Bulk commodity price prediction system and method based on multi-model fusion

Abstract

The invention relates to the field of artificial intelligence, and discloses a bulk commodity price prediction system and method based on multi-model fusion, and the system comprises a data processing module, a hierarchical training module, a dynamic fusion module, and a prediction execution and verification module. The hierarchical training module carries out classification training on the base model according to the predicted time scale; the dynamic fusion module dynamically calculates the weight according to the historical prediction accuracy and stability index of each base model, and performs weighted combination on the prediction output of each model to generate a fusion prediction result; and the prediction execution and verification module establishes a closed-loop feedback mechanism, and triggers retraining when the performance is lower than a preset threshold value. According to the scheme, targeted optimization of different prediction periods is realized, market changes can be dynamically adapted, and the accuracy, robustness and automation level of bulk commodity price prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a commodity price forecasting system and method based on multi-model fusion. Background Technology

[0002] Commodities are the fundamental raw materials constituting the global economy, and their prices are characterized by dramatic fluctuations. This volatility is influenced by a complex interplay of factors, including global supply and demand, geopolitics, the macro-financial environment, and supply chain conditions, making accurate market predictions difficult. For companies involved in commodity trading, price volatility is the most critical market risk in their operations. Therefore, building a scientific and intelligent price forecasting system is of great significance for stabilizing operations and managing risks.

[0003] Currently, commodity price forecasting mainly relies on manual analysis or existing software systems, but both methods have significant limitations.

[0004] On the one hand, traditional manual forecasting methods rely heavily on analysts' personal experience and subjective judgment. This model lacks a systematic methodology, making the forecasting process difficult to standardize and replicate. Developing forecasting capabilities requires a long period of experience accumulation, and forecast quality can be inconsistent due to personnel changes, failing to provide stable and reliable decision support for enterprises.

[0005] On the other hand, to address the shortcomings of manual forecasting, the industry has developed various automated forecasting software programs. However, existing software systems still need improvement in terms of forecasting accuracy and adaptability. Many systems employ a single forecasting model, but any single model struggles to capture all market influencing factors and lacks generalization ability in the face of complex market changes, resulting in limited forecasting accuracy. While some systems attempt to integrate multiple models, they often use static weighting strategies, failing to dynamically adjust the contribution of each model based on changes in market conditions and the recent performance of each model. Consequently, the system's forecasting performance significantly declines when market conditions change.

[0006] Furthermore, existing technologies rarely employ differentiated modeling strategies for forecasting tasks across different time dimensions, such as short-term, medium-term, and long-term. This makes it difficult for the system to simultaneously ensure the accuracy of short-term price fluctuations and the precision of long-term price trends. Summary of the Invention

[0007] The purpose of this invention is to provide a commodity price forecasting system and method based on multi-model fusion, which solves the technical problems of low forecasting accuracy and poor market adaptability caused by insufficient generalization ability of single models, insufficient utilization of multi-source heterogeneous data, and static model weight allocation in existing forecasting technologies.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The first aspect of the present invention provides a commodity price forecasting system based on multi-model fusion, comprising: a data processing module, a hierarchical training module, a dynamic fusion module, and a forecast execution and verification module.

[0010] The data processing module is used to acquire multi-source heterogeneous data related to bulk commodities from multiple preset data sources and perform feature engineering operations to generate standardized feature sets. In this scheme, the feature engineering operations include constructing time-periodic features from the raw data. To overcome the discontinuity problem caused by linear time variables at periodic critical points, such as abrupt changes at the end of the year and the beginning of the year, the data processing module adopts a periodic feature transformation formula, which maps the time variable to sine and cosine components, converting it into continuous periodic features, thereby enabling the model to accurately learn the periodic patterns of the data.

[0011] The hierarchical training module, connected to the data processing module, is used to receive the standardized feature set, independently train and validate the base models in the base model library, and generate trained base models and corresponding historical prediction performance data.

[0012] In this scheme, the base model library is not a simple collection of models, but rather logically hierarchical based on the time scale of the prediction task. For example, statistical models are assigned to short-term prediction tasks, recurrent neural networks such as Long Short-Term Memory networks are assigned to medium- and long-term prediction tasks, and models based on the Transformer architecture are assigned to long-term prediction tasks. This hierarchical structure ensures that the system calls upon the most advantageous model type for different prediction periods, forming a specialized model combination.

[0013] The dynamic fusion module, connected to the hierarchical training module, receives the trained base model and its corresponding historical prediction performance data, calculates dynamic weights, and ultimately generates a fused prediction result. This module is the core of the dynamic adaptability of this scheme, and its calculation of dynamic weights involves multiple steps.

[0014] First, the historical prediction accuracy and historical prediction stability of each base model are quantified by using the mean absolute percentage error formula and the standard deviation formula of the prediction error.

[0015] Next, the comprehensive performance score formula is used to weight and combine the two different dimensions of accuracy and stability to generate a comprehensive performance score for each base model.

[0016] Finally, a dynamic weight normalization formula is used to normalize the overall performance score of each model, generating a set of dynamic weights. This weight allocation mechanism enables the system to dynamically adjust the contribution of each model to the final prediction result in real time based on its recent actual performance. Models with better performance receive higher weights, while those with poor performance receive lower weights, thus achieving a rapid response to market changes.

[0017] The prediction execution and verification module is connected to the dynamic fusion module and is used to receive the fusion prediction results and execute the prediction task.

[0018] The output of this scheme includes not only a single predicted price value, but also the predicted direction and the range of prediction uncertainty. The range of prediction uncertainty is calculated based on the degree of divergence among the independent prediction sets of all base models; a large divergence indicates significant differences in conclusions between models and high uncertainty.

[0019] In addition, the module has established a closed-loop feedback mechanism. When the long-term prediction performance of the system is detected to be continuously lower than the preset threshold, the hierarchical training module and the dynamic fusion module will be automatically triggered to perform the retraining process, thus ensuring the long-term effectiveness of the system.

[0020] A second aspect of this invention provides a method for predicting commodity prices based on multi-model fusion, comprising the following steps:

[0021] S1: Perform data processing and feature engineering to acquire and process multi-source heterogeneous data and generate standardized feature sets;

[0022] S2: Based on the standardized feature set, perform hierarchical and parallel model training, independently train the base models in the base model library, and generate trained base models and corresponding historical prediction performance data.

[0023] S3: Based on the trained base model and the corresponding historical prediction performance data, perform dynamic fusion weight calculation and prediction generation, calculate dynamic weights for each base model, and weight and combine the independent prediction outputs of each base model with the corresponding dynamic weights to generate the final fusion prediction result.

[0024] S4: Based on the fusion prediction results, perform multi-dimensional prediction task output and closed-loop verification, output structured prediction results, and periodically compare historical prediction results with actual market prices. When the performance is lower than a preset threshold, trigger the retraining of the base model library.

[0025] In summary, the present invention has at least one of the following beneficial technical effects:

[0026] 1. This invention improves the accuracy and robustness of the prediction system by establishing a dynamic fusion mechanism based on accuracy and stability. The system utilizes feature engineering processing of multi-source heterogeneous data, and in particular, uses a periodic feature transformation formula to solve the problem of discontinuity in time series. It also dynamically calculates normalized weights based on the historical prediction accuracy and stability indices of each base model. This mechanism enables the final fusion prediction result to automatically adapt to changes in the market environment, effectively overcomes the deficiency of insufficient generalization ability of existing single models, and solves the problem of limited accuracy in traditional prediction software.

[0027] 2. This invention adopts a hierarchical modeling strategy to achieve targeted optimization for prediction tasks in different time dimensions. By logically dividing the base model library into short-term, medium-to-long-term, and long-sequence prediction groups, the system can respectively use statistical models to capture short-term high-frequency fluctuations, use recurrent neural networks to fit nonlinear relationships, and use the Transformer architecture to grasp long-term trends. This method of matching the optimal model combination for different time scales takes into account both the accuracy of short-term predictions and the trend of long-term predictions, ensuring that the system can maintain excellent prediction performance in all time dimensions.

[0028] 3. This invention constructs a systematic prediction process that includes closed-loop verification and automatic optimization, which solves the problems of lack of standardization and reproducibility in manual prediction. The prediction execution and verification module continuously monitors the long-term performance indicators of the system and automatically triggers the retraining process of the base model when the performance is lower than a preset threshold. Combined with the quantitative output of the prediction uncertainty interval, it realizes real-time monitoring and consistency guarantee of prediction quality, thereby ensuring that the system can output high-quality prediction results in a long-term and stable manner. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the structure of the commodity price forecasting system based on multi-model fusion of the present invention;

[0030] Figure 2 This is a schematic diagram illustrating the dynamic weight changes of each base model in this invention;

[0031] Figure 3 This is a bar chart comparing the accuracy of the present invention.

[0032] Figure 4 This is a comparison chart of the prediction error distribution of the present invention;

[0033] Figure 5 This is a flowchart illustrating the commodity price forecasting method based on multi-model fusion of the present invention.

[0034] The module consists of: 10. Data processing module; 20. Layered training module; 30. Dynamic fusion module; and 40. Prediction execution and verification module. Detailed Implementation

[0035] Example 1:

[0036] See attached document Figure 1 The present invention provides a commodity price prediction system based on multi-model fusion, which may include: a data processing module 10, a hierarchical training module 20, a dynamic fusion module 30, and a prediction execution and verification module 40.

[0037] The prediction system can be deployed on one or more servers. The server includes a processor, memory, and a communication bus for transferring data between the processor and memory. The memory stores a computer program that, when executed by the processor, implements the prediction method steps described later.

[0038] The data processing module 10 is used to acquire multi-source heterogeneous data related to commodities from multiple preset data sources. The data processing module 10 performs feature engineering operations, including constructing time-series features, technical indicator features, and fundamental features from the raw data. After construction, the module 10 performs standardization processing on all features and outputs the processed feature set to the hierarchical training module 20.

[0039] The hierarchical training module 20, connected to the data processing module 10, receives the standardized feature set. Internally, the hierarchical training module 20 contains a base model library with multiple prediction models suitable for different prediction periods. Following a preset training strategy, the hierarchical training module 20 independently trains and validates each model in the base model library using the feature set, generating a series of trained base models.

[0040] The dynamic fusion module 30, connected to the hierarchical training module 20, receives the trained base models and their historical prediction performance data. Based on the historical prediction accuracy and stability metrics of each base model, the dynamic fusion module 30 calculates a dynamic weight for each base model using a preset weight calculation algorithm. Subsequently, the dynamic fusion module 30 weights and combines the prediction outputs of each base model with the corresponding dynamic weights to generate a final fused prediction result.

[0041] The prediction execution and verification module 40, connected to the dynamic fusion module 30, receives the fused prediction results. The prediction execution and verification module 40 is responsible for executing specific prediction tasks, such as outputting price predictions for a specific future point in time. Simultaneously, the prediction execution and verification module 40 also has a verification function, capable of comparing the prediction results with subsequently acquired actual market prices to evaluate and record the system's long-term prediction performance.

[0042] The specific implementation of the data processing module 10 is described below. The functions of the data processing module 10 are accomplished by the processor executing preset computer program instructions, which may include the following steps:

[0043] S11: Multi-source heterogeneous data acquisition. The data processing module 10 collects data from multiple data sources, including but not limited to: market data databases published by commodity exchanges, fundamental data databases provided by industry information websites, macroeconomic databases published by national statistical agencies, and supply chain databases provided by third-party data service providers. The acquired data types cover price data (such as opening price, closing price, highest price, and lowest price), trading volume data, inventory data, production capacity data, operating rate data, import and export data, macroeconomic indicators (such as the Consumer Price Index (CPI) and Purchasing Managers' Index (PMI), and price data of related commodities (such as crude oil).

[0044] S12: Feature Construction. The data processing module 10 processes the acquired raw data to construct a feature set for model training. This process includes:

[0045] First, construct time-period features. For periodic patterns present in time-series data, such as weekly, monthly, or annual patterns, a periodic feature transformation formula is used. The periodic feature transformation formula is:

[0046] ;

[0047] ;

[0048] in, For time distance variables, their units are related to the total length of the period. The units must be consistent; for example, when analyzing weekly patterns, It can be any day of the week. It is 7; Pi; It is a sine function; It is a cosine function; This indicates that the linear time variable The radian value mapped onto the unit circle; and These are the calculated sine and cosine eigenvalues, respectively. The combination of these two values ​​can uniquely represent the position of a time point within the period, thus solving the problem of discontinuity at the critical point of the period.

[0049] Secondly, technical indicator features are constructed by calculating a series of technical analysis indicators based on price and trading volume data. For the calculation of commonly used technical indicators such as moving averages (MA), relative strength index (RSI), and moving average converging / distracting line (MACD), those skilled in the art can use existing standard algorithms, and the calculation methods are well-known in the field, so they will not be elaborated here.

[0050] Third, construct fundamental features. Perform month-on-month and year-on-year calculations on absolute numerical data such as inventory and production capacity, or perform ratio calculations (such as inventory-to-consumption ratio) to generate derived features that reflect changes in fundamentals.

[0051] S13: Feature Standardization. To eliminate the impact of differences in the units of measurement between different features on model training, data processing module 10 uses the min-max normalization formula to normalize each feature in the feature set. The min-max normalization formula is:

[0052] ;

[0053] in, The original value of a certain feature; This is the minimum value of the feature in the training dataset; This represents the maximum value of the feature in the training dataset. This is the difference between the original value and the minimum value. This represents the range of the feature in the training dataset. These are the standardized eigenvalues ​​calculated and whose values ​​are in the range [0,1].

[0054] In a specific implementation, if the following occurs equal In the case where the eigenvalues ​​are constant, then all corresponding... The value is assigned to a preset constant (e.g., 0.5) to ensure the stability of the calculation.

[0055] S14: Feature Importance Analysis. After constructing a massive number of features, the data processing module 10 can perform feature importance analysis to quantify the contribution of various features to price prediction and verify the effectiveness of feature engineering. Referring to Appendix Table 1, this table represents the distribution of feature importance weights in a specific commodity (such as polyolefin) price prediction scenario according to an embodiment of the present invention.

[0056] Table 1. Feature Importance Weight Distribution Table

[0057] Feature categories Weight Main features Contribution stability Price technical indicators 28.5% Moving average, RSI, MACD high 0.89 Supply and demand fundamentals 24.3% Inventory, operating rate, and production capacity high 0.92 Macroeconomics 18.7% GDP, CPI, PMI middle 0.76 crude oil prices 15.2% Brent, WTI high 0.88 Time period 8.9% Seasonal, Holidays middle 0.94 Other factors 4.4% exchange rates and policies Low 0.67

[0058] in conclusion:

[0059] As shown in Table 1, the combined weight of price technical indicators and supply and demand fundamentals exceeds 50%, indicating that market trading behavior and the underlying supply and demand relationship of commodities are the core driving factors influencing price trends. This invention, through multi-source data collection and feature engineering, can effectively capture and quantify these core factors. Furthermore, crude oil prices, as an important upstream cost factor, also have a weight of 15.2%, verifying the effectiveness of incorporating cross-commodity correlation features in this scheme. This importance analysis result intuitively reveals the core driving factors affecting prices, providing a basis for understanding model decisions and demonstrating the comprehensiveness of the feature engineering scheme of this invention. In an optional implementation, this analysis result can also be used to guide subsequent feature optimization iterations, such as removing redundant features with consistently zero or extremely low contribution.

[0060] Next, the specific implementation of the hierarchical training module 20 will be described. The hierarchical training module 20 receives the standardized feature set output by the data processing module 10. Its function is also accomplished by the processor executing preset computer program instructions, specifically including the following steps:

[0061] S21: Construct a classification / grouping base model library. The hierarchical training module 20 internally configures a base model library containing multiple prediction models. Here, "hierarchy" (or "logical grouping") refers to the logical classification and division of models in the library based on the time scale (horizon) of the prediction task and the characteristics of the models themselves. Its purpose is to ensure the diversity and relevance of the model library. Based on the algorithmic characteristics of each model and its performance at different prediction periods, this invention has formulated a targeted model combination strategy. Specific divisions include:

[0062] Short-term forecast group (e.g., 1–7 days): This group contains models suitable for capturing short-term high-frequency fluctuations and nonlinear patterns. In a specific embodiment, this group may preferentially select NHITS models, Long Short-Term Memory (LSTM) models, and TSMier models. These models perform well in handling shorter sequences and accurately capturing local price movements.

[0063] Medium- to long-term forecasting group (e.g., 8–30 days): This group includes models that offer advantages in balancing forecasting accuracy and long-term trend capture. In a specific implementation, this group may preferentially select Informer, iTransformer, and NHITS models. These models are effective at handling medium-length time series and have a good fit to trend changes.

[0064] Long sequence prediction group (e.g., 30 days or more): This group contains advanced models designed to improve the performance of long sequence prediction, better capturing long-term trends driven by macroeconomic factors. In one specific embodiment, this group may preferentially select iTransformer models, Informer models, and Long Short-Term Memory (LSTM) network models. These models, through their unique attention mechanisms or recurrent structures, effectively capture long-term dependencies between sequences.

[0065] Through this hierarchical and targeted model combination strategy, the present invention ensures that the system calls the most advantageous model type in different prediction periods, forming a professional model matrix, and providing high-quality benchmark predictions for subsequent dynamic fusion.

[0066] S22: Execute a parallel and independent training strategy. The hierarchical training module 20 employs a parallel and independent training strategy. It provides the complete training dataset to each model in the base model library and trains them independently. This training method allows each model to extract and fit specific patterns in the data according to its algorithmic structure. For example, the LSTM model excels at capturing sequence dependencies, while the NHITS model is better at decomposing and predicting the trend and seasonal components of sequences.

[0067] S23: Model Training and Performance Recording. During training, the hierarchical training module 20 divides the feature set into a training set and a validation set. Each base model undergoes parameter optimization on the training set and periodic performance evaluation on the validation set. Evaluation metrics include, but are not limited to, mean absolute percentage error (MAPE) and root mean square error (RMSE). After each base model completes training, its final model parameters and historical performance evaluation data on the validation set are saved and output to the subsequent dynamic fusion module 30 as the basis for dynamic weight calculation.

[0068] The specific implementation of the dynamic fusion module 30 is described below. This module receives the base models and their historical prediction performance data output by the hierarchical training module 20. Its function is accomplished by the processor executing preset computer program instructions, which may include the following steps:

[0069] S31: Quantitative Evaluation of Historical Model Performance. The dynamic fusion module 30 first quantitatively evaluates the historical prediction performance of each base model, with evaluation dimensions including accuracy and stability. For accuracy, the Mean Absolute Percentage Error (MAPE) formula is used for calculation. The MAPE formula is as follows:

[0070] MAPE ;

[0071] in, The number of sample points during the historical evaluation period; For at a certain point in time The actual price of the goods; For the model at time points Predicted price; This represents the prediction error at a single point. The prediction error rate for a single point; This represents the absolute percentage error at a single point.

[0072] The smaller the value calculated by this formula, the higher the prediction accuracy of the model. In a specific implementation, to prevent the actual value from being... A value of zero or close to zero can cause instability in calculations; a very small positive number can be added to the denominator. , or in When the value is zero, the error term at that point is not included in the summation.

[0073] For stability, the standard deviation formula of the prediction error is used for calculation. First, the prediction error sequence is calculated. :

[0074] ;

[0075] Subsequently, the dispersion of the error sequence is calculated using the standard deviation formula for the prediction error:

[0076] ;

[0077] in,

[0078] in, For at a certain point in time The prediction error; This is the arithmetic mean of all prediction errors; This represents the deviation of a single error value from the mean error. This represents the standard deviation of the prediction error. The smaller the value, the smaller the fluctuation of the model's prediction error and the more stable the prediction performance.

[0079] S32: Calculate dynamic weights. The dynamic fusion module 30 calculates a dynamic weight for each base model based on the quantitative evaluation results. This process first uses the model comprehensive performance score formula to combine the accuracy and stability indicators into a single score. The model comprehensive performance score formula is:

[0080] ;

[0081] in, For the first The overall performance score of each base model; and The first The mean absolute percentage error and standard deviation of the prediction error for each model; It is a very small positive number (e.g.) This is used to prevent the denominator from being zero; and These are the preset weighting coefficients for accuracy and stability indicators, respectively. In this embodiment, it can be set , This formula converts performance metrics that are better when the value is smaller into scores that are better when the value is larger, and then weights them according to preset preferences.

[0082] This formula converts performance metrics that prioritize smaller values ​​into scores that prioritize larger values, and then weights these scores according to preset preferences. In practical applications of financial and commodity forecasting, an ideal model not only needs a small average error (high accuracy) but also stable performance, avoiding extreme prediction errors (high stability), because sporadic, large errors can lead to catastrophic decision-making failures. The comprehensive performance scoring formula of this invention is designed based on this business insight. By weighting and combining accuracy and stability, it can more comprehensively evaluate the "reliability" of the model in real-world business scenarios.

[0083] Then, a dynamic weight normalization (Softma) formula is used to convert the combined performance scores of all base models into a set of weights that sum to 1. The dynamic weight normalization formula is:

[0084] ;

[0085] in, Is assigned to the first The final dynamic weights of each base model; It is the first The overall performance score of each base model; It represents the total number of base models participating in the fusion; It is the base of the natural logarithm; Indicates the first The performance scores of each base model are exponentially amplified to highlight the contributions of high-performing models. This formula assigns greater weight to models with higher performance scores.

[0086] See attached document Figure 2 The figure shows the fusion weights of multiple base models (such as NHITS, Informer, etc.) over a period of time. The trajectory changes over time. As can be seen from the graph, no single model can consistently maintain the highest weight; the weights of each model are dynamically adjusted according to changes in the market environment. This intuitively demonstrates the principle of the dynamic fusion mechanism of this invention: instead of pre-setting any particular model as optimal, it adaptively adjusts the contribution ratio of each model to the final prediction result based on its actual performance on recent historical data, thereby improving the robustness and environmental adaptability of the entire prediction system.

[0087] S33: Generate fused prediction results. The dynamic fusion module 30 combines the independent predictions from each base model for the same future time point with the dynamic weights calculated in S32. The final fused prediction result is generated using the weighted fusion prediction formula. The weighted fusion prediction formula is:

[0088] ;

[0089] in, It is the final output of the fused predicted price; and They are the first Dynamic weights and independent price predictions of each base model; Indicates the first The contribution of each base model to the overall prediction result.

[0090] Please refer to Appendix 2 and Appendix 3. Figure 3 .

[0091] Table 2. Comparison of Prediction Accuracy Across Eight Dimensions

[0092] Predictive Dimensions Single model best Multi-model fusion Improved accuracy Directional accuracy Confidence interval Next day's forecast 85.2% (LSTM) 91.8% +6.6% 78.2% ±2.1% This week's forecast 78.5% (GBoost) 84.7% +6.2% 74.8% ±3.1% Week 2 Forecast 75.3% (NHITS) 81.2% +5.9% 72.1% ±4.0% Next week's forecast 76.8% (Informer) 82.9% +6.1% 73.5% ±3.3% Next month's forecast 72.1% (iTransformer) 78.6% +6.5% 70.8% ±5.2% February forecast 68.9% (LSTM) 75.4% +6.5% 68.9% ±6.1% March forecast 65.7% (NHITS) 72.3% +6.6% 67.2% ±7.0% April forecast 62.4% (Informer) 69.1% +6.7% 65.7% ±7.9%

[0093] in conclusion:

[0094] From Table 2 and Figure 3 The data shows that the multi-model fusion method of this invention achieves higher accuracy across all prediction dimensions than the best-performing single model in the corresponding dimension. The average accuracy improvement exceeds 6%. This indicates that the performance of a single model has limitations across different prediction periods, while the dynamic fusion mechanism of this invention, by combining the outputs of multiple models, can effectively compensate for the performance shortcomings of a single model, thereby achieving stable performance gains across various prediction periods.

[0095] Refer to Appendix 3 and Appendix Figure 4 .

[0096] Table 3. Statistical Table of Model Performance Indicators

[0097] Performance indicators Single model Multi-model fusion degree of improvement Statistical significance Mean Absolute Error (MAPE) 7.8% 4.9% -37.2% p<0.001 Root Mean Square Error (RMSE) 0.156 0.098 -37.2% p<0.001 Coefficient of determination (R²) 0.742 0.891 +20.1% p<0.001 Directional accuracy 68.5% 76.8% +8.3% p<0.001 Maximum prediction bias 18.7% 12.3% -34.2% p<0.001 Predicted stability 0.68 0.85 +25.0% p<0.001

[0098] in conclusion:

[0099] Table 3 further confirms the effectiveness of the technical solution of this invention from multiple perspectives. Compared with a single model, the mean absolute percentage error (MAPE) and root mean square error (RMSE) of the multi-model fusion method are both reduced by approximately 37%, indicating that the average bias of the fusion prediction results is smaller, and the magnitude and frequency of extreme errors are also controlled. The coefficient of determination (R²) is improved by more than 20%, indicating that the fusion model has a stronger explanatory power for price fluctuations. In addition, the improvements in indicators such as directional accuracy, maximum prediction bias, and prediction stability, and the fact that all improvements are statistically significant (p<0.001), jointly demonstrate that the dynamic fusion method of this invention can generate more accurate and robust prediction outputs than any single base model, which has significant value in commercial applications requiring high-reliability decision support.

[0100] Finally, the specific implementation of the prediction execution and verification module 40 is described. This module is connected to the dynamic fusion module 30, and its function is accomplished by the processor executing preset computer program instructions, which may include the following steps:

[0101] S41: Perform multi-dimensional prediction task. Prediction execution and verification module 40 receives the fused prediction results generated by dynamic fusion module 30. The prediction execution and verification module 40 executes prediction tasks for multiple different time dimensions based on preset business requirements.

[0102] In one specific embodiment, the time dimension may include eight dimensions: the next day, this week, the second week, next week, next month, next February, next March, and next April. For the prediction task of each dimension, the prediction execution and verification module 40 outputs a structured prediction result, which may include the predicted price value, the predicted direction (up, down, or sideways), and a prediction uncertainty range.

[0103] The uncertainty interval is used to quantify the reliability of this prediction, and it is constructed based on the independent predictions of all base models. The degree of divergence. For example, this uncertainty interval can be calculated as:

[0104] ;

[0105] in, It is a set of independent predictions from the base model. The standard deviation of the σ². This value measures the consistency of predictions among different models and is often referred to as "ensemble spread". The larger the value, the greater the discrepancy in predictions between the models, which means a more uncertain market outlook and a higher degree of uncertainty in the results of this fusion prediction. This is a preset interval coefficient (e.g., set to 1 or 2), used to control the width of the interval. It is important to emphasize that this prediction uncertainty interval is not a confidence interval in the traditional statistical sense, but rather a measure that intuitively reflects the consensus of the model's group of opinions.

[0106] The width of the interval provides decision-makers with an important reference for the reliability of the forecast; a narrow interval means that multiple models agree and the forecast results are more reliable; while a wide interval suggests that there is a large disagreement in the forecast and decision-makers should take a more cautious approach.

[0107] The prediction results can be provided to downstream business systems through an application programming interface (API), written to a database for querying, or visualized on a user interface.

[0108] S42: Continuous verification of forecast results. During system operation, the forecast execution and verification module 40 periodically retrieves actual market price data that has already occurred from the data source. The forecast execution and verification module 40 compares the stored forecast results for a certain point in the past with the corresponding actual market price.

[0109] By continuously calculating and recording performance metrics such as MAPE, RMSE, and orientation accuracy, a long-term prediction performance tracking record of the system is formed. This verification mechanism is used to monitor whether the model performance degrades due to changes in the underlying market logic, i.e., model drift. When one or more performance metrics are continuously lower than a preset threshold, the system can generate an alarm, prompting that the base model library needs to be updated or retrained. Alternatively, in another embodiment, it can be configured to directly trigger the hierarchical training module 20 and the dynamic fusion module 30 to perform a retraining process.

[0110] S43: Evaluate the overall engineering performance of the system. In addition to accuracy verification at the algorithm level, this invention also focuses on the overall performance of the system in engineering applications.

[0111] Referring to Appendix Table 4, this table contains statistical data on key performance indicators of the prediction system of this invention within a deployment cycle.

[0112] Table 4. Comparison of System Performance Indicators

[0113] Performance indicators Target value actual value Achievement rate industry average Exceeding range System Availability 99.5% 99.7% 100.2% 98.2% +1.5% Response time <3 seconds 1.2 seconds >100% 4.8 seconds -75.0% Concurrent processing >1000 / second 1247 / second 124.7% 856 / second +45.7% Data delay <5 minutes 2.3 minutes >100% 8.7 minutes -73.6% Prediction accuracy >80% 91.8% 114.8% 76.3% +20.3% User satisfaction >8.0 points 8.6 points 107.5% 7.1 points +21.1%

[0114] in conclusion:

[0115] The data in Table 4 shows that the system implementation of this invention has achieved a high standard at the engineering level. The analysis is as follows: System availability reaches 99.7%, exceeding the target value and the industry average. This is attributed to the clear modular design of this invention (data processing module 10, hierarchical training module 20, dynamic fusion module 30, prediction execution and verification module 40), which allows each unit to be maintained independently and is fault-tolerant, improving overall stability. The response time is 1.2 seconds, significantly better than the target value of less than 3 seconds, which is attributed to the prediction execution process of this invention. Since model training is completed offline in advance, online prediction only involves feature extraction and a series of efficient matrix operations. In particular, the dynamic fusion step is only a single weighted summation, with low computational overhead, thus enabling rapid response. The concurrent processing capacity exceeds 1200 times / second, indicating that the system architecture has good scalability, can handle high-frequency prediction requests, and meets the needs of large-scale commercial applications. The data latency is 2.3 minutes, also significantly better than the target value of less than 5 minutes, reflecting the efficiency of the data processing module 10, which can quickly complete the entire process from multi-source data acquisition to feature engineering, ensuring the high timeliness of the data used for prediction. In summary, the high availability, low latency, and high concurrency at the system level, combined with the high prediction accuracy at the algorithm level, demonstrate that the technical solution of this invention not only ensures algorithm performance but also achieves high efficiency and high reliability in its system implementation, thus meeting the requirements of actual commercial deployment.

[0116] See attached document Figure 5 The present invention also provides a method for predicting commodity prices based on multi-model fusion, which can be executed by the aforementioned system, and its specific process may include the following steps:

[0117] S100: Perform data processing and feature engineering. Acquire multi-source heterogeneous data, including market data, fundamental data, and macroeconomic data. Perform feature engineering operations on the data, including constructing time-cycle features, technical indicator features, and fundamental features. Finally, standardize the constructed feature set to generate a standardized feature set for subsequent model use.

[0118] S200: Perform hierarchical and parallel model training. Based on a standardized feature set, a parallel and independent training strategy is used to train a pre-defined base model library containing multiple different types of base models (e.g., LSTM, NHITS, Informer, etc.). This training process generates model parameters for each base model and records its historical prediction performance data on the validation set.

[0119] S300: Performs dynamic fusion weight calculation and prediction generation. It receives trained base models and their historical prediction performance data. Based on the prediction accuracy and stability metrics of each base model on recent historical data, it calculates a dynamic weight for each base model using a preset performance score formula and weight normalization formula. Subsequently, it weights and combines the independent prediction outputs of each base model for the same future time point with the corresponding dynamic weights to generate a final fusion prediction result.

[0120] S400: Performs multi-dimensional prediction tasks and outputs results. Based on the fused prediction results, it performs prediction tasks for one or more preset time dimensions (e.g., next day, next week, next month, etc.). For each time dimension, it outputs structured prediction results including predicted price value, prediction direction, and confidence interval.

[0121] S500: Perform closed-loop validation and model optimization. This step uses a structured quality control process to verify the reasonableness of the prediction results and continuously monitor and optimize model performance to ensure the commercial usability of the output. This process may specifically include:

[0122] S501: Perform a time-series logical consistency check. Check for logical gaps in the forecast results across adjacent time dimensions. For example, verify whether the forecast value on the start day of "This Week's Forecast" is continuous with the result of "Next Day's Forecast," and that the difference should be within a preset reasonable fluctuation range (e.g., 5%); verify whether the forecast value on the end day of "This Week's Forecast" is smoothly connected with the forecast value on the start day of "Second Week's Forecast." This step ensures the continuity and logical rationality of the multi-dimensional forecast results over time.

[0123] S502: Perform a historical data reasonableness check. Compare the forecast results for each dimension with the distribution range of historical price data. Calculate the statistical interval of historical prices (e.g., based on the highest price, lowest price, mean, and standard deviation of the past year) and check whether the forecast value falls within this statistically reasonable range. For forecast values ​​that deviate significantly from the historical interval, the system marks them to indicate potential market structural changes or model forecast bias.

[0124] S503: Perform a trend consistency check. Check whether the trend direction of the short-term forecast is consistent with the trend direction of the medium- and long-term forecasts on a macro level, avoiding logical contradictions such as a short-term forecast of a sharp rise suddenly turning into a sudden, unexplained sharp fall. For example, by calculating the trend slope of the forecast sequences in different dimensions and analyzing their correlation, ensure that the underlying trend of the forecast results is consistent.

[0125] S504: Calculate the prediction quality score and handle anomalies. Based on the inspection results of S501 to S503, a quantified quality score is generated for the prediction result in each dimension. When the quality score of a prediction result is lower than a preset threshold, or is identified as an anomaly in any of the above inspections, the system executes an anomaly handling mechanism. This mechanism may include: marking the prediction result with a warning to prompt the user to use it with caution; or automatically triggering a retraining process for the relevant base model to correct model bias.

[0126] S505: Performs long-term performance monitoring and retraining. It periodically compares historical prediction results with subsequently acquired actual market prices, continuously calculating and monitoring long-term performance metrics such as MAPE, RMSE, and directional accuracy. When any key metric is detected to be continuously below a preset threshold, it triggers retraining of the entire base model library (corresponding to S200) and updating of the dynamic weight benchmark (corresponding to S300), achieving closed-loop self-optimization of the system.

[0127] By performing steps S100 to S505, the method of the present invention forms a complete workflow from data processing to model training, dynamic fusion, multidimensional prediction and structured quality control, thereby achieving continuous, high-precision and commercially reliable prediction of commodity prices.

[0128] Example 2:

[0129] This embodiment is used to specifically illustrate the application of the present invention in the scenario of predicting the price of polyolefins (such as linear low-density polyethylene LLDPE or polypropylene PP). This application follows the system architecture and method flow of Embodiment 1.

[0130] In the data processing and feature engineering step (corresponding to S100), the specific multi-source heterogeneous data collected by the data processing module 10 for the characteristics of polyolefins include:

[0131] Upstream cost data: international crude oil (Brent, WTI) prices, naphtha prices.

[0132] Supply-side data: operating rates of domestic polyolefin production facilities, planned and unplanned maintenance information, progress of new capacity commissioning, and import and export data of polyolefins.

[0133] Demand-side data: Operating rates and seasonal demand changes in major downstream application areas (such as agricultural film, packaging film, injection molding, pipes, etc.) (e.g., peak and off-peak seasons for agricultural film).

[0134] Inventory data: Production enterprise inventory, port inventory, and social inventory.

[0135] Market trading data: Prices (opening price, closing price, highest price, lowest price), trading volume, and open interest of futures contracts corresponding to the Dalian Commodity Exchange, as well as spot market quotation data.

[0136] In constructing the features, in addition to the general technical indicators and time cycle features of Example 1, derivative features reflecting the characteristics of the polyolefin industry chain were also specifically constructed. For example, crude oil prices were used as an important cost-driven feature, the operating rate and maintenance losses of petrochemical enterprises were quantified as features reflecting supply changes, and the ratio of social inventory to downstream operating rate was calculated as the inventory-to-consumption ratio feature.

[0137] In the model training and fusion steps (corresponding to S200 and S300), the hierarchical training module 20 uses the feature set constructed for polyolefins to train each model in the base model library. When generating fused predictions, the dynamic fusion module 30 adaptively reflects the characteristics of the polyolefin market with its dynamic weights. For example, during periods of concentrated new capacity release or large-scale maintenance, significant changes occur in the market supply, and the weights of models more sensitive to fundamental data dynamically increase; conversely, during periods of sharp fluctuations in crude oil prices, the weights of models more correlated with costs increase accordingly.

[0138] In the prediction execution and verification steps (corresponding to S400 and S500), the system outputs price predictions for LLDPE or PP across multiple future time dimensions. Application test results show that when applied to polyolefin price prediction, the performance indicators of this invention also reach the high levels shown in Appendix 2, Appendix 3, and Appendix 4 of Example 1. For example, the MAPE after multi-model fusion is significantly lower than any single base model, and the system engineering indicators are robust. This fully demonstrates the good applicability and high accuracy of this invention in this commodity field.

[0139] Example 3:

[0140] This embodiment is used to specifically illustrate the application of the present invention in the scenario of crude oil (e.g., WTI or Brent crude oil) price prediction.

[0141] In the data processing and feature engineering step (corresponding to S100), considering the characteristics of crude oil as a global commodity, its influencing factors are more macroscopic and complex. The specific multi-source heterogeneous data collected by data processing module 10 includes:

[0142] Supply-side data includes: production quotas and actual output of the Organization of the Petroleum Exporting Countries and its allies (OPEC+); weekly crude oil inventory and production data released by the U.S. Energy Information Administration (EIA) and the American Petroleum Institute (API); the number of active drilling rigs in the United States released by Baker Hughes; and geopolitical events in major oil-producing countries (such as Saudi Arabia, Russia, and Iran).

[0143] Demand-side data includes macroeconomic indicators (PMI, GDP, etc.) of major global economies (such as China, the United States, and the European Union), monthly / annual demand forecast reports released by the International Energy Agency (IEA) and OPEC, global flight numbers, and transportation indices of major countries, reflecting end-consumer data.

[0144] Financial and market data: US Dollar Index (DY), US Treasury yields, Commitment of Traders (COT) report released by the Commodity Futures Trading Commission (CFTC), and Market Fear Index (VI).

[0145] Logistics data: the status of key shipping lanes (such as the Strait of Hormuz and the Suez Canal) and global Very Large Crude Carrier (VLCC) freight indices.

[0146] In constructing the features, this invention uses the inverse change of the US dollar index as an important financial feature affecting the price of crude oil denominated in US dollars, and uses the net long positions of commercial and non-commercial entities in the CFTC report as a feature reflecting market sentiment and capital flows.

[0147] In the model training and fusion steps (corresponding to S200 and S300), the feature set constructed for the crude oil market is used for training. The weight allocation of the dynamic fusion module 30 can reflect the adaptability to different driving modes of the crude oil market. For example, before and after the OPEC+ meeting to decide on production policy, the market focus is on the supply side, and the weight of models sensitive to relevant features will increase at this time; while during the centralized release of global economic data, the weight of models reflecting macroeconomic demand will increase accordingly.

[0148] In the prediction execution and verification steps (corresponding to S400 and S500), the system outputs a future price prediction for WTI or Brent crude oil. Application results show that despite the numerous and volatile factors influencing the crude oil market, this invention, through comprehensive capture of multi-source heterogeneous data and adaptive adjustment of a dynamic fusion mechanism, can still achieve stable and reliable predictions. Its performance is also consistent with the performance indicators shown in the appendices of Example 1, further verifying the versatility and robustness of the method of this invention.

[0149] Example 4:

[0150] This embodiment illustrates the specific application of the present invention in the scenario of petroleum coke price prediction. As a byproduct of crude oil refining, the price of petroleum coke is influenced by both upstream refineries and downstream demand.

[0151] In the data processing and feature engineering step (corresponding to S100), considering the industrial characteristics of petroleum coke, the data processing module 10 collects and processes specific multi-source heterogeneous data, including:

[0152] Upstream supply and cost data: operating rates of atmospheric and vacuum distillation units at major domestic and overseas refineries, processing ratios of different types of crude oil (e.g., heavy crude oil, sulfur-containing crude oil), refinery maintenance plans, port inventory and import volume of petroleum coke.

[0153] Downstream demand data: Operating rates and output data of major downstream industries such as electrolytic aluminum, cement, and steel, especially the output and inventory data of prebaked anodes (a key consumable in electrolytic aluminum production).

[0154] Prices of related products and substitutes: Price data for alternative fuels such as thermal coal and anthracite.

[0155] Market transaction data: ex-factory prices and port spot prices of petroleum coke of different grades (such as low sulfur, medium sulfur, and high sulfur), as well as freight indices.

[0156] In constructing the features, in addition to general features, derivative features reflecting the market logic of petroleum coke were also constructed. For example, features reflecting downstream absorption capacity were constructed by calculating "electrolytic aluminum profit and prebaked anode price", and features reflecting the substitution effect of petroleum coke in the fuel sector were constructed by calculating "petroleum coke price and thermal coal price".

[0157] In the model training and fusion steps (corresponding to S200 and S300), the feature set constructed for petroleum coke as described above is used for training. The weight allocation of the dynamic fusion module 30 can adapt to changes in the dominant driving factors of the petroleum coke market. For example, when environmental policies lead to a decrease in demand for high-sulfur coke, the weights of models that are highly sensitive to downstream demand and environmental policy characteristics will be increased accordingly; when concentrated refinery maintenance leads to a contraction in supply, the weights of models that are sensitive to supply-side data will increase.

[0158] As can be seen from the above embodiments, the prediction system and method of the present invention possess good versatility and robustness, and can adapt to the market characteristics of different commodities. More importantly, the high-precision prediction results generated can be deeply integrated with real business decision-making processes, creating quantifiable value for market participants.

[0159] To further illustrate the practicality and commercial value of the present invention, please refer to Appendix Table 5, which compares the improvement effects of applying the prediction results of the present invention to different business scenarios (such as futures trading, corporate procurement, trade, etc.) on key business indicators.

[0160] Table 5. Statistical Table of Business Application Effectiveness

[0161] Application scenarios Indicator Name Before application After application Improvement range Sample size Futures trading Trading win rate 65.2% 78.9% +13.7% 1,356 people Futures trading Annualized rate of return 8.3% 12.7% +4.4% 1,356 people Futures trading Maximum drawdown 15.8% 9.2% -41.8% 1,356 people Corporate Procurement Procurement costs 100% of the benchmark 93.3% -6.7% 89 companies Corporate Procurement Inventory turnover rate 6.2 times / year 8.1 times / year +30.6% 89 companies Corporate Procurement Out-of-stock rate 3.8% 1.2% -68.4% 89 companies Traders Inventory turnover rate 4.8 times / year 6.9 times / year +43.8% 156 traders Traders Capital occupation 100% of the benchmark 91.7% -8.3% 156 traders Traders operational efficiency 100% of the benchmark 128.6% +28.6% 156 traders

[0162] in conclusion:

[0163] Table 5 shows the changes in key performance indicators under different business scenarios before and after applying the prediction results of this invention. For futures trading, the improvement in trading win rate and annualized return can be attributed to the high-precision price prediction and direction judgment capabilities provided by this invention, offering a reference for trading decisions. The significant decrease in maximum drawdown reflects the stability of the predictions provided by this invention; the dynamic fusion mechanism helps avoid extreme erroneous predictions, thereby helping to control trading risks. For enterprise procurement scenarios (such as electrolytic aluminum plants purchasing petroleum coke), the reduction in procurement costs is due to the ability of enterprises to choose to purchase at predicted low prices based on the price predictions provided by this invention. The improvement in inventory turnover and the reduction in stockout rate are because reliable predictions allow enterprises to implement more refined inventory management, reduce unnecessary safety stock, and avoid raw material shortages caused by misjudging the market. For trading scenarios, the improvement in inventory turnover, capital occupation, and operating efficiency is because the prediction results of this invention help traders more accurately grasp market rhythms and conduct rapid in-and-out operations, thereby accelerating capital turnover and improving profitability per unit time.

[0164] In summary, the results of this embodiment demonstrate that the present invention is not only technically applicable to the specific commodity of petroleum coke, but also that its output prediction results can be combined with actual business decisions, bringing quantifiable business indicator improvements to different types of market participants, thus reflecting the practicality of the method of the present invention.

Claims

1. A commodity price forecasting system based on multi-model fusion, characterized in that, include: The data processing module is used to acquire multi-source heterogeneous data related to bulk commodities from multiple preset data sources and perform feature engineering operations to generate standardized feature sets. A hierarchical training module, connected to the data processing module, is used to receive the standardized feature set and independently train and validate the base models in the base model library according to a preset training strategy, generating trained base models and their historical prediction performance data. The dynamic fusion module, which is connected to the hierarchical training module, is used to receive the trained base models and their corresponding historical prediction performance data, calculate dynamic weights based on the historical prediction accuracy and historical prediction stability indices of each base model, and weight and combine the prediction outputs of each base model with the corresponding dynamic weights to generate the final fusion prediction result. The prediction execution and verification module, which is connected to the dynamic fusion module, is used to receive the fusion prediction results, execute specific prediction tasks, compare the prediction results with the subsequently obtained actual market prices, and evaluate the long-term prediction performance of the system.

2. The commodity price forecasting system based on multi-model fusion according to claim 1, characterized in that, The feature engineering operations performed by the data processing module specifically include: Time period features, technical indicator features, and fundamental features are constructed from the multi-source heterogeneous data.

3. The commodity price forecasting system based on multi-model fusion according to claim 1, characterized in that, When constructing the time periodic features, the data processing module uses a periodic feature transformation formula to map linear time variables into sine and cosine feature values.

4. The commodity price forecasting system based on multi-model fusion according to claim 1, characterized in that, The hierarchical training module independently trains and validates the base models in the base model library. The base model library logically categorizes and divides the base models according to the time scale of the prediction task. This categorization and division includes: Short-term forecasting groups, including statistical models or lightweight machine learning models; Medium- to long-term prediction groups, which include long short-term memory networks or gated recurrent units; Long sequence prediction group, which includes models based on the Transformer architecture.

5. The commodity price forecasting system based on multi-model fusion according to claim 1, characterized in that, The dynamic fusion module determines the historical prediction accuracy and stability indicators in the following ways: The historical prediction accuracy of each base model is calculated using the mean absolute percentage error formula; The historical prediction stability of each base model is calculated using the standard deviation formula of the prediction error.

6. The commodity price forecasting system based on multi-model fusion according to claim 5, characterized in that, The process by which the dynamic fusion module calculates dynamic weights based on the historical prediction accuracy and stability indices of each base model further includes: The historical prediction accuracy and historical prediction stability are weighted and combined into a comprehensive performance score for each base model using a model comprehensive performance score formula.

7. The commodity price forecasting system based on multi-model fusion according to claim 6, characterized in that, When calculating the dynamic weights, the dynamic fusion module further employs a dynamic weight normalization formula to convert the comprehensive performance scores of all base models into a set of normalized dynamic weights.

8. The commodity price forecasting system based on multi-model fusion according to claim 1, characterized in that, When the prediction execution and verification module executes the specific prediction task, the prediction results output include: Predict the price value, predict the direction, and predict the range of uncertainty. The prediction uncertainty interval is calculated based on the degree of divergence of the set of independent predictions from all base models, and is used to quantify the reliability of the prediction results.

9. The commodity price forecasting system based on multi-model fusion according to claim 1, characterized in that, After evaluating the long-term predictive performance of the system, the prediction execution and verification module is further used for: When the long-term prediction performance is detected to be continuously lower than a preset threshold, the hierarchical training module and the dynamic fusion module are triggered to perform a retraining process.

10. A commodity price forecasting method based on multi-model fusion, characterized in that, Includes the following steps: S1. Perform data processing and feature engineering to acquire and process multi-source heterogeneous data and generate standardized feature sets; S2. Based on the standardized feature set, perform hierarchical and parallel model training, independently train the base models in the base model library, and generate the trained base models and their historical prediction performance data. S3. Based on the trained base models and their historical prediction performance data, perform dynamic fusion weight calculation and prediction generation, calculate dynamic weights for each base model, and weight and combine the independent prediction outputs of each base model with the corresponding dynamic weights to generate the final fusion prediction result. S4. Based on the fusion prediction results, perform multi-dimensional prediction task output and closed-loop verification, output structured prediction results, and periodically compare historical prediction results with actual market prices. When the performance is lower than a preset threshold, trigger the retraining of the base model library.

Images