Multi-Scenario-Based Bid Price Prediction System and Method

Abstract

The present invention relates to a multi-scenario-based bid price prediction system and method that learns monetary information such as base amounts included in past bid data, competition information related to the number of competitors or competition intensity, information on the ordering agency, and the winning bid amount using an artificial intelligence model, and provides a user with a bid price corresponding to one or more bidding scenarios among conservative, balanced, and aggressive types based on the winning bid price predicted for a new bid.

Description

Technology Field

[0001] The present invention relates to a multi-scenario-based bid price prediction system and method, and more specifically, to a multi-scenario-based bid price prediction system and method that learns monetary information such as base amounts included in past bid data, competition information related to the number of competitors or competition intensity, information on the ordering agency, and the winning bid amount using an artificial intelligence model, and provides a user with a bid price corresponding to one or more of a bidding scenario among conservative, balanced, and aggressive types based on the winning bid price predicted for a new bid. Background Technology

[0002] Generally, the process of selecting contractors for construction, services, or goods procurement projects commissioned by the government, local governments, and various public institutions is carried out through electronic bidding.

[0003] This electronic bidding is a method that allows bidders to participate in bidding for the procurement of goods or construction of facilities via the internet without the need to visit the bidding location in person to submit bidding documents. The electronic bidding system was first introduced by the Public Procurement Service in November 2000 and has since been expanded to government agencies at all levels, local governments, and public institutions. Since September 2002, most public institutions have been required to post bid announcements through the Public Procurement Service's National Electronic Procurement System (Nara Jangteo).

[0004] The bidding officer of a public institution prepares the bid announcement using an electronic bidding management system, electronically receives bids submitted by participating companies, performs automatic bid opening, and selects the successful bidder.

[0005] These electronic bidding announcements include various information such as the announcement name, announcement number, region, license information, base price, estimated price, net construction cost, bid rate, evaluation rate range, bidding date and time, and bid deadline.

[0006] The above base amount is a standard amount for calculating the bid amount and serves as the reference amount for determining multiple preliminary prices. Generally, the base amount is calculated by adding value-added tax to the estimated price.

[0007] The above estimated price is an amount that determines the scope of application of the qualification criteria and is used as a standard amount to determine the bidding method and contract method.

[0008] The assessment rate range is generally set to 97% to 103% or 98% to 102%, and the ordering agency system generates multiple preliminary prices based on the base amount.

[0009] Generally, multiple preliminary prices consist of 15 preliminary prices generated by dividing the evaluation rate range into specific intervals, and the estimated price is calculated by drawing four of these preliminary prices during the bidding process and calculating the arithmetic mean. These 15 preliminary prices are generated by the contracting authority's system and are generally not disclosed externally.

[0010] Since past bidding data must be classified and analyzed according to various conditions such as region, industry, license type, and characteristics of the ordering agency, it is very difficult to derive the optimal bid amount during the actual bidding process.

[0011] Korean Registered Patent No. 10-2676472, instead of a method for deriving the optimal bid amount, sets the number of data points secured by the user from multiple preliminary price data as the "Bid Score" and uses it to predict the probability distribution of the estimated price. However, while the aforementioned registered patent predicts the probabilistic distribution of the estimated price, it still has limitations in that it cannot directly calculate or present the bid amount that the actual bidder must submit.

[0012] In addition, Korean Published Patent No. 10-2023-0152517 discloses a prediction system that calculates a high-probability bid evaluation rate by applying a time-series prediction model using data on multiple preliminary prices opened in the past and preliminary prices drawn. The aforementioned published patent proposes a method for predicting the estimated price evaluation rate using multiple preliminary prices and draw numbers, and estimating a competitor's bid amount based on this.

[0013] However, the aforementioned registered and published patents have a problem in that they merely utilize the estimated price or the estimated price assessment rate as the primary variables and fail to sufficiently consider various variables—such as competition intensity, characteristics of the contracting authority, and past winning bid patterns—which are important factors influencing actual bidding results. Furthermore, they are limited in that they provide only a single predicted value and fail to present diverse bidding strategies based on the strategic needs of users or participating companies.

[0014] Furthermore, since the aforementioned registered and disclosed patents rely on a method of predicting the estimated price, there is a problem in that it is difficult to directly predict the bid amount or winning bid amount of actual bidders, and there are limitations in that they fail to provide information necessary for decision-making, such as the probability of winning the predicted winning bid or the expected profit margin. The problem to be solved

[0015] The present invention provides a bid price prediction system that predicts the bid price by considering not only monetary information such as the base price and the estimated price, but also various variables that affect the actual bidding results, such as competition intensity, characteristics of the ordering agency, and past winning bid patterns.

[0016] The present invention provides a bid price prediction system that provides a user with information such as bid prices corresponding to various bidding strategies, the probability of winning the bid, or the expected profit margin. means of solving the problem

[0017] In one aspect, the present invention

[0018] A data collection unit (10) that collects multiple past bidding data;

[0019] A feature extraction unit (20) that extracts feature variables including price-related variables, competition-related variables and ordering agency-related variables, and the winning bid amount of the corresponding past bid from the collected past bid data;

[0020] A machine learning unit (30) having a learning unit (31) that learns one or more machine learning models using the above feature variables as input data and the above winning bid amount as a target value, and a prediction unit (32) that derives the winning bid amount through the machine learning model learned in the learning unit (31) using input data including the base amount and ordering agency information included in the new bid announcement; and

[0021] The present invention relates to a multi-scenario-based bid price prediction system comprising a scenario generation unit (40) that generates a bid price corresponding to one or more bidding scenarios based on the predicted winning bid price.

[0022] In another aspect, the present invention

[0023] Step of collecting multiple past bidding data;

[0024] A step of extracting feature variables including amount-related variables, competition-related variables, and ordering agency-related variables, and the winning bid amount of the corresponding past bid;

[0025] A step of training one or more machine learning models using the above feature variables as input data and the above winning bid amount as the target value;

[0026] A step of predicting the winning bid price through the above-mentioned trained machine learning model using input data including the base amount and ordering agency information included in the new bid announcement;

[0027] The present invention relates to a multi-scenario-based bid price (bid price) prediction method comprising the step of generating bid prices corresponding to one or more bidding scenarios based on the predicted winning bid price. Effects of the invention

[0028] The multi-scenario-based bid price prediction system and method of the present invention can provide a more accurate winning bid price compared to existing technologies that rely on the estimated price and its probability distribution by predicting the winning bid price through a machine learning model that learns various variables, such as the number of competitors, competition intensity, characteristics of the ordering agency, and past winning bid patterns, which affect not only the estimated price but also the actual winning bid result.

[0029] More specifically, the present invention can improve prediction accuracy by predicting the winning bid price using a plurality of machine learning models including Random Forest and XGBoost, and calculating the final winning bid price by weighting the prediction results of the plurality of machine learning models.

[0030] In addition, the present invention enables the prediction of the winning bid price that reflects the characteristics of the ordering agency by training an independent machine learning model (submodel) for each ordering agency using past bidding data grouped by ordering agency, thereby further improving the accuracy of the winning bid price prediction.

[0031] The present invention can calculate one or more bidding scenarios among conservative, balanced, and aggressive types, and bid prices corresponding to each bidding scenario, by considering the probability of winning the bid and the expected profit margin. Accordingly, the present invention has the effect of enabling a user to select a reasonable bid price according to their bidding strategy by providing bid prices corresponding to various bidding strategies based on the winning bid price predicted by an artificial intelligence model.

[0032] In addition, the system of the present invention can improve the interpretability and reliability of the prediction results by analyzing the contribution of each feature variable to the prediction result using an explainable artificial intelligence analysis method and providing this to the user. Brief explanation of the drawing

[0033] Figure 1 is a configuration diagram of the multi-scenario-based bid price prediction system of the present invention. Figure 2 is a flowchart of the multi-scenario-based bid price prediction method of the present invention. Figure 3 is a conceptual diagram of bidding strategies such as conservative and aggressive types according to percentiles in a normal distribution. Figure 4 is an example of a scenario-based bid price according to a position corresponding to a percentile. Specific details for implementing the invention

[0034] Hereinafter, preferred embodiments of the present invention are described with reference to the drawings. However, the scope of the present invention is not limited to the description of the embodiments or the drawings below. That is, the terms used in this specification are used merely to describe specific embodiments and are not intended to limit the present invention. Unless the context clearly indicates otherwise, the singular expression includes the plural expression. Furthermore, terms such as "comprising" or "having" as used in this specification are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0035] FIG. 1 is a configuration diagram of a multi-scenario-based bid price prediction system of the present invention, FIG. 2 is a flowchart of a multi-scenario-based bid price prediction method of the present invention, FIG. 3 is a conceptual diagram of bidding strategies such as conservative and aggressive types according to percentiles in a normal distribution, and FIG. 4 is an example of a scenario-based bid price according to a position corresponding to a percentile.

[0036] Referring to FIGS. 1 to 4, the multi-scenario-based bid price prediction system of the present invention may include a data collection unit (10), a feature extraction unit (20), a machine learning unit (30), a scenario generation unit (40), and a contribution analysis unit (50).

[0037] In addition, the system may include a control unit (60) for controlling the operation of each of the above components.

[0038] The control unit (60) can control the data flow between the data collection unit (10), feature extraction unit (20), machine learning unit (30), scenario generation unit (40), and contribution analysis unit (50), and can manage the operation of each component in an integrated manner.

[0039] For example, the control unit (60) can control the transmission of bid data collected by the data collection unit (10) to the feature extraction unit (20), and the transmission of feature variables generated by the feature extraction unit (20) to the machine learning unit (30) to perform a winning bid price prediction.

[0040] In addition, the control unit (60) can control the generation of multiple bid price scenarios by transmitting the predicted winning bid price calculated by the machine learning unit (30) to the scenario generation unit (40).

[0041] The above data collection unit (10), feature extraction unit (20), machine learning unit (30), scenario generation unit (40), and contribution analysis unit (50) may be implemented in the form of a software program or module, and the above control unit (60) may be implemented by program code executed by a processor.

[0042] The above data collection unit (10) performs the function of collecting past data regarding public bidding from an external system and storing it in a database. The above data collection unit may be configured to automatically or periodically collect bidding-related data from one or more of the Government-to-Business (G2B), the Public Procurement Service system, the public institution bidding announcement system, or the private bidding information provision system.

[0043] The above data collection unit can collect bidding information using an API (Application Programming Interface), web crawling, or data file download methods, and the collected data can be received in JSON, XML, or CSV format.

[0044] The bidding data collected by the above-mentioned data collection unit may include the announcement number, bidding announcement name, base price, estimated price, winning bid amount, winning bid rate, number of participating companies, ordering agency, etc. Table 1 below is an example of bidding data.

[0045] Column name Example value explanation Announcement No. 20250107-001 Unique ID for Bid Announcement Bid Announcement Name Establishment of security system for OO City Hall Bid Title base amount 500,000,000 won The base amount calculated by the contracting authority Estimated price 487,325,000 won Actual winning bid standard amount Winning bid amount 475,123,456 won Final winning bid amount Successful bid rate 97.49% (Winning Bid Amount / Estimated Price) × 100 Number of bidding companies 7 Number of companies participating in the bid Ordering agency OO City Hall Name of the public institution that placed the order Bid Date 2024-12-15 Bidding Date

[0046] The above data collection unit can perform missing value processing and outlier removal processes to preprocess the collected data. For example, if the percentage of missing values ​​in a specific column exceeds a preset threshold (e.g., 30%), the column may be removed, or the missing values ​​may be replaced using the median or mean.

[0047] In addition, to remove abnormal data where the winning bid rate is outside the normal range, data where the winning bid rate is outside the preset range (e.g., 70% to 120%) can be determined as an abnormal value and removed.

[0048] The above data collection unit may be configured to store preprocessed data in a database, and the database may be a PostgreSQL, MySQL, NoSQL database, or a cloud-based data storage.

[0049] In addition, the data collection unit can automatically detect when new bidding data is generated and add it to the database, and the stored data can subsequently be converted into feature variables for machine learning training by the feature extraction unit and used.

[0050] By automatically collecting and managing various bidding-related data through such a data collection unit, training data for analyzing bidding patterns and predicting winning bid prices can be reliably secured.

[0051] Meanwhile, the above data collection unit can generate data groups based on the collected data, such as ordering agency, industry, region, and base price range, and the data groups can be used as filter conditions to form a set of similar bid data when training a machine learning model.

[0052] The above feature extraction unit (20) performs the function of extracting feature variables used for machine learning model training and the winning bid amount of the corresponding past bid from past bid data collected from the above data collection unit.

[0053] The above feature extraction unit (20) can perform data transformations such as normalization or log transformation on the feature variables to make them suitable for training a machine learning model.

[0054] The above feature extraction unit (20) can generate a plurality of feature variables including an amount-related variable, a competition-related variable, and an ordering agency-related variable from the collected data.

[0055] First, the feature extraction unit (20) may extract or generate a base amount, a ratio of the planned price to the base amount, and a relative base amount as amount-related variables, provided, but is not limited thereto.

[0056] The base amount refers to the standard amount presented by the contracting authority in the bid announcement, which serves as the basis for calculating multiple preliminary prices.

[0057] The ratio of the estimated price to the base price can be defined as the value obtained by dividing the estimated price by the base price, and can be calculated as follows.

[0058] Ratio of Estimated Price to Base Price = Estimated Price / Base Price

[0059] The above ratio of the estimated price to the base amount can be used as a variable to normalize price deviations by bid announcement, thereby enabling the comparison of bid data of different scales.

[0060] The relative base amount is a variable used to indicate the scale of a specific bid's base amount within bids of the same industry or type, and can be calculated by dividing the base amount by the median base amount of the same industry and then log-transforming it.

[0061] The relative base amount can be calculated as follows.

[0062] Relative base amount = log(base amount / median base amount by industry)

[0063] Such relative base amount variables can play a role in mitigating the impact of differences in bid size and enabling a single model to learn bid data of various sizes.

[0064] Next, the feature extraction unit (20) may include, as competition-related variables, the number of participating companies, the average number of participating companies by industry, and the intensity of competition, but is not limited thereto.

[0065] The number of participating companies refers to the number of companies that actually participated in and submitted bids for the relevant tender.

[0066] The average number of participating companies by industry refers to the average number of companies that participated in past bidding data for the same industry or type of work, and can be calculated from past bidding data.

[0067] Competition intensity is a variable used to quantitatively represent the level of competition in a specific bid, and it can be calculated by applying a log transformation to the ratio of the number of participating companies to the average number of participating companies by industry.

[0068] Competition intensity can be calculated as follows.

[0069] Competition intensity = log(Number of participating companies / Average number of participating companies by industry)

[0070] The above competition intensity variable can be used to reflect the degree to which the level of competition is higher or lower compared to the average of the same industry.

[0071] In addition, the above feature extraction unit (20) may include, as variables related to the ordering agency, the ordering agency ID, the ordering agency’s past winning bid rate, or the standard deviation of the winning bid rate in the ordering agency’s recent N bids, but is not limited thereto.

[0072] The ordering agency ID is identification information for identifying the ordering agency and can be converted using one-hot encoding for processing by a machine learning model.

[0073] The past success rate of a contracting agency refers to the average success rate of bids previously conducted by the relevant contracting agency, and can be calculated, for example, as the average success rate of the last N bid data points.

[0074] The past winning bid rate of the ordering agency can be calculated as follows.

[0075] Ordering Agency's Past Success Rate = Average of the Success Rates for the Last N Cases

[0076] Here, the successful bid rate can be defined as the ratio of the successful bid amount to the estimated price.

[0077] The standard deviation of the above successful bid rate can be used as a variable representing the price volatility of the bidding results of the relevant ordering agency.

[0078] In addition, the feature extraction unit (20) may use bidding classification information, such as construction, goods, and services, as feature variables to reflect bidding type information, and the classification information may be converted into a one-hot encoding method for machine learning model training.

[0079] In addition, the feature extraction unit (20) may use information on the month in which the bidding was performed as a feature variable to reflect the seasonality or seasonal pattern of the bidding period.

[0080] Examples of the above feature variables are as shown in Table 2 below.

[0081] variable name explanation Example value X1: Base amount The base amount calculated by the contracting authority 500,000,000 won X2: Ratio of estimated price to base price Estimated Price / Base Price 0.9745 X3: Number of participating companies Number of bidding companies 7 X4: Ordering Agency's Past Success Rate Average winning bid rate of the institution's last 10 cases 97.2% X5: Average number of participating companies by industry Average number of participating companies in the same industry 6.3 X6: Ordering Agency ID One-hot encoding [1, 0, 0, ...] X7: Month Reflecting bidding seasonality (1~12) 12 X8: Classification Construction / Goods / Services (One-Hot Encoding) [0, 1, 0] X9: Competition Intensity log(Number of participating companies / Industry average) 0.15 X10: Relative base amount log(base amount / industry median) 0.23

[0083] Meanwhile, the key terms used in the present invention are defined as follows.

[0084] The winning bid amount refers to the amount submitted by the company that was ultimately awarded in past bids; generally, it is the lowest valid bid amount determined within the range that satisfies the eligibility or price evaluation criteria among the bid amounts submitted by participating companies.

[0085] The winning bid price refers to the amount predicted by an artificial intelligence machine learning model. The winning bid price is calculated by a machine learning model trained using feature variables generated through the data collection unit and feature extraction unit, and is the amount predicted by the trained machine learning model when input data for a new bid announcement is provided.

[0086] The bid price refers to the amount submitted by a user participating in the bidding process during the actual bidding process. In the present invention, bid prices corresponding to various bidding strategies can be generated based on the winning bid price predicted by an artificial intelligence model.

[0087] For example, in a balanced scenario, the predicted winning bid price can be used as the bid price, and in a conservative or aggressive scenario, the bid price can be generated by performing addition or subtraction operations within a certain range based on the winning bid price. For example, the bid price can be calculated by adding or subtracting the value obtained by multiplying by a weight (k) to the predicted winning bid price.

[0088] The bid price generally refers to a price formed within a certain percentage range based on the estimated price set in the bid announcement, and in the present invention, the bid price can be defined as a price range that a user can select and submit during the actual bidding process.

[0089] The machine learning unit (30) above can predict the winning bid price by linking with an AI model built on a server or an external AI service via an API (Application Programming Interface).

[0090] The machine learning unit (30) may include a learning unit (31) that learns one or more artificial intelligence models and a prediction unit (32) that predicts the winning bid price using the learned models.

[0091] The above artificial intelligence machine learning model may use various known machine learning algorithms, such as Random Forest, Gradient Boosting Machine, XGBoost (eXtreme Gradient Boosting), Support Vector Machine, and Neural Network.

[0092] Preferably, multiple machine learning models including Random Forest and XGBoost may be used to improve prediction performance.

[0093] Random Forest is an ensemble learning method that generates multiple decision trees and averages the prediction results of each tree to produce a final prediction value, and can provide stable prediction performance while preventing overfitting of the data.

[0094] XGBoost (eXtreme Gradient Boosting) is a gradient boosting-based machine learning algorithm that can provide high prediction accuracy by sequentially training trees to compensate for errors in the previous model.

[0095] The above learning unit (31) can train one or more machine learning models using feature variables generated by the above feature extraction unit as input data and the actual winning bid amount from past bids as the target value.

[0096] The above-mentioned learning unit can perform model training and performance verification by separating training data and validation data, and can improve the model's prediction performance through cross-validation or hyperparameter tuning.

[0097] In addition, the learning unit can train a plurality of machine learning models individually and generate a final prediction model by combining the prediction results of the plurality of models using a weighted average method.

[0098] The prediction unit (32) can predict the winning bid price using a machine learning model learned in the learning unit (31) by using input data including the base amount, ordering agency information, bid type, expected number of participating companies, etc. included in the new bid announcement.

[0099] The prediction unit receives winning bid prices predicted by a plurality of machine learning models, and can calculate the final winning bid price by weighting the prediction results or combining them in an ensemble manner.

[0100] For example, the prediction unit (32) can provide input data corresponding to the base amount of a new bid, information on the ordering agency, the bid type, and the expected number of participating companies to the Random Forest and XGBoost models, respectively, and the Random Forest model and the XGBoost model can each predict the winning bid price.

[0101] The prediction unit (32) can calculate an average value by summing the winning bid prices predicted from the two models, or calculate the final winning bid price by applying a pre-set weight to each model. For example, the final winning bid price can be calculated using a weighted average method by assigning a weight of 0.4 to the Random Forest model and a weight of 0.6 to the XGBoost model.

[0102] In addition, the prediction unit (32) may predict the winning bid price by sequentially combining a Random Forest and an XGBoost model.

[0103] For example, an initial winning bid price can be predicted using a random forest model, and a correction value can be calculated by the XGBoost model learning the residual corresponding to the difference between the initial winning bid price predicted by the random forest model and the actual winning bid price.

[0104] The prediction unit (32) can calculate the final winning bid price by combining the initial winning bid price predicted by the random forest model and the correction value calculated by the XGBoost model.

[0105] In this way, by combining multiple machine learning models to predict the winning bid price, prediction accuracy can be improved compared to using a single model.

[0106] The machine learning unit (30) above can learn independent sub-models for each ordering agency using past bidding data grouped by ordering agency ID.

[0107] For example, the data collection unit can collect past bidding data ordered by a specific ordering agency (Agency A), and the feature extraction unit can extract feature variables and the winning bid amount of the corresponding past bid from the past bidding data.

[0108] The above learning unit (31) can train one or more machine learning models using the above feature variables as input data and the above winning bid amount as the target value.

[0109] That is, the system of the present invention can learn independent sub-models for each ordering agency using the artificial intelligence model.

[0110] When a new bid announcement is issued by a specific ordering agency (Agency A), the prediction unit (32) can predict the winning bid price by selecting a submodel learned using past bid data corresponding to the ordering agency ID.

[0111] For example, if the number of training data for the ordering agency is greater than a pre-set threshold (e.g., 30 or more), the prediction unit (32) can predict the winning bid price using a sub-model corresponding to the ordering agency ID.

[0112] Meanwhile, if the number of past bidding data for a specific ordering agency is less than the threshold value, the prediction unit can predict the winning bid price using a common model learned using all bidding data.

[0114] The system of the present invention may include a multi-scenario generation unit (40) for generating multiple bid prices corresponding to various bidding strategies based on the winning bid price predicted by an artificial intelligence model.

[0115] First, the scenario generation unit (40) can calculate a statistical distribution of winning bid amounts using a plurality of past winning bid amounts extracted from past bidding data, and calculate the standard deviation (σ) of the winning bid amounts based on the statistical distribution.

[0116] For example, the above standard deviation can be calculated using the deviation between each winning bid amount and the average, based on the average of the winning bid amounts of all past bidding data used for learning in the above learning unit.

[0117] In addition, the above scenario generation unit (40) can reflect the distribution characteristics of the winning bid amount by each ordering agency by calculating the average and standard deviation of the winning bid amount for past bidding data grouped by ordering agency.

[0118] In addition, the above scenario generation unit (40) can select past bidding data similar to the new bidding announcement and ordering conditions to be predicted, and calculate the standard deviation using the winning bid amount included in the similar bidding data.

[0119] For example, the above similar past bidding data may be selected based on one or more conditions, such as the ordering agency, order name or type of construction, base amount, project amount, and bid type.

[0120] More specifically, the scenario generation unit (40) can select the past bidding data as similar bidding data even if the ordering agency is different, if the ordering name or type of construction is the same or similar, and the difference between the base amount of the new bidding announcement and the base amount of the past bidding data is within a preset range (e.g., ±10%).

[0121] The above scenario generation unit (40) can provide various bid price scenarios in response to the user's bidding strategy.

[0122] More specifically, the above scenario generation unit (40) can provide one or more bidding scenarios among a conservative scenario, a balanced scenario, and an aggressive scenario, taking into account the probability of winning the bid and the expected profit margin.

[0123] For example, the conservative scenario is a strategy that proposes a relatively high bid price; it is characterized by a high probability of winning the bid but a low expected profit margin, and can be a suitable strategy when aiming to secure a winning bid stably.

[0124] The aggressive scenario is a strategy of submitting a relatively low bid price; while the expected profit margin is high, the probability of winning the bid may be low, and it can be applied when aiming to secure high profits through price competition.

[0125] In addition, the balanced scenario is a strategy generated by considering the balance between the probability of winning the bid and the expected profit margin, and can be utilized in a general bidding environment.

[0126] For example, the above scenario generation unit (40) can be selected as a conservative type when the probability of winning the bid is 75% or higher, as a balanced type when the probability of winning the bid is 40-60%, and as an aggressive type when the probability of winning the bid is 25% or lower.

[0127] In addition, the above scenario generation unit (40) can generate conservative, balanced, and aggressive bid prices based on one or more of the following: a position corresponding to a percentile in the distribution of past winning bid amounts, a range of variation based on the standard deviation, competition intensity, or winning bid patterns by ordering agency.

[0128] The above scenario generation unit (40) can generate a bid price corresponding to each bidding scenario (aggressive type or conservative type, etc.) by adding or subtracting a value obtained by multiplying the standard deviation by a weight (k) using the above predicted winning bid price as a reference value.

[0129] The bid price for each scenario can be calculated as shown in the following mathematical formula 1.

[0130] [Mathematical Formula 1]

[0131] T=T pred + kσ

[0132] Here,

[0133] T : Bid price

[0134] T pred : Predicted winning bid price

[0135] σ : Standard deviation of past winning bid amounts

[0136] k: is the weight

[0137] The above weights may be selected based on the position (Z-score) corresponding to the percentile, the range of variation based on the standard deviation, or the intensity of competition.

[0138] Here, the percentile refers to the value at the position corresponding to a specific ratio when the total past winning bid amounts are sorted in order of size.

[0139] Figure 3 illustrates a conceptual diagram explaining the positions corresponding to each percentile when a general statistical distribution is assumed to be in the form of a normal distribution.

[0140] The x-axis of Figure 3 represents standardized values ​​(z-scores) indicating where each value is located relative to the mean, and this can serve as an indicator representing the relative position corresponding to a specific quantile in a normal distribution.

[0141] As shown in FIG. 3, the z-score corresponding to the 25th percentile can be approximately -1.08, the 50th percentile can be 0, and the 75th percentile can be approximately +0.67, and in the present invention, these position values ​​can be used as examples of weights (k) for determining a bidding strategy.

[0142] For example, for an aggressive bidding strategy, the weight (k) can be set to approximately -1.08, for a balanced bidding strategy, the weight (k) can be set to 0, and for a conservative bidding strategy, the weight (k) can be set to approximately +0.67.

[0143] Through this, the above scenario generation unit (40) can generate multiple bid prices having different risk levels centered on the predicted winning bid price.

[0144] Meanwhile, the above scenario generation unit (40) can generate conservative, balanced, and aggressive bidding scenarios using a range of variation based on the standard deviation of the past winning bid amount.

[0145] That is, the above scenario generation unit can generate multiple bid price ranges using the winning bid price calculated by the above prediction unit (32) as a reference value and the standard deviation (σ) representing the statistical dispersion of past winning bid amounts.

[0146] The above scenario generation unit can calculate the bid price corresponding to each scenario as shown in the above mathematical formula 1.

[0147] Here, the weight (k) is a value corresponding to a standard deviation-based fluctuation range, and can be set to, for example, any one of a range of multiples of the standard deviation centered on the predicted winning bid price ±1σ, ±2σ, or ±3σ.

[0148] Assuming a normal distribution, the mean ±1σ range contains about 68.3% of the total data, the mean ±2σ range contains about 95.4%, and the mean ±3σ range contains about 99.7% of the data, so a statistical range of variation around the predicted winning bid price can be set using multiples of the standard deviation.

[0149] For example, in an aggressive scenario, k can be set in a negative range to set a bid price lower than the predicted winning bid price, in an balanced scenario, k can be set in a range of 0 or close to 0 to set a bid price similar to the predicted winning bid price, and in a conservative scenario, k can be set in a positive range to increase the probability of winning the bid.

[0150] In one embodiment, the weight k (standard deviation-based variation range) may be set to a range of -1.5 to -0.5 for an aggressive scenario, a range of -0.5 to +0.5 for an balanced scenario, and a range of +0.5 to +1.5 for a conservative scenario.

[0151] However, the value of the above weight k is not limited to a fixed value and may be adjusted according to the distribution characteristics of past winning bid amounts, competition intensity, winning patterns by ordering agency, the expected number of participating companies, or the user's risk preference.

[0152] In addition, the above scenario generation unit (40) can generate conservative, balanced, and aggressive bidding scenarios based on competition intensity.

[0153] As described above, the above competition intensity is a variable for quantitatively representing the level of competition in a specific bid, and can be calculated by applying a log transformation to the ratio of the number of participating companies to the average number of participating companies by industry.

[0154] Competition Intensity (C) = log(Number of participating firms / Average number of participating firms by industry)

[0155] Competition intensity increases as the number of participating companies in a specific bid exceeds the industry average, and conversely, decreases as the number of participating companies falls below the average.

[0156] For example, when competition intensity is high, it implies a situation of fierce bidding competition, so a conservative scenario may be applied to secure a relatively stable probability of winning the bid; when competition intensity is at an average level, a balanced scenario that considers both the probability of winning the bid and the expected profit margin may be applied; and when competition intensity is low, it is judged as a situation of relatively low competitive pressure, so an aggressive scenario may be applied to increase the expected profit margin.

[0157] The above scenario generation unit can set a weight k for calculating the bid price according to the intensity of competition, as shown in Table 3.

[0158] Competition intensity C Bidding Scenario Weight k example C ≥ 0.5 conservative type +0.5 ~ +1.2 -0.5 < C < 0.5 Balanced type -0.3 ~ +0.3 C ≤ -0.5 offensive type -1.2 ~ -0.5

[0159] The above scenario generation unit can calculate the bid price by inputting a weight k based on competition intensity into the above mathematical formula 1.

[0160] The above scenario generation unit (40) can calculate at least one of the probability of winning the bid and the expected profit margin for a bid price corresponding to each bidding scenario and provide it to the user. Here, the probability of winning the bid refers to the possibility of winning the bid in the corresponding bid when bidding at a specific bid price.

[0161] The probability of winning the bid can be calculated using various publicly available methods.

[0162] For example, the scenario generation unit can estimate the probability density function of the winning bid amount using a set of winning bid amounts collected from past bidding data. The probability distribution can be generated using a normal distribution, a log-normal distribution, or kernel density estimation.

[0163] For example, assuming a normal distribution, the probability density function of the winning bid amount can be expressed as shown in Equation 3 below.

[0164] [Mathematical Formula 2]

[0165]

[0166] Here

[0167] x is the winning bid amount

[0168] μ is the average of past winning bid amounts

[0169] σ is the standard deviation of past winning bid amounts.

[0170] The above scenario generation unit is a specific bid price T bid Given , the probability of a winning bid occurring at or below the corresponding bid price can be calculated using the cumulative distribution function.

[0171] For example, the probability of winning the bid, Pwin, can be calculated as shown in Equation 3 below.

[0172] [Mathematical Formula 3]

[0173]

[0174] Here, P win The probability of winning the bid, T bid is the bid price, X is a random variable representing the winning bid amount, and F(x) is the cumulative distribution function (CDF) of the winning bid amount distribution.

[0175] The above cumulative distribution function F(x) is a function representing the probability that the winning bid amount will be less than or equal to a specific value x, where F(T bid ) is the winning bid amount is the above bid price T bid It means the probability of becoming less than or equal to.

[0176] Therefore, the above mathematical formula 3 means that the probability of the winning bid amount being greater than or equal to the above bid price is calculated as the winning bid probability.

[0177] In addition, the above scenario generation unit (40) may also calculate the probability of winning the bid using Monte Carlo simulation.

[0178] For example, the above scenario generation unit generates multiple virtual winning bid amounts based on the statistical distribution of past winning bid amounts, and in each simulation, a specific bid price T bid The decision on whether to award the bid can be made by determining whether the generated winning bid amount is less than or equal to the amount.

[0179] In this way, when N simulations are performed, the probability of winning the bid can be calculated as shown in Equation 4 below.

[0180] [Mathematical Formula 4]

[0181]

[0182] N is the number of simulations, Xi is the winning bid amount generated in the i-th simulation, and I(x) is an indicator function that returns 1 if the condition is satisfied and 0 otherwise, and I(Tbid≤Xi) is a condition function that determines whether the proposed bid price is less than the winning bid amount generated in the i-th simulation (i.e., whether it is a winning bid).

[0183] In addition, the above scenario generation unit can calculate the expected profit margin using the bid price and winning probability corresponding to each scenario.

[0184] For example, the expected profit margin can be calculated as follows using the expected profit upon winning the bid and the probability of winning.

[0185] Expected Profit Margin = Probability of Winning the Bid × (Bid Price - Cost) / Cost

[0186] Through this, the present invention can provide the user with the economic expected value for each bidding scenario.

[0187] In addition, the system may include a contribution analysis unit (50) for analyzing the influence of each feature variable on the prediction result for the predicted winning bid price.

[0188] The above contribution analysis unit (50) can calculate the contribution of each feature variable to the predicted winning bid price using an explainable artificial intelligence (XAI)-based analysis method to provide explainability for the prediction results of the artificial intelligence model.

[0189] Analysis methods for calculating the contribution of the aforementioned feature variables include feature importance-based analysis, SHAP (SHapley Additive exPlanations)-based analysis, LIME (Local Interpretable Model-agnostic Explanations)-based analysis, permutation importance-based analysis, or Partial Dependence Plot (PDP)-based analysis.

[0190] For example, the SHAP-based analysis method is based on cooperative game theory and quantitatively decomposes and calculates the extent to which each feature variable contributes to the model's prediction results, offering the advantage of reliably interpreting the contribution of each feature variable.

[0191] For example, in the SHAP analysis method, the output value of the prediction model can be expressed as the sum of the contributions of each feature variable as shown in Equation 5 below.

[0192] [Mathematical Formula 5]

[0193]

[0194] Here, f(x) is the winning bid price predicted by the artificial intelligence model, Φ0 is the average predicted value of the entire data, Φi is the SHAP contribution of the i-th feature variable, and n is the number of feature variables.

[0195] In other words, the SHAP value of each feature variable indicates the extent to which that variable increases or decreases the predicted winning bid price, allowing users to intuitively identify which factors influenced the bid price prediction for a specific bid.

[0196] The above contribution analysis unit may provide contribution values, influence directions, and explanatory information for each feature variable in the form of visualization graphs, importance rankings, or variable influence charts in the user interface.

[0197] In another aspect, the present invention relates to a multi-scenario-based bid price prediction method. Referring to FIG. 2, the multi-scenario-based bid price prediction method includes a data collection step, a feature variable extraction step, a learning step, a winning bid price prediction step, and a bid price generation step.

[0198] The above data collection step is a step of collecting multiple past bidding data.

[0199] The above past bidding data may include past bidding announcement information, base price, estimated price, winning bid amount, number of participating companies, information on the ordering agency, information on construction type or industry, bidding timing information, and other bidding-related statistical data, and the above data may be collected from a public bidding system, a procurement database, or a bidding information providing server.

[0200] The above feature variable extraction step is a step of extracting multiple feature variables from the collected past bidding data to be used for training an artificial intelligence model and predicting the winning bid price.

[0201] For example, the above feature variables may include one or more of the following: base price, bid success rate relative to the estimated price, number of participating companies, average number of participating companies by industry, characteristics of the ordering agency, construction type information, seasonality information, competition intensity, past winning patterns, and other variables representing the bidding environment.

[0202] In addition, the above feature variables can be generated through preprocessing processes such as data normalization, log transformation, categorical variable encoding, or the generation of statistically derived variables.

[0203] The above learning step is a step of training an artificial intelligence model for predicting the winning bid price using the above feature variables and past winning bid amount data.

[0204] The above artificial intelligence model can be implemented using one or more of a regression-based model, an ensemble model, or a neural network model, for example, Random Forest, Gradient Boosting, XGBoost, LightGBM, or a deep learning-based regression model may be used.

[0205] In the above learning step, the prediction accuracy of the model is evaluated using training data and validation data, and an optimal prediction model can be generated through the hyperparameter optimization process.

[0206] The above winning bid price prediction step is a step of calculating the predicted winning bid price for a specific bid using the above-mentioned trained artificial intelligence model.

[0207] Specifically, by inputting feature variables for new bids into an artificial intelligence model, it is possible to predict the expected winning bid amount or bid threshold value for the bid.

[0208] The above bid price generation step is a step of generating bid price scenarios corresponding to multiple bidding strategies using the above predicted winning bid price as a reference value.

[0209] For example, in the bid price generation step described above, conservative, balanced, and aggressive bidding scenarios can be generated using one or more of the standard deviation, percentile, competition intensity, or winning patterns by contracting authority of past winning bid amounts, and the bid price corresponding to each scenario can be calculated by applying the standard deviation and weights to the predicted winning bid price.

[0210] In addition, in the bid price generation step, at least one of the probability of winning the bid and the expected profit margin can be calculated for the bid price corresponding to each scenario and provided to the user.

[0211] The above multi-scenario-based bid price prediction method can be implemented by referring to the configuration and operation of the multi-scenario-based bid price prediction system described above.

[0213] The present invention will be described in more detail below through examples. These examples are solely for the purpose of further explaining the present invention, and it is obvious to those skilled in the art that the scope of the present invention is not limited by these examples according to the gist of the invention.

[0214] [Example]

[0215] First, past bidding data was collected from the Korea Public Procurement Service (g2b.go.kr), the Public Procurement Service bidding database, and the Public Bidding Information System (see Table 1 above). The above past bidding data includes the announcement number, base price, estimated price, winning bid amount, ordering agency, number of participating companies, bidding period, and information on construction type or industry.

[0216] The collected data was used as training data after performing data preprocessing steps such as removing missing values, removing outliers, and normalization. Subsequently, feature variables for training an artificial intelligence model were extracted from the data (see Table 2 above).

[0217] Some of the above feature variables were used as input variables for the artificial intelligence model after undergoing preprocessing steps such as log transformation, normalization, or categorical encoding.

[0218] Next, a winning bid prediction model was trained using two artificial intelligence models. The models used were the Random Forest model and the XGBoost model, and each model was trained with hyperparameter conditions as shown in Table 4 below.

[0219] model Random Forest XGBoost parameters n_estimators=100, max_depth=15, min_samples_split=50, random_state=42 learning_rate=0.05, max_depth=6, n_estimators=200, subsample=0.8, colsample_bytree=0.8

[0220] The two models mentioned above were trained using historical bidding data and feature variables, and were configured to predict the winning bid price independently.

[0221] Next, the winning bid price was predicted by inputting feature variables for the new bid announcement. The base price of the new announcement was 492,000,000 won, and the data was input into two trained artificial intelligence models to predict the winning bid price.

[0222] To combine the prediction results of each model, a weighted average was calculated using the following formula.

[0223] Final Predicted Winning Bid = 0.4 × RF_Prediction + 0.6 × XGBoost_Prediction

[0224] At this time, the final predicted winning bid was calculated to be 468,000,000 won.

[0225] In addition, the standard deviation of the winning bid amount was calculated from past bidding data of the same industry and similar types of work, and the value was calculated to be 12,000,000 won.

[0226] Next, three bidding scenarios were generated based on the predicted winning bid price. Each scenario was calculated using Equation 1, which was explained earlier.

[0227] In mathematical formula 1, the weights were set to position values ​​corresponding to the percentiles of the past winning bid amount distribution, and 0.67 was applied for the conservative scenario, 0 for the balanced scenario, and -1.08 for the aggressive scenario.

[0228] The bid price calculated by applying these weights, along with the corresponding probability of winning and expected profit margin, is shown in Table 5 below.

[0229] scenario Bid price (T) Winning bid probability Expected profit margin explanation conservative type 476,000,000 won 82% 4.2% High probability of winning the bid, low profit Balanced type 468,000,000 won 65% 7.5% Medium probability, medium profit offensive type 455,000,000 won 45% 11.8% Low probability, high profit

[0230] The aforementioned winning bid probability was calculated by performing Monte Carlo simulations based on the distribution of past winning bid amounts. For example, multiple virtual winning bid amounts were generated based on the past distribution, and the winning bid status was determined by assessing whether the amount fell below a specific bid price in each simulation; the winning bid probability was calculated as the ratio of winning bids to the total number of simulations.

[0231] In addition, the expected profit margin for each scenario was calculated using the probability of winning the bid and the expected profit upon winning.

[0232] Next, SHAP analysis was performed to ensure the explainability of the AI ​​model's prediction results. Specifically, the SHAP TreeExplainer algorithm was used to input the test data (X_test) and calculate the SHAP values ​​for each feature variable.

[0233] For the balanced scenario (468,000,000 won), the model's average predicted value (Base Value) was 470,000,000 won, and the contribution of each feature variable was calculated as shown in Table 6 below.

[0234] Feature variables SHAP value Direction of influence explanation base amount +5,200,000 rising Bid prices rise as base amounts are higher than average Ordering agency (OO City Hall) -3,800,000 degradation OO City Hall prefers low successful bid rates Number of participating companies (7) -1,500,000 degradation High competition intensity leads to lower bid prices Seasonality (December) -1,200,000 degradation Low prices preferred due to year-end budget exhaustion etc -700,000 degradation The sum of the remaining variables

[0235] Through the SHAP analysis results above, the impact of each feature variable on the prediction of the winning bid price can be quantitatively verified, and users can formulate bidding strategies more rationally based on this.

[0237] [Example 2]

[0238] For 500 actual new bids during the period from October to December 2024 (3 months), the winning bid price (lowest valid bid price) was predicted by the present invention, Company A (prediction based on the probability distribution of the estimated price), Company B (prediction based on the bid evaluation rate of other companies using preliminary price data), and human experts. After the winning bid results for the case were disclosed, the difference between the actual winning bid amount and the predicted winning bid price was calculated, and the mean error (MAE) between them was calculated and shown in Table 7 below.

[0239] method MAE (Mean Error) Scenario provided explainability User satisfaction The present invention 8.5% 3 types SHAP 4.6 / 5.0 Company A 12.3% single score doesn't exist 3.2 / 5.0 Company B 11.8% Single prediction doesn't exist 3.4 / 5.0 human experts 10.2% △ Manual presentation △ Oral explanation 3.9 / 5.0

[0240] Referring to Table 6, it can be seen that the multi-scenario-based bid price prediction method according to the present invention has significantly improved winning bid price prediction accuracy compared to the existing prediction methods of Company A and Company B.

[0241] Specifically, the mean error (MAE) of the present invention was found to be 8.5%, which is about 3 to 4 percentage points lower than Company A's method (12.3%) and Company B's method (11.8%), resulting in a significant improvement in the accuracy of winning bid price prediction. Furthermore, even when compared to the prediction method of human experts (10.2%), the method of the present invention showed a lower mean error, confirming the effectiveness of the data-based artificial intelligence prediction method.

[0242] In addition, while most existing systems provide only a single winning bid prediction value, the present invention can present various bidding strategies to users by simultaneously providing three bidding scenarios: conservative, balanced, and aggressive. This method of providing multiple scenarios has the advantage of allowing bidders to select a bidding strategy based on their risk preference and profit target.

[0243] Furthermore, the present invention utilizes an explainable AI analysis method to provide the user with the influence of each feature variable on the winning bid prediction result, thereby presenting the basis for the prediction result. Accordingly, the user can understand the prediction result of the AI ​​model more intuitively and improve confidence in establishing bidding strategies.

[0244] For this reason, the system of the present invention provides superior performance in terms of prediction accuracy, decision support, and explainability compared to existing single prediction methods, and also showed a high satisfaction rating of 4.6 / 5.0 in actual user evaluations.

[0246] Although preferred embodiments of the present invention have been described in detail above, such description is merely for describing and disclosing exemplary embodiments of the invention. Those skilled in the art will readily recognize that various changes, modifications, and variations are possible from the above description and the accompanying drawings without departing from the scope and essence of the invention.

Claims

Claim 1 A data collection unit (10) for collecting multiple past bidding data; a feature extraction unit (20) for extracting feature variables including price-related variables, competition-related variables, and ordering agency-related variables, and the winning bid amount of the corresponding past bid from the collected past bidding data; a machine learning unit (30) having a learning unit (31) for learning one or more machine learning models using the feature variables as input data and the winning bid amount as a target value, and a prediction unit (32) for deriving the winning bid price through the machine learning model learned in the learning unit (31) using input data including the base amount and ordering agency information included in the new bidding announcement; The system includes a scenario generation unit (40) that generates a bid price corresponding to one or more bidding scenarios based on the predicted winning bid price, wherein the amount-related variables include a base amount and a ratio of the estimated price to the base amount, the competition-related variables include the number of participating companies, the average number of participating companies by industry, and the ratio of the average number of participating companies by industry to the number of participating companies, and the ordering agency-related variables include the past winning bid rate of the ordering agency, and the scenario generation unit (40) calculates the standard deviation of the winning bid amount using a plurality of past winning bid amounts extracted from past bidding data, and provides one or more bidding scenarios among conservative, balanced, and aggressive types by considering the probability of winning and the expected profit margin, and the predicted winning bid price (T pred A multi-scenario-based bid price prediction system characterized by generating a bid price (T) corresponding to the bid scenario by adding or subtracting the value obtained by multiplying the standard deviation by a weight (k) according to the following mathematical formula 1, using ) as a reference value. [Mathematical Formula 1] T=T pred + kσ where T is the bid price, T pred is the predicted winning bid price, σ is the standard deviation of past winning bid amounts, and k is the weight. Claim 2 delete Claim 3 A multi-scenario-based bid price prediction system according to claim 1, wherein the machine learning unit (30) predicts the winning bid price using a plurality of machine learning models including Random Forest and XGBoost, and calculates the final winning bid price by weighted averaging the prediction results of the plurality of machine learning models. Claim 4 A multi-scenario-based bid price prediction system according to claim 1, wherein the machine learning unit (30) predicts an initial winning bid price using a Random Forest, and the XGBoost model learns the residual corresponding to the difference between the initial winning bid price predicted by the Random Forest and the actual winning bid price to calculate a correction value, and the final winning bid price prediction value is calculated by combining the initial winning bid price and the correction value. Claim 5 A multi-scenario-based bid price prediction system characterized in that, in the first paragraph, the variable related to the ordering agency includes an ordering agency ID, and the machine learning unit (30) learns an independent sub-model for each ordering agency using past bid data grouped based on the ordering agency ID, and when the number of training data for each ordering agency is greater than or equal to a preset threshold, the machine learning unit (30) selects a sub-model corresponding to the ordering agency ID for a new bid to predict the winning bid price. Claim 6 delete Claim 7 In claim 1, the scenario generation unit (40) is characterized by generating the conservative, balanced, and aggressive bidding scenarios based on one or more of the following: a position corresponding to a percentile in the distribution of past winning bid amounts, a range of variation based on the standard deviation, competition intensity, or winning bid patterns by ordering agency. Claim 8 delete Claim 9 A multi-scenario-based bid price prediction system according to claim 1, wherein the scenario generation unit (40) calculates at least one of a winning bid probability and an expected profit margin for a bid price corresponding to each bid scenario and provides the same to the user. Claim 10 A multi-scenario-based bid price prediction system according to claim 1, wherein the system includes a contribution analysis unit (50) that calculates the contribution of each feature variable to the prediction result using an explainable AI analysis method that calculates the contribution of each feature variable to the prediction result for the predicted winning bid price, and the contribution analysis unit (50) provides the user with the influence of each feature variable on the prediction result based on the contribution of each feature variable. Claim 11 A multi-scenario-based bid price prediction method performed by a bid price prediction system comprising a data collection unit (10), a feature extraction unit (20), a machine learning unit (30), a scenario generation unit (40), and a control unit (60), wherein the data collection unit collects a plurality of past bid data; the feature extraction unit extracts from the collected past bid data a feature variable including a price-related variable, a competition-related variable, and a contracting authority-related variable, and the winning bid amount of the corresponding past bid; the machine learning unit learns one or more machine learning models using the feature variable as input data and the winning bid amount as a target value, and predicts the winning bid price through the learned machine learning model using input data including a base price and contracting authority information included in a new bid announcement; the scenario generation unit generates a bid price corresponding to one or more bid scenarios based on the predicted winning bid price, wherein the price-related variable includes a base price and a ratio of the estimated price to the base price, and the competition-related variable includes the number of participating companies, the average number of participating companies by industry, and the ratio of the average number of participating companies by industry to the number of participating companies, and Variables related to the ordering agency include the ordering agency's past winning bid rate, and the scenario generation unit (40) calculates the standard deviation of the winning bid amount using a plurality of past winning bid amounts extracted from past bidding data, and provides one or more bidding scenarios among conservative, balanced, and aggressive types by considering the probability of winning and the expected profit margin, and the predicted winning bid price (T pred A multi-scenario-based bid price prediction method characterized by generating a bid price (T) corresponding to the bid scenario by adding or subtracting the value obtained by multiplying the standard deviation by a weight (k) according to the following mathematical formula 1, using ) as a reference value. [Mathematical Formula 1] T=T pred + kσ where T is the bid price, T pred is the predicted winning bid price, σ is the standard deviation of past winning bid amounts, and k is the weight.

Images