Intelligent enterprise cash management method and system

Abstract

The invention provides an enterprise cash intelligent management method and system, and belongs to the technical field of financial science and technology and artificial intelligence. The method comprises the steps that historical cash flow data of an enterprise is acquired, a multi-dimensional time feature space of the data is constructed, and the feature space at least comprises a first class of time features based on a preset calendar period and a second class of time features based on a specific business period of the enterprise; inputting the multi-dimensional time feature space into a prediction model to obtain a cash flow prediction result of a future target time period; dynamically calibrating a lowest control line parameter and a cash flow volatility parameter in the cash management model based on the Miller-Oyer principle based on the cash flow prediction result and the historical cash flow data; and generating a decision basis of cash management according to the dynamically calibrated cash management model. According to the method and the device, the problems of static stiffness and incapability of adaptive adjustment of cash management model parameters in the prior art are solved, the dynamic adaptation of the management strategy is realized, and the real-time performance and the accuracy of decision making are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial technology and artificial intelligence, in particular to a method and system for enterprise cash flow prediction and dynamic optimization management based on data driving. BACKGROUND

[0002] Enterprise cash management is a core financial link to maintain the continuity and stability of its business activities. Adequate cash flow can guarantee the daily operation of the enterprise, repay debts and make investments, while effective cash management aims to maintain an appropriate cash level at the lowest cost, balancing liquidity and profitability.

[0003] When managing cash holdings, enterprises usually use classic financial models such as the Miller-Or model. These models set upper and lower limits and return lines for cash holdings to guide the conversion between cash and valuable securities. However, the key control parameters of these models, such as the lower limit of cash holdings, cash flow volatility, and transaction costs, are usually set as fixed values or adjusted manually on a regular basis. This static parameter mechanism makes the cash management strategy rigid and unable to adapt to real-time changes in market conditions and future funding needs of the enterprise, making it difficult to achieve the best balance between ensuring liquidity safety and improving capital utilization efficiency. SUMMARY

[0004] Therefore, embodiments of the present application aim to provide an enterprise cash intelligent management method and system to solve the technical problems of insufficient quantification of multi-period time characteristics in the existing cash flow prediction model, low prediction accuracy, and static rigidity of cash management model parameters that cannot be adjusted adaptively, thereby realizing the automation, dynamization and optimization of cash management strategy.

[0005] To achieve the above-mentioned purpose, the present application provides an enterprise cash intelligent management method, comprising the following steps: obtaining historical cash flow data of an enterprise, and constructing a multi-dimensional time feature space of the cash flow data of the enterprise, the feature space comprising at least a first type of time feature based on a preset calendar period and a second type of time feature based on a specific business period of the enterprise; inputting the multi-dimensional time feature space into a prediction model to obtain a cash flow prediction result of a future target period; wherein the prediction model is a deep learning model; based on the cash flow prediction result and the historical cash flow data, dynamically calibrating a minimum control line parameter and a cash flow volatility parameter in a cash management model; wherein the cash management model is a financial model based on the Miller-Or principle; generating a decision basis for cash management according to the cash management model after dynamic calibration.

[0006] Optionally, the first type of time features comprises at least one of a basic timing layer feature, a periodic fluctuation layer feature, and a social cycle layer feature.

[0007] Optionally, the basic timing layer feature comprises year, month, day, week sequence number, and generated periodic trigonometric function decomposed by date. The periodic fluctuation layer feature comprises time point labels for identifying weekends, month ends, or quarter ends, and continuous working day count. The social cycle layer feature comprises statutory holiday density and effective working day proportion.

[0008] Optionally, the second type of time features is a business cycle layer feature, comprising fund settlement key days marked according to industry characteristics of the enterprise, and / or business peak period labels marked according to historical business data.

[0009] Optionally, the deep learning model is an improved random forest model. The improvement of the improved random forest model comprises: The maximum depth of the decision tree in the improved random forest model is adjustable. The improved random forest model reassigns weights of the multi-dimensional time features. The loss function of the improved random forest model adds a penalty term for fund prediction gap on the basis of standard mean square error.

[0010] Optionally, the improved random forest model comprises a dynamic tree depth control strategy, wherein the maximum depth of the decision tree is positively correlated with the number of training samples.

[0011] Optionally, the reassignment of weights of the multi-dimensional time features comprises: Based on a sliding window entropy value algorithm, dynamic weights are assigned to features in the feature space, wherein the higher the entropy value of a feature, the greater the assigned weight.

[0012] Optionally, the dynamic calibration of the minimum control line parameter and the cash flow volatility rate parameter in the cash management model comprises: Using a rolling window method, the maximum single-day net outflow in a preset historical time window is identified as the minimum control line of the current period. A first volatility rate of historical cash flow data is calculated; a second volatility rate of the cash flow prediction result is calculated; the first volatility rate and the second volatility rate are weighted and fused to obtain a hybrid volatility rate, which is taken as the cash flow volatility rate parameter.

[0013] Optionally, it further comprises: Obtaining the market value of the valuable securities; adjusting, in real time, a market value fluctuation rate of the marketable securities in the cash management model based on the market value of the marketable securities.

[0014] The application also provides an intelligent cash management system for an enterprise, comprising: an acquisition module, configured to acquire historical cash flow data of the enterprise, and construct a multi-dimensional time feature space of the cash flow data of the enterprise, wherein the feature space comprises at least a first type of time feature based on a preset calendar cycle and a second type of time feature based on a specific business cycle of the enterprise; a prediction module, configured to input the multi-dimensional time feature space into a prediction model to obtain a cash flow prediction result of a future target period, wherein the prediction model is a deep learning model; a calibration module, configured to dynamically calibrate a minimum control line parameter and a cash flow fluctuation rate parameter in a cash management model based on the cash flow prediction result and the historical cash flow data, wherein the cash management model is a financial model based on the Miller-Or principle; a generation module, configured to generate a decision basis for cash management according to the cash management model after dynamic calibration.

[0015] Compared with the prior art, the technical scheme provided by the application has the following beneficial effects: 1. The prediction accuracy and scene adaptability are improved. By constructing a multi-dimensional time feature space, various periodic factors affecting cash flow are quantified, and the prediction error at special time points such as holidays and industry settlement days is significantly reduced.

[0016] 2. Dynamic self-adaptation of management strategies is achieved. The application deeply couples the high-precision prediction result with the classic Miller-Or model, and dynamically calculates the core parameters such as the minimum control line and the fluctuation rate in real time, so that the cash management strategy can automatically adapt to market changes and future fund demand, breaking through the limitations of the static parameters of traditional models. The dynamically adjusted cash holding threshold enables the enterprise to ensure sufficient liquidity to cope with risks while avoiding holding too much idle cash, thereby improving the overall fund utilization efficiency and achieving a better balance between risk control and profitability. BRIEF DESCRIPTION OF DRAWINGS The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application, taken in conjunction with the accompanying drawings. The drawings provided in the specification and the embodiments of the present application together serve to explain the present application and, therefore, to provide further understanding of the present application, and form a part of the specification. The drawings do not limit the present application, but serve to explain the present application together with the embodiments of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1is a flowchart of an enterprise cash intelligent management method provided by an embodiment of the present application.

[0018] Figure 2 is a structural diagram of an enterprise cash intelligent management system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0020] Figure 1 is a whole flowchart of a cash management method provided by an embodiment of the present application. It can be understood that the method can be executed by one or more computer devices (for example, servers or cloud servers deployed in an enterprise). The computer device can include hardware components such as a processor, a memory, a communication interface and a bus. The memory stores computer program instructions executable by the processor, and the processor executes the instructions to implement the method steps described in the embodiments of the present application.

[0021] Specifically, the method includes the following steps: Step S110: obtaining historical cash flow data of an enterprise, and constructing a multi-dimensional time feature space of the cash flow data of the enterprise, wherein the feature space at least includes a first type of time feature based on a preset calendar period and a second type of time feature based on an enterprise-specific business period; In an application scenario for a large retail enterprise, the system accesses the databases of the enterprise resource planning system, the customer relationship management system and the financial management software of the enterprise through a data interface module 211 (see Figure 2 ). The data interface module 211 can be configured with a standard database connection driver, for example, through open database connection or Java database connection technology, to automatically extract the total amount of cash inflow, the total amount of cash outflow, the account balance and other core financial data every day in the past three years at a preset frequency (such as every morning). The collected data format can be structured table data, wherein each row represents a date, and the columns include date, inflow amount, outflow amount and other fields. After data collection, data preprocessing can also be performed, for example, forward or backward filling method is used to complete the missing data, and 3-sigma principle is used to identify and smooth the extreme abnormal values caused by accounting errors, so as to ensure the data quality.

[0022] The feature engineering module 212 receives the raw time series data from the data interface module 211 and performs a four-layer temporal feature decomposition, thereby constructing a multi-dimensional temporal feature space that can fully represent the cash flow fluctuation pattern. The feature space includes at least a first type of temporal feature based on a preset calendar period and a second type of temporal feature based on a business-specific cycle.

[0023] As a specific implementation, the first type of temporal feature can be further divided into the following three levels: Basic time series layer features: Basic decomposition of the timestamp. For each date, the feature engineering module 212 extracts its corresponding year, month, quarter, week of the year, day of the week, etc. In addition, to help the model understand the periodicity, periodic trigonometric function features can also be generated. For example, for the month feature, the sine and cosine transformations can be calculated: and This transformation maps the linear month (1 to 12) to a unit circle, so that the distance between December and January in the feature space is closer than in the original numerical space, thereby better representing the annual periodicity. Similarly, similar trigonometric function transformations can also be performed on the week, quarter, etc.

[0024] Periodic fluctuation layer features: used to identify time points with special significance in the standard calendar. The feature engineering module 212 generates a Boolean label for each date according to the calendar rules, such as "whether it is a weekend", "whether it is the end of the month", "whether it is the end of the quarter", "whether it is the end of the year". These time points are usually accompanied by concentrated settlement or turnover of funds, which are important factors affecting cash flow. In addition, features such as "number of consecutive working days" or "number of days to the next holiday" can also be calculated to quantify the work rhythm and fund preparation behavior before and after the holiday.

[0025] Social cycle layer features: mainly used to quantify the impact of social activities (especially statutory holidays). The feature engineering module 212 has a built-in statutory holiday library to determine whether each date is a holiday. Further, more macro statistical features can be calculated, such as "monthly statutory holiday density", whose calculation formula is: In addition, features such as "monthly effective working day proportion" can also be calculated. These features help the model capture changes in consumer behavior or changes in cash flow patterns caused by the suspension of business activities due to holidays, such as the significant fluctuations in cash flow during the Spring Festival and National Day holidays.

[0026] The second type of time feature, i.e., the business cycle layer feature, can be customized according to the business characteristics of the enterprise. For the retail enterprise in the present embodiment, the feature engineering module 212 can mark out industry-specific business peak period labels, such as "Double 11", "618" and other large e-commerce promotion days and a period of time before and after the promotion days (such as a week of preparation before the event and a week of payment after the event). These labels can be Boolean values or numerical values representing the intensity of the event. In addition, according to historical data analysis, the sales peak period within the enterprise's monthly sales can be marked, for example, for some offline retailers, the weekend or the days after the pay day may be the sales peak. These features closely related to the specific business activity cycle of the enterprise help to improve the accuracy of the prediction model.

[0027] After step S110 is completed, the generated feature vector containing multi-dimensional time features is passed to the subsequent steps.

[0028] Step S120: inputting the multi-dimensional time feature space into a prediction model to obtain a cash flow prediction result of a future target period; wherein the prediction model is a deep learning model; The purpose of this step is to use the multi-dimensional time feature space constructed in step S110, which is rich in business semantics, to generate a high-precision and forward-looking cash flow prediction result through an advanced prediction model.

[0029] This step realizes accurate and robust prediction of future cash flow by inputting the multi-dimensional time feature space rich in information into an improved random forest model with three core improvements, i.e., "dynamic structure", "adaptive attention" and "risk aversion preference". The prediction result is no longer a simple statistical extrapolation, but an intelligent inference deeply integrating time rules and business knowledge, providing high-quality and reliable input basis for the dynamic parameter calibration of the next step.

[0030] Step S130: dynamically calibrating the minimum control line parameter and the cash flow volatility parameter in the cash management model based on the cash flow prediction result and the historical cash flow data; wherein the cash management model is a financial model based on the Miller-Orr principle; This step constitutes the core optimization link of the intelligent decision-making of the system, and its main goal is to deeply integrate the prediction data with the traditional financial model. Through the dynamic calibration mechanism, the classic Miller-Orr model is transformed from a static calculation tool into an intelligent decision-making system that can adapt to market changes and actual needs of the enterprise.

[0031] Step S140: generating the decision basis for cash management according to the cash management model after dynamic calibration.

[0032] The step is based on the dynamically calibrated cash management model, and outputs the operable fund management strategy, and completes the final conversion from data prediction to decision execution. Through the organic combination of S130 and S140, the leap from traditional static cash management to dynamic intelligent decision is realized. The system can not only adapt to the business cycle changes of the enterprise itself, but also respond to the fluctuations of the financial market in real time, so that the cash management strategy has business compatibility and market sensitivity. This dynamic calibration mechanism based on multi-dimensional prediction effectively solves the industry pain points such as parameter rigidity and response lag of traditional financial models, and provides a more accurate and agile fund management solution for modern enterprises.

[0033] Further, the deep learning model is an improved random forest model; The improvement of the improved random forest model includes: The maximum depth of the decision tree in the improved random forest model is adjustable; The improved random forest model reassigns the weights of the multi-dimensional time features; The loss function of the improved random forest model adds a penalty term for the fund prediction gap on the basis of the standard mean square error.

[0034] The improved random forest model includes a dynamic tree depth control strategy, in which the maximum depth of the decision tree is positively correlated with the number of training samples.

[0035] The reassignment of the weights of the multi-dimensional time features includes: Based on the sliding window entropy algorithm, dynamic weights are assigned to the features in the feature space, and the higher the entropy value of the feature, the greater the assigned weight.

[0036] Specifically, the tree depth of the traditional random forest model is usually determined as a fixed value through cross-validation, which cannot adapt to the complexity requirements of the sample number and time period in time series data. The present application introduces a dynamic tree depth control strategy, which establishes a positive correlation function relationship between the maximum depth of the decision tree and the number of training samples: Embodiment: Set the maximum depth calculation function as max_depth = F(N), where N is the number of training samples, and F is a monotonically increasing function; Preferred scheme: Use the form max_depth = log 2 (N) + C, where C is an adjustment term based on the business cycle characteristics.

[0037] In this way, the model can adaptively adjust the complexity according to the data size, effectively control the risk of overfitting while ensuring the expression ability of the model.

[0038] 2. Dynamic feature weight allocation based on sliding window entropy value: To overcome the static limitations of traditional feature importance evaluation methods, the present application proposes a dynamic feature weight allocation algorithm based on sliding window entropy value: Entropy calculation includes: maintaining a fixed-length sliding time window for each feature, and calculating the information entropy value of the feature sequence in the window; Weight allocation: based on the entropy calculation results of each feature, the Softmax function is used for normalization processing to generate a dynamic weight vector; Allocation principle: the higher the entropy value of the feature, the greater the uncertainty of the feature in the recent window, and the weight allocated is correspondingly increased; Technical advantages: enable the model to automatically focus on the feature dimension with significant recent fluctuations, and improve the perception ability of sudden changes in funds.

[0039] 3. Risk-sensitive asymmetric loss function: In view of the asymmetric business characteristics that "the risk of fund gap is higher than that of fund redundancy" in enterprise cash management, the standard loss function is improved: Infrastructure: introduce a fund gap penalty term based on the standard mean square error (MSE); Mathematical expression: L = MSE + λ·Σ(max(0, actual value-predicted value)) 2 ; Parameter description: λ is the penalty coefficient, which is determined by the validation set; max(0, actual value-predicted value) monomial ensures that only the fund gap is penalized.

[0040] Business value: guide the model to prioritize the prevention of fund shortage risk during training, which is highly consistent with the actual risk management needs of enterprises, and has synergistic effect and overall advantage; The above three improvements work together to build an intelligent learning system optimized for cash flow prediction: dynamic tree depth control ensures the rationality of the model structure, feature weight allocation enhances the adaptability of the model to time series changes, and asymmetric loss function ensures the consistency of the model output with the business goal. This multi-level optimization makes the improved random forest model perform excellently in cash flow prediction tasks, providing reliable technical support for subsequent financial decision-making.

[0041] Further, the dynamic calibration of the minimum control line parameter and the cash flow volatility parameter in the cash management model includes: Using the rolling window method, identify the maximum single-day net outflow in the preset historical time window, and use it as the minimum control line for the current period; A first volatility of historical cash flow data is calculated; a second volatility of the cash flow prediction result is calculated; the first volatility and the second volatility are weighted and fused to obtain a mixed volatility, and the mixed volatility is taken as a cash flow volatility parameter.

[0042] Specifically, the minimum control line in the traditional Miller-Or model is usually set based on experience and remains fixed, and cannot adapt to the dynamic changes of enterprise fund demand. Window setting: a historical time window of a preset length (preferably 30 trading days) is set to ensure covering a complete business cycle; Calculation method: in the time window, the difference between the total daily cash outflow and the total daily cash inflow is calculated day by day to identify the maximum value; Parameter determination: the identified maximum single-day net outflow is taken as the minimum control line of the current period, that is, in this way, the minimum cash reserve held by the enterprise can always cover the maximum fund demand that may occur in the near future, effectively preventing payment risk; To overcome the limitation of traditional methods relying only on historical data, the application proposes a mixed volatility estimation method that combines historical rules and future insights: First volatility calculation: based on historical cash flow data, the standard deviation is calculated in the same time window, denoted as std(Xhistory); Second volatility calculation: based on the future cash flow prediction results output by the prediction model, the standard deviation in the prediction period is calculated, denoted as std(Xpredicted); Weighted fusion mechanism: entropy weighted method is used to fuse the two volatilities: δt= αstd(Xhistory) + (1-α)std(Xpredicted); Wherein, α is a weighting coefficient determined based on information entropy theory, and the preferred value is 0.6; Parameter output: the calculated mixed volatility δt is taken as the current cash flow volatility parameter; In this way, the calibration mechanism realizes the following technical breakthroughs through the synergistic optimization of the above two aspects: The minimum control line is dynamically adjusted according to the actual fund demand of the enterprise in the recent period, avoiding the risk of idle or short funds caused by traditional fixed setting: the volatility parameter fuses prediction information, enabling the model to look forward and respond to expected market changes and business fluctuations in advance; through reasonable setting of the weighting coefficient, the best balance between the robustness of historical experience and the sensitivity of future prediction is achieved; This dynamic calibration mechanism effectively solves the technical problem of parameter rigidity in traditional cash management models, making the Miller-Or model regain its practical value in modern complex business environments and providing more accurate and reliable technical support for enterprise fund management.

[0043] In some embodiments, further comprising: obtaining the market value of the valuable securities; and adjusting the market value fluctuation rate of the valuable securities in the cash management model in real time based on the market value of the valuable securities.

[0044] As a further optimization of the dynamic calibration process, the present application introduces a real-time adjustment mechanism for the market value fluctuation rate of valuable securities in some preferred embodiments to address the technical defect of traditional cash management models that ignore the impact of securities market fluctuations on the conversion cost of funds.

[0045] The specific implementation includes: Market value data acquisition: Real-time acquisition of market value information of valuable securities portfolio held by the enterprise through a financial data interface; the types of securities covered include but are not limited to: treasury bonds, money market funds, short-term corporate bonds, and other liquid assets; the data update frequency is synchronized with the cash management decision-making cycle to ensure the timeliness of the information; Market value fluctuation rate calculation: Based on the historical market value data of the securities portfolio, calculate its fluctuation rate index within the preset period; Use the standardized volatility calculation formula to ensure comparability between different types of securities; normalize the volatility rate to map it to a pre-set reasonable interval; Dynamic conversion coefficient generation: Convert the calculated market value fluctuation rate to a dynamic conversion coefficient ξ; Set the effective range of the coefficient to [0.8, 1.2], corresponding to different market conditions: ξ>1.0: indicates that the securities market is in an upward channel and the realization environment is favorable; ξ = 1.0: indicates that the market conditions are stable and the normal cost is estimated; ξ<1.0: indicates that the market is under downward pressure and realization may incur additional costs; Real-time calibration of conversion cost: Establish a functional relationship between the dynamic conversion coefficient ξ and the fixed conversion cost parameter in the traditional cash management model: implementation formula example: bt = Lt-1 × k × ξt / T where k is the basic conversion cost rate, T is the period parameter; bt is the dynamic conversion cost parameter; Lt-1 is the last period's lowest control line; Through this linkage mechanism, the conversion cost is dynamically adjusted according to the market conditions: the supplementary mechanism realizes the following technical breakthroughs by introducing the real-time fluctuation factor of the value of the securities market: risk perception ability is improved: the cash management system can perceive and respond to real-time changes in the financial market, significantly enhancing the environmental adaptability of the system; decision accuracy is optimized: by dynamically adjusting the conversion cost, the conversion decision between cash and securities is more in line with the actual market, avoiding unnecessary asset conversion in unfavorable market conditions; liquidity management is enhanced: considering both cash demand and securities market fluctuations, the efficiency of fund allocation is optimized under the premise of ensuring payment safety. This innovation effectively solves the problem of the applicability of classical financial models in dynamic market environments, providing more accurate and reliable decision support for enterprise cash management in complex financial market environments.

[0046] Reference Figure 2 The application also provides an enterprise cash intelligent management system, which can be used to execute the method embodiments of the application. For details not disclosed in the system embodiments of the application, please refer to the method embodiments of the application.

[0047] The enterprise cash intelligent management system provided by the application comprises: An acquisition module 21 is configured to acquire historical cash flow data of an enterprise, and construct a multi-dimensional time feature space of the cash flow data of the enterprise, wherein the feature space comprises at least a first type of time feature based on a preset calendar period and a second type of time feature based on a specific business period of the enterprise. A prediction module 22 is configured to input the multi-dimensional time feature space into a prediction model to obtain a cash flow prediction result of a future target period, wherein the prediction model is a deep learning model. A calibration module 23 is configured to dynamically calibrate a minimum control line parameter and a cash flow volatility parameter in a cash management model based on the cash flow prediction result and the historical cash flow data, wherein the cash management model is a financial model based on the Miller-Or principle. A generation module 24 is configured to generate a decision basis for cash management according to the cash management model after dynamic calibration.

[0048] In addition to the above method and device, the embodiments of the application can also be a computer program product, which comprises computer program instructions, which, when executed by a processor, cause the processor to execute the steps of the enterprise cash intelligent management method according to various embodiments of the application described in the above "Exemplary Method" section of the specification.

[0049] The computer program product can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the user's computing device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device or entirely on the remote computing device or server. The embodiments of the present application are not limited by the

[0050] In addition, an embodiment of the present application can also be a computer readable storage medium, having stored thereon computer program instructions which, when executed by a processor, cause the processor to perform the steps described in the above "Exemplary Method" section of the present specification for the enterprise cash intelligent management method according to various embodiments of the present application.

[0051] The computer readable storage medium can be any combination of one or more computer readable media. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0052] The above description is given for illustrative and descriptive purposes. Furthermore, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although various example aspects and embodiments have been discussed above, those of ordinary skill in the art will appreciate certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for intelligent cash management in enterprises, characterized in that, Includes the following steps: Acquire historical cash flow data of the enterprise, and construct a multi-dimensional time feature space of the enterprise's cash flow data. The feature space includes at least a first type of time feature based on a preset calendar period and a second type of time feature based on a specific business cycle of the enterprise. The multi-dimensional time feature space is input into the prediction model to obtain the cash flow prediction results for the future target period; wherein, the prediction model is a deep learning model; Based on the aforementioned cash flow forecast results and historical cash flow data, the minimum control line parameter and cash flow volatility parameter in the cash management model are dynamically calibrated; wherein, the cash management model is a financial model based on the Miller-Olé principle; Based on the dynamically calibrated cash management model, a decision-making basis for cash management is generated.

2. The method according to claim 1, characterized in that, The first type of time feature includes at least one of the following: basic time series layer features, periodic fluctuation layer features, and social cycle layer features.

3. The method according to claim 2, characterized in that, The basic time-series layer features include: year, month, day, and week ordinal numbers decomposed from the date, and the generated periodic trigonometric functions; The cyclical fluctuation layer features include: time point labels for identifying weekends, end of months or quarters, and consecutive working day counts; The characteristics of the social cycle layer include: the density of statutory holidays and the percentage of effective working days.

4. The method according to claim 1, characterized in that, The second type of time feature is a business cycle layer feature, including: key days for fund settlement marked according to the characteristics of the industry to which the enterprise belongs, and / or business peak period labels marked according to historical business data.

5. The method according to claim 1, characterized in that, The deep learning model is an improved random forest model; The improvements to the improved random forest model include: In the improved random forest model, the maximum depth of the decision tree is adjustable; The weights of the multi-dimensional temporal features are redistributed in the improved random forest model; The improved random forest model's loss function adds a penalty term for funding prediction gaps to the standard mean squared error.

6. The method according to claim 5, characterized in that, The improved random forest model includes a dynamic tree depth control strategy, where the maximum depth of the decision tree is positively correlated with the number of training samples.

7. The method according to claim 5, characterized in that, The redistribution of weights for the multi-dimensional time features includes: Based on the sliding window entropy algorithm, dynamic weights are assigned to features in the feature space, wherein the higher the entropy value of a feature, the greater the weight assigned.

8. The method according to claim 1, characterized in that, The minimum control line parameters and cash flow volatility parameters in the dynamic calibration cash management model include: The rolling window method is used to identify the maximum single-day net capital outflow within a preset historical time window and use it as the minimum control line for the current period. Calculate the first volatility of historical cash flow data; calculate the second volatility of the cash flow forecast result; weight and fuse the first volatility and the second volatility to obtain a mixed volatility, and use the mixed volatility as the cash flow volatility parameter.

9. The method according to claim 1, characterized in that, Also includes; To obtain the market value of securities; Based on the market value of the securities, the volatility of the market value of the securities in the cash management model is adjusted in real time.

10. An intelligent cash management system for enterprises, characterized in that, include: The acquisition module is used to acquire the historical cash flow data of the enterprise and construct a multi-dimensional time feature space of the enterprise's cash flow data. The feature space includes at least a first type of time feature based on a preset calendar period and a second type of time feature based on a specific business cycle of the enterprise. The prediction module is used to input the multi-dimensional time feature space into the prediction model to obtain the cash flow prediction result for the future target period; wherein, the prediction model is a deep learning model; The calibration module is used to dynamically calibrate the minimum control line parameter and cash flow volatility parameter in the cash management model based on the cash flow forecast results and historical cash flow data; wherein, the cash management model is a financial model based on the Miller-Olé principle; The generation module is used to generate decision-making criteria for cash management based on the dynamically calibrated cash management model.

Images