Exchange rate and agricultural product price prediction system

The system addresses the challenge of inaccurate exchange rate and agricultural price prediction by employing advanced data processing and modeling techniques, resulting in reliable forecasts that support informed decision-making and economic stability.

KR103044705B1Active Publication Date: 2026-11-11ACORN CO
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
KR1020240023278
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Priority Date
2023-03-08
Filing Date
2024-02-19
Publication Date
2026-11-11
Estimated Expiration
2044-02-19

AI Technical Summary

Technical Problem

Existing systems lack the capability to accurately predict exchange rates and agricultural product prices, which are crucial for managing risks and making informed decisions in international trade and finance.

Method used

A system incorporating data preprocessing, converter model training, exchange rate prediction, data normalization, data smoothing, feature selection, price prediction model training, and hyperparameter optimization, utilizing advanced models like the Temporal Fusion Transformer to enhance forecasting accuracy.

Benefits of technology

Provides accurate and reliable forecasts for exchange rates and agricultural prices, enabling better decision-making, risk management, and promoting economic stability and growth.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112024018687178-PAT00001_ABST
    Figure 112024018687178-PAT00001_ABST
Patent Text Reader

Abstract

The exchange rate and agricultural product price prediction system according to the present invention includes a data preprocessing unit, a converter model training unit, and an exchange rate prediction unit. The data preprocessing unit preprocesses input foreign exchange-related data, the converter model training unit trains a converter model using the foreign exchange-related data, and the exchange rate prediction unit predicts the exchange rate using the trained converter model.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to prediction-related technology, and more specifically, to a system for predicting exchange rates and agricultural product prices. Background Technology

[0003] What is an exchange rate?

[0004] The exchange rate represents the price at which one currency can be exchanged for another. It is the ratio at which one currency is valued in relation to another. Exchange rates are determined by the foreign exchange market, where currencies are bought and sold. Exchange rates can fluctuate over time due to changes in the supply and demand of various currencies, as well as economic and political events that affect the currency market. Exchange rates are important for international trade and finance because they determine the relative value of various currencies, which in turn affects import and export costs as well as the profitability of international investments.

[0006] What is exchange rate forecasting?

[0007] Exchange rate forecasting is the process of predicting future changes in exchange rates between different currencies. The goal of exchange rate forecasting is to provide insights into potential changes in currency values. This can be useful for individuals and organizations involved in international trade, finance, and investment. Exchange rate forecasting can be achieved through various methods, including statistical models, economic models, and machine learning algorithms. These methods predict how exchange rates will change over time by analyzing factors such as economic indicators, geopolitical events, and market trends.

[0009] The Importance of Exchange Rate Forecasting

[0010] Exchange rate forecasting is a critical aspect of international trade, investment, and economic policy. By predicting changes in currency values, companies can adjust pricing and hedging strategies to mitigate risks associated with exchange rate fluctuations. Investors can also use exchange rate forecasts to make informed decisions regarding foreign investment. Governments and central banks can communicate monetary and fiscal policies based on exchange rate forecasts. Individuals and organizations can manage risks associated with exchange rate fluctuations by implementing hedging strategies using exchange rate forecasts. Overall, exchange rate forecasting is an essential tool for managing risk and making informed decisions in the context of global commerce.

[0012] What are agricultural product prices?

[0013] Agricultural prices refer to the prices at which crops and livestock are bought and sold in the market. These prices are influenced by various factors, including weather conditions, global supply and demand, government policies, and economic conditions. Fluctuations in agricultural prices can affect farmers' incomes and the cost of food for consumers. Governments frequently intervene in agricultural markets to stabilize prices and ensure food security, such as by setting minimum prices or providing subsidies to farmers. Overall, agricultural prices are a critical element of the economy, impacting farmers' livelihoods as well as food availability and affordability for consumers.

[0015] What is agricultural product price forecasting?

[0016] Agricultural price forecasting is a critical process that utilizes statistical and mathematical models to predict future prices. By analyzing historical data on factors such as supply and demand, weather patterns, global economic conditions, and government policies, these models assist farmers, traders, and policymakers in making informed decisions regarding production, marketing, and policy intervention. The accuracy of agricultural price forecasting depends on the quality of the data used, the selection of forecasting models, and the ability to predict unexpected events. This tool is used by farmers to plan production, by traders to manage inventory and prepare for price fluctuations, and by policymakers to design effective policies to ensure food security and stabilize prices.

[0018] The Importance of Agricultural Product Price Forecasting

[0019] Agricultural price forecasting plays a crucial role in various aspects, including ensuring stable income for farmers, meeting food demand, managing risks, promoting economic growth, and facilitating policy decisions. By accurately predicting future prices, agricultural price forecasting facilitates the efficient allocation of resources and guarantees food security, thereby benefiting farmers, consumers, policymakers, and the economy as a whole. Prior art literature

[0021] (Patent Document 0001) KR 102324142 B1 The problem to be solved

[0022] The present invention aims to provide a system capable of predicting exchange rates and the prices of agricultural products. means of solving the problem

[0024] To achieve the above objective, the exchange rate and agricultural product price prediction system according to the present invention includes a data preprocessing unit, a converter model training unit, and an exchange rate prediction unit. The data preprocessing unit preprocesses input foreign exchange-related data, the converter model training unit trains a converter model using the foreign exchange-related data, and the exchange rate prediction unit predicts the exchange rate using the trained converter model.

[0026] In addition, it may further include a data normalization unit that normalizes foreign exchange-related data, and a data smoothing unit that smooths foreign exchange-related data using a moving average.

[0028] In addition, the data preprocessing unit further preprocesses the input agricultural product price-related data, and

[0029] It may further include a feature selection unit that selects features among agricultural product price-related data that have a correlation with the price of agricultural products above a preset standard, a price prediction model training unit that trains a price prediction model incorporating the selected features, and an agricultural product price prediction unit that predicts the price of agricultural products using the price prediction model.

[0031] In addition, it may further include a hyperparameter optimization unit that optimizes the hyperparameters of the price prediction model. Effects of the invention

[0033] According to the present invention, it is possible to provide accurate and reliable foreign exchange forecasts that can provide users with information on the decision-making process, mitigate currency risk, and potentially bring about greater financial returns.

[0035] Furthermore, by providing accurate and reliable information on the future prices of agricultural products, it becomes possible to promote a stable and sustainable agricultural sector and contribute to overall economic growth and development. Brief explanation of the drawing

[0037] FIG. 1 is a flowchart illustrating the exchange rate prediction process performed by the exchange rate and agricultural product price prediction system of FIG. 8. FIG. 2 is a flowchart illustrating the agricultural product price prediction process performed by the exchange rate and agricultural product price prediction system of FIG. 8. Figure 3 is a flowchart for selecting useful features. Figure 4 is a flowchart for hyperparameter optimization. Figure 5 is a diagram illustrating a machine learning model. Figures 6 and 7 are diagrams for illustrating a deep learning model. FIG. 8 is a schematic block diagram of an exchange rate and agricultural product price prediction system according to one embodiment of the present invention. Specific details for implementing the invention

[0038] Hereinafter, preferred embodiments of the present invention will be described with reference to the attached drawings.

[0039] FIG. 8 is a schematic block diagram of an exchange rate and agricultural product price prediction system according to an embodiment of the present invention. In FIG. 8, the exchange rate and agricultural product price prediction system includes a data preprocessing unit (110), a converter model learning unit (120), an exchange rate prediction unit (130), a data normalization unit (140), a data smoothing unit (150), a feature selection unit (160), a price prediction model learning unit (170), an agricultural product price prediction unit (180), and a hyperparameter optimization unit (190).

[0041] The data preprocessing unit (110) preprocesses the input foreign exchange-related data, the converter model training unit (120) trains a converter model using the foreign exchange-related data, and the exchange rate prediction unit (130) predicts the exchange rate using the trained converter model. The data normalization unit (140) normalizes the foreign exchange-related data, and the data smoothing unit (150) smooths the foreign exchange-related data using a moving average.

[0043] The data preprocessing unit (110) further preprocesses the input data related to agricultural product prices, the feature selection unit (160) selects features among the data related to agricultural product prices that have a correlation with the price of the agricultural product above a preset standard, the price prediction model training unit (170) trains a price prediction model incorporating the selected features, and the agricultural product price prediction unit (180) predicts the price of the agricultural product using the trained price prediction model. Additionally, the hyperparameter optimization unit (190) optimizes the hyperparameters of the price prediction model.

[0045] The present invention will be explained in more detail below with more specific examples.

[0047] 1. Methodology of Exchange Rate Forecasting

[0048] As shown in the following flowchart, a comprehensive data preparation process is required to predict exchange rates. Figure 2 is a flowchart illustrating the exchange rate prediction process performed by the exchange rate and agricultural product price prediction system of Figure 1. First, relevant data is collected and preprocessed by handling missing values ​​and removing outliers. Second, the data is normalized if there are noticeable scale differences across various variables. Third, a moving average is applied to smooth the data and reduce noise. Next, a transducer model capable of effectively capturing the temporal dependency of the data is constructed. This model is trained using the preprocessed data with appropriate hyperparameter adjustments to optimize performance. Finally, the trained model is deployed to predict exchange rates for various periods, such as the next day, a week, or a month ahead.

[0050] The exchange rate forecasting process of Fig. 1 is distinguished from existing methods in two respects. First, data smoothing techniques are used to reduce noise and complexity in the input data. This helps improve the accuracy of the forecast by removing unwanted fluctuations that can obscure the underlying trend of the data. Second, a transducer model containing time embeddings can be utilized to accurately identify the temporal dependency between input and output variables. This, combined with the powerful learning capabilities of the transducer model, enables more accurate exchange rate forecasting.

[0052] 1.1. Data Smoothing

[0053] Data is smoothed using moving averages because it can reduce noise and volatility in time-series data, such as currency exchange rates. Smoothing involves applying mathematical functions to raw data to generate smoother and more consistent trends. This allows analysts to identify underlying patterns and trends that might be obscured by short-term fluctuations, which can aid in exchange rate forecasting. By smoothing data, analysts can make more accurate predictions of future exchange rates, which can be useful for traders, investors, and other stakeholders who rely on accurate rate forecasts. Additionally, smoothing helps reduce the impact of data outliers and other anomalies, thereby improving the overall quality of forecasting models.

[0055] 1.2. Transformers and Time Embeds

[0056] We predict exchange rates using state-of-the-art transducer models and time embeddings. This model is particularly effective because it is designed to process time-series data, a data type commonly used in the financial and economic fields.

[0057] The key difference between this model for exchange rate forecasting and historical methods is the use of a transducer model and time embeddings. Traditional methods for exchange rate forecasting often relied on simple statistical models that were not designed to handle time-series data with long-term dependencies. In contrast, the transducer model used in this approach is specifically designed to capture the complex relationships between data points in time-series data, making it highly suitable for forecasting exchange rates over long periods.

[0059] 2. Methodology for Forecasting Agricultural Product Prices

[0060] Figure 3 illustrates the process of predicting agricultural product prices. Figure 3 is a flowchart illustrating the agricultural product price prediction process performed by the exchange rate and agricultural product price prediction system of Figure 1. Initially, data related to agricultural product prices is loaded, and then the data is cleaned by processing missing values ​​or outliers. Next, features are evaluated, and features exhibiting a strong correlation with prices are selected. Then, the construction of a statistical, machine learning, or deep learning model is carried out. The selected features are input into the model, and model training is started. Hyperparameter optimization is used to improve the prediction accuracy of the model. Through this optimization process, parameters such as batch size, sequence length, learning rate, network depth, and hidden size can be automatically optimized. Ultimately, we use the model trained with optimized hyperparameters to predict daily, weekly, and monthly agricultural product prices.

[0062] The agricultural product price prediction process shown in Figure 2 is distinguished from existing methods in three aspects. First, feature selection is integrated into the process to automatically select the most useful features for price prediction when processing large amounts of data. Second, hyperparameter optimization is included to enable machine learning and deep learning models to automatically select the most appropriate hyperparameters. This improves the accuracy of the prediction. Third, a more advanced Temporal Fusion Transformer model is utilized, which can not only provide more accurate predictions but also predict price fluctuation intervals. This is particularly useful for risk management purposes.

[0064] 2.1. Feature Selection

[0065] Feature selection involves identifying the most relevant features or variables that have a significant impact on the target variable (in this case, agricultural product prices). By selecting the most useful features, data dimensionality can be reduced, model performance improved, and the risk of overfitting lowered. The feature selection process used in this agricultural product price prediction model is performed automatically based on statistical techniques such as correlation analysis, mutual information, and regression coefficients. By integrating feature selection into the model, accurate price predictions can be achieved by utilizing the most important features.

[0066] As shown in Figure 3, clean data must first be loaded to select useful features. Figure 3 is a flowchart for selecting useful features. If the data includes text, images, or other unstructured data, features must be extracted from that data. Various feature selection algorithms can be used depending on the types of input and output variables. For numerical inputs and outputs, Pearson or Spearman correlation can be used. If the input is numeric and the output is categorical, ANOVA can be used. For categorical inputs and numerical outputs, Kendall's correlation can be used. If both the input and output are categorical, chi-square or mutual information can be used. Ultimately, we will significantly improve the model by identifying features that have a high correlation with the output variable.

[0068] 2.2. Hyperparameter Optimization

[0069] Hyperparameter optimization is another critical step in building accurate machine learning and deep learning models. Hyperparameters are settings that are established before training, even though the model does not learn them. Since these settings significantly impact model performance, they must be carefully tuned to achieve optimal results. The hyperparameter optimization process used in this model involves selecting the most suitable values ​​for parameters such as the learning rate, batch size, and the number of hidden layers. This is accomplished using techniques such as grid search, random search, or Bayesian optimization. By optimizing hyperparameters, one can verify that the model is properly configured to make accurate price predictions. This can lead to improved performance, reduced training time, and increased reliability of predictions.

[0071] As shown in Figure 4, to optimize hyperparameters, we first define the hyperparameters to be adjusted, such as batch size, hidden size, number of layers, and learning rate. Figure 4 is a flowchart for hyperparameter optimization. Next, we set up a search space for these hyperparameters. We select an optimization algorithm based on three conditions. If the amount of hyperparameters or the size of the search space is small, we use grid search; otherwise, we use random search. If computational resources are insufficient, we use random or grid search, but if resources are sufficient, we can use Bayesian optimization or evolutionary algorithms. Finally, if we lack domain expertise regarding Bayesian optimization or evolutionary algorithms, we must stick to random or grid search.

[0073] 2.3. Building Statistical, Machine Learning, and Deep Learning Models

[0074] Statistical Models: Both ARIMA and multiple linear regression are statistical models. In an ARIMA model, time-series data must be fixed; that is, statistical characteristics remain constant over time. Failure to meet this requirement may result in biased or inaccurate outcomes. Meanwhile, multiple linear regression models assume a linear relationship between the dependent and independent variables, which may not be accurate when the relationship is more complex.

[0076] Machine Learning Models: Support Vector Regression (SVR) is a machine learning algorithm used for regression tasks that predict continuous values ​​based on input data by splitting data into two classes using lines or hyperplanes. Random Forest and Gradient Boosting are two ensemble learning methods that combine multiple models to perform more accurate predictions. Random Forest uses bagging to train multiple models independently and combines them by taking the average of their predictions, whereas Gradient Boosting trains models sequentially, with each model learning from the error of the previous model.

[0077] SVR is expensive, sensitive to kernel selection, and may not perform well on non-linear data. Random Forest may not perform well on datasets with high-dimensional space and can be affected by noisy data. Interpreting the model's predictions can also be difficult. On the other hand, Gradient Boosting is prone to overfitting if the model is too complex or the dataset is too small. Additionally, it is computationally expensive and may require extensive hyperparameter tuning to achieve optimal performance. Figure 5 is a diagram illustrating machine learning models.

[0079] Deep Learning Models: Recurrent Neural Networks (RNNs) are a type of neural network designed to process sequential data, such as time series or natural language. RNNs can utilize previous input information to the network for more accurate predictions. Long Short-Term Memory (LSTM) is a special type of RNN that can learn long-term dependencies from sequential data by using memory cells capable of storing information for a long period. Gated Recurrent Unit (GRU) is another type of RNN that controls the flow of information into and out of memory cells using a gating mechanism. Figures 6 and 7 are diagrams illustrating deep learning models.

[0081] The Temporal Fusion Transformer (TFT) is a neural network model designed to process sequential data. The TFT is based on a transformer architecture that processes input data in a parallelizable manner using a self-attention mechanism. The model first selects the most prominent variables and extracts features from static metadata associated with the input variables using a static covariate encoder. It then uses static reinforcement layers to enhance the temporal characteristics of the input variables. Temporal self-attention layers learn the long-term dependencies of the time-series data, enabling the model to make more accurate predictions and provide interpretability. The output is then processed by a positional feedforward layer, and the TFT outputs the interval of the predicted values ​​for the time-series variables using a quantile loss function.

[0083] 3. Conclusion

[0084] 3.1. Exchange Rate Forecast

[0085] Exchange rate forecasting is crucial for various users, including businesses, investors, governments, and individuals, to manage currency risk and make informed decisions. The ultimate goal of exchange rate forecasting is to provide users with accurate and reliable predictions that inform the decision-making process, mitigate currency risk, and potentially lead to greater financial returns.

[0086] Proposed smoothing techniques, such as moving averages or exponential smoothing, can help identify underlying trends and patterns more easily by removing noise and volatility from the data. This can lead to more accurate and stable predictions, particularly when the data is noisy or exhibits extreme fluctuations. Furthermore, transducer models can bring several advantages to exchange rate forecasting. First, transducer models can process multiple inputs and capture non-linear trends and patterns in the data. This is important for finance, where data can be complex and dynamic. Second, transducer models incorporate time embeddings to account for temporal dependencies between data points, which are crucial for predicting trends over time. Third, due to their ability to capture complex relationships between data points, transducer models have the potential to outperform existing statistical models in terms of accuracy and reliability.

[0088] 3.2. Agricultural Product Price Forecasting

[0089] Accurate forecasting of agricultural prices plays a crucial role in enabling farmers, merchants, policymakers, and consumers to make informed decisions and manage risks efficiently. The ultimate goal is to promote a stable and sustainable agricultural sector and contribute to overall economic growth and development by providing accurate and reliable information on future prices.

[0090] The proposed feature extraction and hyperparameter optimization not only improve the performance of all models but also significantly reduce labor and time consumption. Furthermore, the TFT model is advantageous for risk management as it can predict not only accurate forecasts but also fluctuation intervals. Based on reliable agricultural price forecasts, farmers can plan production levels, select crops, and manage finances, while traders and processors can manage risks and plan supply chains more efficiently. Policymakers can gain valuable insights into market trends to develop policies that support the agricultural sector, and consumers can benefit from a stable food supply at affordable prices.

[0092] Although the present invention has been described by some preferred embodiments, the scope of the invention should not be limited thereto and should extend to variations or improvements of the embodiments supported by the claims. Explanation of the symbols

[0094] 110: Data Preprocessing Unit 120: Transformer model learning section 130: Exchange Rate Forecasting Department 140: Data Normalization Section 150: Data Smoothing Section 160: Feature selection section 170: Price Prediction Model Training Unit 180: Agricultural Product Price Forecasting Department 190: Hyperparameter Optimization Section

Claims

Claim 1 An exchange rate and agricultural product price prediction system comprising: a data preprocessing unit for preprocessing input foreign exchange-related data; a converter model training unit for training a converter model using the foreign exchange-related data; and an exchange rate prediction unit for predicting an exchange rate using the converter model, wherein the data preprocessing unit further preprocesses input agricultural product price-related data and a feature selection unit for selecting features among the agricultural product price-related data that have a correlation with the price of the agricultural product above a preset standard; a price prediction model training unit for training a price prediction model incorporating the selected features; and an agricultural product price prediction unit for predicting the price of the agricultural product using the price prediction model.The system further includes a hyperparameter optimization unit that selects an optimization algorithm according to preset conditions and updates the hyperparameters of the price prediction model by the selected optimization algorithm; the feature selection unit selects the feature when the agricultural product price-related data includes unstructured data, and selects the feature using different feature selection algorithms depending on the types of input and output variables using preset statistical techniques; it uses Pearson or Spearman correlation when both input and output variables are numeric, ANOVA when the input variable is numeric and the output variable is categorical, Kendall correlation when the input variable is categorical and the output variable is numeric, and a chi-square or mutual information algorithm when both input and output variables are categorical; the hyperparameter optimization unit uses grid search or random search when computational resources are below a preset threshold, and uses a Bayesian optimization algorithm or an evolutionary algorithm when computational resources are above the preset threshold; the agricultural product price prediction unit further predicts the fluctuation interval of agricultural product prices using a quantile loss function based on a Temporal Fusion Transformer (TFT); and the transformer model [is based on] input variables and output variables An exchange rate and agricultural product price prediction system characterized by including time embeddings to identify temporal dependencies between, and designed to predict the exchange rate by capturing the relationship between points of time series data. Claim 2 An exchange rate and agricultural product price prediction system according to claim 1, further comprising: a data normalization unit for normalizing foreign exchange-related data; and a data smoothing unit for smoothing foreign exchange-related data using a moving average. Claim 3 delete Claim 4 delete

Citation Information

Patent Citations

  • Method for Predicting Price of Product

    KR1020220151453A

  • Method and server for predicting exchange rate

    KR102171252B1

  • Matching system for residual agricultural product

    KR102324148B1

  • Method and apparatus of predicting price of artwork based on artificial intelligence

    KR102491607B1