decision signal generation unit 410 outputs a non-"holding" decision signal (i.e., "buy" or "sell"), the decision interpretation unit 420 starts from the corresponding product node and backtracks through the dynamic heterogeneous causal graph. This unit utilizes the recorded attention weights and employs a greedy algorithm to find the top N (e.g., N=3) causal transmission paths that contribute the most to the final decision. A causal path is a sequence of nodes and edges that starts from a source node (usually an event node), proceeds through K steps, and finally reaches the target product node. The backtracking algorithm selects the upstream neighbor node with the highest attention weight at each step. For example, for the decision to "buy PTA", the system might find a high-weighted path like this: "[Event node: ID-E123, text: 'Conflict escalation'] - (dynamic edge, weight 0.91) -> [commodity node: crude oil] - (static edge, relationship: upstream) -> [commodity node: PTA]". The decision interpretation unit 420 then converts these high-weighted causal paths into a structured, human-readable description using a pre-defined natural language generation template. For example, the output interpretation might be: "Decision signal: Buy PTA futures. Attribution analysis: A high-impact event 'conflict escalation' was detected. This impact, through the upstream cost transmission path 'crude oil -> PTA', is expected to push up PTA production costs, thus exerting significant upward pressure on PTA futures prices. The contribution score of this path to the decision is 9.2 / 10." Example: To verify the effectiveness of the method provided by this invention, this example selects a real historical event—the outbreak of a conflict on February 24, 2022—for backtesting analysis. This event had a severe impact on the global energy, food, and some industrial product markets.

[0051] In the lead-up to the event, the preprocessing module 100 of this invention continuously monitored global news. Around 3:00 AM GMT on February 24, the unstructured data processing unit 120 captured a large amount of breaking news about "a certain conflict" from multiple sources, including Reuters and the Associated Press, via an API interface. Simultaneously, the structured data processing unit 110 recorded pre-market price data for relevant futures contracts such as European natural gas (TTF), Brent crude oil, wheat, and aluminum.

[0052] After receiving the text data, the causal graph construction module 200 immediately extracts events of type "a certain conflict" and identifies commodity entities such as "natural gas" and "crude oil" as potential affected objects. A high-dimensional event node is created and initialized. Simultaneously, the static and dynamic edge generation unit 230 establishes high-weight dynamic edges between this event node and the commodity nodes "Brent crude oil" and "European natural gas". Furthermore, pre-stored relationships in the static knowledge graph, such as <natural gas, a factor of production, aluminum> and <natural gas, a factor of production, nitrogen fertilizer>, ensure that static edges exist between the corresponding commodity nodes.

[0053] After receiving the updated graph, the causal neural network inference module 300 begins inference. In the first step of propagation (K=1), a strong signal from the "certain conflict" event node significantly affects the hidden states of the "European natural gas" and "Brent crude oil" nodes through high-weight dynamic edges. In the second step of propagation (K=2), the updated state information of the "European natural gas" node is further propagated to the "aluminum" node and the "nitrogen fertilizer" concept node, which is a key element in wheat production, through static edges.

[0054] Finally, based on the final hidden state of each commodity node after three steps of transmission, the futures decision signal generation module 400 outputs the following decision signals before the market price begins to fluctuate sharply: 1. Decision signal: Buy Brent crude oil futures. Confidence level: 0.95.

[0055] 2. Decision signal: Buy European natural gas futures. Confidence level: 0.98.

[0056] 3. Decision signal: Buy wheat futures. Confidence level: 0.88.

[0057] 4. Decision signal: Buy aluminum futures. Confidence level: 0.82.

[0058] Decision Explanation Unit 420 provides the following explanation for the decision to "buy aluminum futures": "Attribution analysis: A high-impact event 'a certain conflict' is detected, and this impact is expected to push up aluminum futures prices through the causal transmission path of 'conflict -> expected surge in natural gas prices -> increased aluminum production costs (electrolytic aluminum)'." As can be clearly seen from the above embodiments, the futures intelligent decision-making agent method based on NLP and multimodal data fusion described in this invention, by constructing and reasoning on a dynamic heterogeneous causal graph, can not only make fast and accurate decisions for commodities directly affected by events, but more importantly, it can predict and quantify the transmission effect of events along the industrial chain through explicit modeling of industrial logic, thereby making forward-looking and highly commercially valuable decisions for indirectly affected commodities, while providing clear decision-making basis, demonstrating significant technical superiority and practical value.

[0059] The above are merely specific embodiments of the present invention, but the technical features of the present invention are not limited thereto. Any simple changes, equivalent substitutions, or modifications made based on the present invention to solve essentially the same technical problems and achieve essentially the same technical effects are all covered within the protection scope of the present invention.

Claims

1. A futures intelligent decision-making agent method based on NLP and multimodal data fusion, characterized in that, Includes the following steps: A preprocessing module (100) acquires and processes multimodal data from multiple heterogeneous data sources in parallel. The multimodal data includes at least: structured futures time series data, unstructured text data, and static knowledge graph data describing the relationships between entities in commodity futures and their related industrial chains. A causal graph construction module (200) receives the multimodal data and maps the separated structured futures time series data, unstructured text data and static knowledge graph data into a dynamic, evolving, heterogeneous causal graph containing nodes and edges. The nodes are used to represent commodity entities, driving event entities and concept entities in the market, and the edges are used to represent the causal relationships between the entities. Through a causal neural network inference module (300), iterative message passing and node state updates are performed on the dynamic heterogeneous causal graph to simulate the transmission process of information impact in the graph, and the final hidden state vector of each commodity entity node is calculated. The final hidden state vector is a unified representation after deep fusion of multimodal information. A futures decision signal generation module (400) generates a trading decision signal for the commodity futures based on the final hidden state vector of each commodity entity node.

2. The futures intelligent decision-making agent method based on NLP and multimodal data fusion as described in claim 1, characterized in that, The steps of acquiring and processing multimodal data through the preprocessing module (100) specifically include: A structured data processing unit (110) is connected to the data distribution system of the futures exchange to obtain structured futures time series data of futures contracts in real time. The data fields of the structured futures time series data include opening price, highest price, lowest price, closing price, trading volume, open interest and basis. The data is then cleaned to remove invalid data points and outliers. Subsequently, the data is normalized using the Z-score standardization method. Finally, the data is cut into fixed-length data panes according to a preset time step. Through an unstructured data processing unit (120), text information such as financial news, policy documents and industry research reports is selectively crawled through a web crawler program or news information API interface; the raw data is parsed to remove irrelevant content, and then a natural language processing toolkit is applied to perform a preprocessing process on the plain text information, including sentence segmentation, lexicalization and removal of stop words. A knowledge graph data processing unit (130) loads and verifies a pre-built static knowledge graph stored in the form of resource description framework triples, which provides the basic industrial logic framework for the construction of subsequent dynamic graphs.

3. The futures intelligent decision-making agent method based on NLP and multimodal data fusion according to claim 2, characterized in that, The steps of constructing a dynamic heterogeneous causal graph through the causal graph construction module (200) specifically include: Through an extraction unit (210), a pre-trained language model based on the Transformer architecture is used to perform parallel named entity recognition and event extraction tasks on the pre-processed unstructured text data, so as to identify futures commodity entities, regional entities, organizational entities, as well as trigger words and related arguments describing market-driven events in the text, and assign unique identifiers to all identified entities and events. An initial feature vector is generated for the nodes in the graph through a node initialization unit (220), and the nodes are divided into three categories: commodity nodes, event nodes and concept nodes. A static edge and a dynamic edge generation unit (230) are used to establish static edges and dynamic edges between the initialized nodes to form the topological structure of the dynamic heterogeneous causal graph.

4. The futures intelligent decision-making agent method based on NLP and multimodal data fusion according to claim 3, characterized in that, The step of generating initial feature vectors for nodes in the graph through the node initialization unit (220) is as follows: For a product node, the latest data pane corresponding to it, generated by the structured data processing unit (110), is input into a one-dimensional convolutional neural network. The one-dimensional convolutional neural network includes at least one convolutional layer, an activation function layer, and a pooling layer. By performing convolution and pooling operations on the data pane, a fixed-length vector is finally output as the initial state embedding of the product node. The initial state embedding encodes the recent price pattern and transaction dynamics of the product. For each event node, the corresponding original sentence text is input into the pre-trained language model to extract semantic base vectors. At the same time, the extracted event type is converted into a one-hot encoded vector, and the extracted quantization index is numerically normalized to form a quantization index vector. Finally, the semantic base vector, the one-hot encoded vector, and the normalized quantization index vector are concatenated to form a comprehensive initial feature vector for the event node. For concept nodes, which come from the static knowledge graph, their feature vectors are initialized by finding the vectors of the corresponding words through a word vector model trained on a large-scale general corpus, or by setting them as a learnable embedding vector.

5. The futures intelligent decision-making agent method based on NLP and multimodal data fusion according to claim 3, characterized in that, The steps for establishing static and dynamic edges through the static and dynamic edge generation unit (230) are as follows: The static edges are directly established based on the entity relationships in the static knowledge graph loaded by the knowledge graph data processing unit (130), and are connected between the corresponding product nodes or concept nodes. A learnable weight parameter that is optimized during model training is set for each static edge. The dynamic edge is generated in real time based on the output of the extraction unit (210). When a new event node is created, a directed dynamic edge is established between the event node and all product nodes that are identified as affected objects in the event text that generated the event node. The weights of the dynamic edges are dynamically generated through a dedicated weight calculation network, which is a multilayer perceptron. Its input is the concatenated initial feature vector of the event node and the initial state embedding of the corresponding product node. Its output is a scalar value between 0 and 1 constrained by the Sigmoid activation function, which is used to quantify the direct influence of the event corresponding to the event node on the product corresponding to the product node.

6. The futures intelligent decision-making agent method based on NLP and multimodal data fusion according to claim 1, characterized in that, The reasoning steps are performed by the causal neural network reasoning module (300). The core of this method is to use a gated graph sequence neural network to update the hidden state of all nodes in the graph by performing iterative message passing for a preset number of steps K.

7. The futures intelligent decision-making agent method based on NLP and multimodal data fusion according to claim 6, characterized in that, The hidden state update process of the nodes in each propagation step k of the gated graph sequence neural network specifically includes: Message aggregation calculation: For any node v in the graph, the hidden states of all its neighboring nodes u in step k-1 are aggregated using an attention-based aggregation function; where the importance of the message passed from node u to node v is calculated by an attention network, which calculates the attention weight α based on the hidden states of nodes v and u in step k-1. vu The aggregated message is the weighted sum of the hidden states of all neighboring nodes according to their corresponding attention weights; Node state update: Taking the aggregated message as input and the hidden state of node v in step k-1 as the initial hidden state, the data is fed into a gated recurrent unit (GRU) for calculation. The output of the GRU is the new hidden state h of node v in step k. v (k) .

8. The futures intelligent decision-making agent method based on NLP and multimodal data fusion according to claim 1, characterized in that, The steps for generating trading decision signals through the futures decision signal generation module (400) are as follows: The final hidden state vector h of each commodity node output by the causal neural network inference module (300) is generated by a decision signal generation unit (410). v (k) Each input is an independent feedforward decision network; the decision network is a multilayer perceptron, whose output layer has multiple neurons corresponding to the "buy", "hold" and "sell" operations respectively, and applies the Softmax activation function to output the probability of each operation; the operation with the highest probability value is selected as the final transaction decision signal for the commodity.

9. The futures intelligent decision-making agent method based on NLP and multimodal data fusion according to claim 8, characterized in that, The futures decision signal generation module (400) further includes a decision interpretation unit (420), whose function is: During the reasoning process of the causal neural network reasoning module (300), all calculated attention weights α are recorded and stored; When the decision signal generation unit (410) outputs a "buy" or "sell" decision signal, it traces back from the commodity node corresponding to the decision signal in the dynamic heterogeneous causal graph. Using the recorded attention weights, find the top N causal transmission paths that contribute the most to the final decision. The causal transmission path is a sequence of nodes and edges that starts from a source event node and goes through multiple transmission steps to finally reach the target product node. The identified high-weight causal transmission paths are transformed into a structured natural language description to clearly explain the source events and industry transmission logic that lead to the decision-making recommendations.

10. A futures intelligent decision-making agent system, characterized in that, include: A preprocessing module (100) is configured to acquire and process multimodal data from multiple heterogeneous data sources in parallel. The multimodal data includes at least: structured futures time series data, unstructured text data, and static knowledge graph data describing the relationships between entities in commodity futures and their related industrial chains. A causal graph construction module (200) is connected to the preprocessing module (100) and is configured to receive the multimodal data and map the separated data in a unified manner to construct a dynamically evolving, heterogeneous causal graph containing nodes and edges. The nodes are used to represent commodity entities, driving event entities and concept entities in the market, and the edges are used to represent the causal relationships between the entities. A causal neural network inference module (300) is connected to the causal graph construction module (200) and configured to perform iterative message passing and node state updates on the dynamic heterogeneous causal graph to simulate the transmission process of information impact in the graph and calculate the final hidden state vector of each commodity entity node. A futures decision signal generation module (400) is connected to the causal neural network inference module (300) and is configured to generate a trading decision signal for the commodity futures based on the final hidden state vector of each commodity entity node.

Images