Objectives of the document

This document provides a technical overview of the solution delivered by D3S. It explains the building blocks of the source code and the methodology followed to obtain insightful predictions for cash recovery.

General presentation of the solution

Business stake and approach

The main business stake is to increase overdue coverage with the existing task force. As of today, dunning and pre-dunning actions focus on the largest outstanding amounts, leaving aside smaller accounts (below a threshold). Pre-dunning include some additional rules applied by cash collection teams through a time-consuming manual process.
Cash collection is steered with End of Month KPIs. Although not necessarily representative of the cost of working capital, EOM metrics are relevant as they are fully aligned with other business steering indicators.
Predictive analytics are a way forward, especially to better address the smaller accounts on which the overdue rate is higher.
Wrap Up
Figure 1: objectives and core principles

• Machine learning technology models deep correlation patterns from past behavior
• A model is trained on 2 years' history, including all documents paid in month or not
• The predictive engine generates risk scores allowing to focus on the main stakes
• Cash collectors follow the proposed priority, applying their expert judgment as required (known event at payer, dunning blocked, etc.)
• Model performance is monitored monthly through a dedicated report (QlikView)
• Statistic on each payer's past payment behavior are also made available without drilling down into SaP

Machine learning methodology

The predictive solution leverage machine learning technology. A model is first trained on payment history to learn customer behavior based on all available caracteristics. For new customers, the model infers behavior based on available data (country, currency, sector, invoice-caracteristics, etc)

Figure 2:overview of the machine learning methodology

Outputs available to Cash collectors

The predictive solution generates cash recovery insights at payer and document levels. The illustration below shows the indicators computed at PRS Customer Zone/GBU/PRS Customer level. Similar information is available to drill down on underlying documents, either open or due in the current month.

Figure 3: illustration of indicators computed by the solution (view per GBU)

Technical description

Solution Architecture

Overview

There are four building blocks in or interacting with the solution:

• BW/SAP server: interaction in/out from the SAP system
• SFTP/SSH server/client: file exchange (in/out) and remotely process execution
• Python server: predictive engine and working list generation
• MSSQL server: data storage, features engineering and processing

Figure 4 is an overview of the solution architecture. BW pushes new data onto the BW SFTP server (1) and launches a remote command to trigger all processing (2). Python server connects to the BW SFTP server (using a sftp client) to download data files (3). These files are inserted in SQL server tables where they are processed and pushed back to the Python server for predictive modeling. Results are then pushed back to the SQL server. Result tables are available through BW, connecting directly to the SQL server (4).

Figure 4. Solution architecture overview

Communication process between BW and the solution server

• Execution is lauched through a SSH command, running a .bat file on the sftp server
• End of processing is indicated through the .bat return code
• BW then read a specific status table in SQL server, with a code indicating the status and the actions required:
  • CODE = 0:
    • Results ready, to be uploaded  BW reads result tables in the SQL server
  • CODE = 1
    • Error#1, data upload issue  BW regenerates the input files and re-launch the SSH command
  • CODE = 2
    • Error#2, issue encountered  specific manual action required (eg. Full reload of the data set)

Table 1 shows an example of the table status. Columns description are the following:

• TIMESTAMP: start timestamp of the command
• TIMESTAMP_END: end of the execution of the command
• COMAND: name of the command executed
• CODE: exist code of the application
• MESSAGE: description of the exit code

Table 1 Example of STATUS table

Inbound connectors

For inbound connector, BW is interacting with the sftp/ssh server (see Figure 5).

Figure 5. Inbound connectors sequence diagram
The BW server pushes csv files onto BW sftp server. The following csv file are expected:

• C_CST_CA2: credit code table (full replacement at each upload)
• C_COMPCDE: company code table (full replacement at each upload)
• C_CUSTID: customer code table (full replacement at each upload)
• GL_ACCOUNT: gl account code table (full replacement at each upload)
• G_CWWE01: sub activity table (full replacement at each upload)
• TCURR: exchange rate table (full replacement at each upload)
• COUNTRY: country table (full replacement at each upload)
• DBFIAR20: credit management table (incremental upload)
• DBFIAR21: order blocked table (incremental upload)

Then a ssh remote command launches the process called "pacm-run.bat"

Outbound connectors

Figure 6. outbound connectors with MSSQL sequence diagram
For outbound connector, BW is interacting with the MSSQL database (Figure 6) to read the results tables (schema cm_data). The end of the name of the results tables are linked to the server:

• PRODUCTION SERVER: <SERVER> = PRD
• PRE-PRODUCTION/TEST SERVER: <SERVER>=TST

The results tables are the following:

• PRIORITY_PER_AMOUNT_<SERVER>: predictive indicators and statistics on each open document within prechasing and chasing scope for the month
• PRIORITY_PER_PAYER_<SERVER>: predictive indicators aggregated by payers, used for cash collection priorities setting
• PERF_MONITORING_<SERVER>: stores predictive performance of the past 4 months (predictive performance indicators are computed on the first day of each month, from previous months' results).

For example for the production server the name of the first result table will be: PRIORITY_PER_AMOUNT_PRD

Solution building blocks

Figure 7 describes all the interaction against the different components.

Figure 7. Sequence diagram

MS SQL server

SQL server is used for data storage, features engineering and processing.
Version : MSQL Server 2012

Connection

Inputs/outputs

The input tables are attached to the schema cm_data:

• C_CST_CA2: credit code table
• C_COMPCDE: company code table
• C_CUSTID: customer code table
• GL_ACCOUNT: gl account code table
• G_CWWE01: sub activity table
• TCURR: exchange rate table
• COUNTRY: country table
• DBFIAR20: credit management table
• DBFIAR21: order blocked table

Results tables are attached to the schema cm_data:

• PRIORITY_PER_AMOUNT_<SERVER>: list of priority per invoices
• PRIORITY_PER_PAYER_<SERVER>: list of priority per payer
• PERF_MONITORING_<SERVER>: follow the performance of the past 4 months

Python

Version :

• python 3.5.2
• scikit-learn 0.17.1
• pandas 0.18.1
• pymssql 2.1.3
• openpyxl 2.3.2
• xlrd 1.0.0

Connection

Inputs/outputs

All interactions with the sftp server go through the pacm-workspace ("F:\pacm_workspace"). The workspace folders are the following:

• archive: folder used to store, for each run, the value in the inbox folder
• bin: folder with all the executable (run, train, reset_db, ping)
• etc: folder with the configuration files for the loging (loging.config) and for the application (pacm.config)
• inbox: composed of one folder (in) used to get new data
• logs: folder to store the logs

SFTP server

Version : Bitvise SSH Server 7.16

Connection

Inputs/outputs

Inputs use the folder workspace inbox folder: F:\pacm_workspace\inbox\in to push new data in the application

SFTP client

Version : Bitvise SSH Client 7.22

Connection

Inputs/outputs

Download all the file from /exploit/BW/PREDICTCM in the workspace inbox folder F:\pacm_workspace\inbox\in using the following command :

Deep-dive on solution building blocks

Table 1 describes the main step of the solution main steps. There is four main steps:

• Get data from BW
• Load data and compute features
• Predictive engine and working list generation
• Performance assessment

Table 2. Main steps

MS SQL server

MS SQL server is used for data storage, features engineering and processing. All the queries are located in: *app\core\sql*
We focus on the online version (app\core\sql\online) of the sql queries in this section. The offline queries (app\core\sql\offline) are used to build-up the training dataset. They can be found in annex "Off line sql queries" (page ). For each source table, all column names and types are described in annex DB table details (page 29).
Figure 8 describes the data model used in the MS SQL server. There are four main steps:

1. Create schema and table in the database
2. Preprocess data to compute predictive features (used by the Python predictive model)
3. Generate result tables, including predictive insights and statistics about past behavior
4. Evaluate predictive performance, comparing what was predicted and what was ultimately observed (monthly)

Figure 8. SQL Data model

Create database

Folder: app\core\sql\create_db
Description: clear and recreate the full schemas:

• cm : containing all input data
• cm_view : temporary views for online
• cm_view_training : temporary views for training
• cm_output: results
• cm_data: view used to interact with BW

List of files:

• 000_create_schema_cm_view_training: : create cm_view_training schema
• 0a_create_schema_cm.sql: create cm schema
• 0b_create_schema_cm_view.sql: create cm_view schema
• 0c_create_schema_cm_output.sql: create cm_output schema
• 1a_create_current_date_table.sql: create a table with the current date
• 1_create_table.sql: create input data tables
• 2_Solvay_functions.sql: define all the specific functions used in the queries
• 3_create_input_data_view.sql: create cm_data schema and view with the input data
• 4_create_payment_table.sql: create PAYMENT_DESC_TABLE with the description of the payment method

Preprocess data to compute predictive features

This section followings the graph numbering in Figure 8.

00 – current_data_table

Python function: app.simul.predictive_model.update_simulation_day()
This table contains the current timestamp in UTC, it is updated by python script after loading the data in the database.
Table 3. current_date_table preview

01 – currency_data_per_month

Folder: app\core\sql\create_view\online
Files: 01_transform_currency.sql
From: TCURR table
Where: -
Group by: -
Description: Fill the currency table based on the TCURR table with one exchange rate per month of the year. If for one month, it is not given, then use the last available
Note: K_FFACT and K_TFACT cannot be zero, zero values are replaced per 1
Table 4. currency_data_per_month table preview

02 – currency

Folder: app\core\sql\create_view\online
Files: 02_Currency_preprocessing.sql
From: currency_data_per_month
Where: -
Group by: -
Description: Compute historical statistics for each currency
Table 5. currency table preview

02a – update currency table

Folder: app\core\sql\create_view\online
Files: 02a_modif_currency_table_for_JPY_KRW
From: currency_data_per_month
Where: -
Group by: -
Description: Update CURRENCY table for currency JPY and KRW. SAP store JPY and KRW currency / 100. This table modify the currency table to add a multiplication per 100 (K_TFACT column).
Warning: Hard coded data modification on specific currency JPY and KRW

03 – credit_limit

Folder: app\core\sql\create_view\online
Files: 03_Credit_limit_preprocessing.sql
From: C_CST_CA2, currency
Where: -
Group by: -
Description: Convert the credit limit in EUR
Table 6. credit_limit table preview

04– documents_view

Folder: app\core\sql\create_view\online
Files: 04_Documents_list.sql
From: DBFIAR20, C_CUSTID, GL_ACCOUNT, current_date_table, currency, C_COMPCDE, G_CWWE01, PAYMENT_DESC_TABLE
Where: GL_ACCOUNT."C_GL_TYPE" = 'RECEIVABLES' and (GL_ACCOUNT."C_GL_STYP" = 'PRODUCT AND SERVICES' or GL_ACCOUNT."C_GL_STYP" = 'SERVICES NON DOUBT' or GL_ACCOUNT."C_GL_STYP" = 'PRODUCT NON DOUBT') and DBFIAR20."0C_CTR_AREA" = 'SOLV'
Group by: -
Description: Preliminary filter, to reduce dimensions and transform key datas used in other views. Convert and consolidate the amount in EUR.
Note: For converting the amount in euro, the rate associated with the "0CLEAR_DATE" is used. If the document is still open, then the most recent exchange rate is used.
Warning: Hard coded document filtering is implemented here

05– documents_perimeter

Folder: app\core\sql\create_view\online
Files: 05_Documents_filtered_perimeter.sql
From: documents_view
Where: (documents_view."0CREATEDON" < documents_view."0CLEAR_DATE" or documents_view."0CLEAR_DATE" is null) and documents_view."Amount EUR" is not null
Group by: -
Description: Full data sample on which customer features are computed. Targets to be predicted are computed, as well as most features directly related to the document or static customer attributes. Condition on documents."Amount EUR" allow to remove the documents outside the scope. Table 7 explains the WHERE clause.
Table 7 Explanation of the where clause for documents_perimeter

06– clearing_list_for_payer_history

Folder: app\core\sql\create_view\online
Files: 06a_Clearing_list_by_payer_for_payer_history.sql
From: documents_view
Where: -
Group by: "Payer ID"
Description: this table computes the average payment day and week of a customer for the last 12 months

07– OB_payer_current_date

Folder: app\core\sql\create_view\online
Files: 07_OB_payer_per_current_date.sql
From: DBFIAR21
Where: -
Group by: "Payer ID "
Description: compute the block orders information per payer (number of blocked order, average resolution time)

08– payer_history_per_month

Folder: app\core\sql\create_view\online
Files: 08_Payer_history_per_month.sql
From: documents_perimeter
Where: ((documents_perimeter.C_FCONNUM is not null and documents_perimeter.[0LOGSYS] = 'PI1_020' and ([0CLEAR_DATE] >= '2017-02-01' or [0PSTNG_DATE] >= '2017-02-01')) or ([0CLEAR_DATE] < '2017-02-01') or (documents_perimeter.C_FCONNUM is null))
Group by: "Payer ID"
Description: past payer behavior for the last 12, 6 and 3 month, accounting for all previous documents keeping only the legal document for the document with a factoring contract number.

10– OD_at_current_Date

Folder: app\core\sql\create_view\online
Files: 10_Open_Documents_at_current_date.sql
From: documents, current_date_table
Where: -
Group by: -
Description: for each payer and current date, this view computes contextual indicators accounting for all open documents

11– core_perimeter

Folder: app\core\sql\create_view\online
Files: 11_Core_perimeter.sql
From: documents, current_date_table, currency, payer_history, payer_history_distinct, OD_at_Due_Date, credit_limit, OB_payer_due_date, DBFIAR20, C_COMPCDE, C_CST_CA2, C_CUSTID, GL_ACCOUNT, G_CWWE01, COUNTRY
Where:

• documents_perimeter."0FI_DOCSTAT" = 'O' : keep only open document
• (current_date_table."Current Date" < C_COMPCDE."C_MERGDAT" or C_COMPCDE."C_MERGDAT" is null): remove company merged
• documents_perimeter."0NETDUEDATE" <= EOMONTH(current_date_table."Current Date"): keep only open document with a net due date in the current month

Group by: -
Description: aggregate all previously computed information into one singe table. This table will then be used to apply the predictive model in Python.

12–result_table

Python function:

• app.core.model.predictive_model_pckg.get_open_documents_from_db (): extract from core_perimeter all the data where "0NETDUEDATE" is before or in the current month + (Apply predictive model)
• app.core.model.predictive_model_pckg. push_pred_in_db(): push the prediction in the result table

13–PRIORITY_PER_AMOUNT

Folder: app\core\sql\report
Files: 0a_outstanding_amount.sql, 0b_priority_per_amount.sql
From: DBFIAR20, result_table
Where: -
Group by: -
Description: for each document, this view computes contextual indicators accounting for all open documents

14–PRIORITY_PER_PAYER

Folder: app\core\sql\report
Files: 00_Clearing_list_by_payer_for_payer_history_per_GBU.sql, 00_Payer_history_per_month_per_zone_logsys_c_custid.sql, priority_payer.sql, 1_priority_C_CUSTID_payer.sql
From: PRIORITY_PER_AMOUNT
Where: -
Group by: [Payer ID], [0LOGSYS], [C_CUSTID], [PRS Cust. Zone (FI)], [GBU]
Description: for each GBU/payer, this view computes contextual indicators accounting for all open documents

15a–result_table_with_truth

Folder: app\core\sql\evaluate_perf
Files: 01_create_truth_table.sql
Description: Merge the result table of the prediction with what was observed during the previous month.

15b–PERF_MONITORING

Use for the performance evaluation – see User guide/Monitoring
Folder: app\core\sql\evaluate_perf
File: 02_perf_per_day.sql
Description: Performance of the algorithm is computed for each day and each zone from the previous month information. This query is executed at the beginning of each month.

Python

Table 8 describes the main python function called during the generation of the working list.
Table 8. Main python function for working list generation

The main parts of the python code are in:

• "app/core": main function for model prediction
• "app/simul": function used for simulation
• "app/sklearn_ext": function used to preprocess the data
• "app/io": function to read or write a files
• "app/utils": useful functions like sql connection, ….

User guide

Installation

Two zipped files are needed to install the solution:

• pacm-0.8.0.zip: pacm application
• pacm-model.zip: predictive model

Step-by-step installation guide:

1. Unzip the application file in the folder "E:\PACM<version>"
2. Copy the folder "E:\PACM<version>\workspace_template" to "F:\pacm_workspace"

3. Update the BIN_FOLDER in : "F:\pacm_workspace\bin\ _env.bat"

   set BIN_FOLDER= E:\PACM<version>\bin

4. Configure the "F:\pacm_workspace\etc\pacm.config":

   <ac:structured-macro ac:name="unmigrated-wiki-markup" ac:schema-version="1" ac:macro-id="858e273f-7ca0-4621-9072-46520c8699d8"><ac:plain-text-body><![CDATA[

   [python server] ]]></ac:plain-text-body></ac:structured-macro>

5. server wher the application is run TST or PRD
   Server=TST
   
   [sql_server]
   server=***
   user=***
   password=***
   database=***
   port=***
   

6. folder used to bulk insert the result in the table
   base_bulk_insert_folder=\\WDCPAAT21\bulk_insert
   
   [model]
   

7. to skip the import_data step in case data is inserted directly in db
   #skip_import_data=True
   
   [predict]
   

8. model folder to use
   model_folder= F:\pacm_workspace\models\20161111
   
   [train]
   

9. base model folder where the new trained model folder will be created,

10. named according to current timestamp
    model_base_folder=F:\pacm_workspace\models\
    training_period_in_month = 36
    testing_period_in_month = 6
    
    
    [dynamic_threshold]
    

11. dynamic threshold use to adjust the risk for customer with large outstanding amount
    amount_threshold_list = [(300000, 0.5), (600000, 0.25), (1200000, 0)]
    
    |

12. To test the connection to the db run: "F:\pacm_workspace\bin\ping.bat"
13. To create the db run: "F:\pacm_workspace\bin\create_db.bat"
14. Unzip the model (pacm-model.zip) in: "F:\pacm_workspace\models\20161111"
15. Launch the executable: "F:\pacm_workspace\bin\run.bat"

Figure 9. pacm workspace
Figure 9 shows the folder in the pacm-workspace

Tunning of the dynamic model

One parameter see the dynamic threshold section in the configuration file. This threshold allows to adapt the risk based on the outstanding amount of a customer. This dynamic threshold is used to compute the Risk Amount for the documents with a positive amount with the following formula:
RiskAmount= Outstanding Amount* Risk Probability
with Risk Probability=if Probability not PIM<threshold then 1if Probability not PIM>threshold then Probability not PIM
Table 9 shows the dynamic threshold.
Table 9 Dynamic probability threshold

These thresholds can be adjust in the configuration file in the "Dynamic treshold" section:

1. dynamic threshold use to adjust the risk for customer with large outstanding amount
   amount_threshold_list = [(300000, 0.5), (600000, 0.25), (1200000, 0)]
   For example:

• If a customer has an outstanding amount of 1 500 K€, then the probability threshold is set to 0%. It means that all documents of the customer with a positive amount will have a risk probability set to 1. The other documents will keep the value of probability computed by the model.
• If a customer has an outstanding amount of 700 K€, then the probability threshold is set to 25%. It means that all documents of the customer with a positive amount and with a probability not PIM over 25% will have a risk probability set to 1. The other documents will keep the value of probability computed by the model.

• If a customer has an outstanding amount of 100 K€, then the probability threshold is set to 100%. It means that all documents of the customer will keep the value of probability computed by the model.

Monitoring

Performance

The solution generates a performance monitoring table at the beginning of each month, comparing predictive insights generated with realized payment behavior. Description of the performing monitoring columns are in annex "Description of the performance monitoring table" (page )
The usual predictive performance monitoring metrics are:

• Recall Score: it quantifies the true positive rate (value between 0 and 1).

• Precision Score: it quantifies how many documents are flagged as positive (depending on the target this will mean 'not paid in month' or 'overdue') are actually so (value between 0 and 1).

• AUC: Area under the Curve, which plots Recall and precision for different decision threshold.

As the main objectif is to prioritize cash collection actions, the solution also generates Lorenz curves. They sort customers with different ordering criteria (simulating the working list ranking) and compare the total amount not paide in month for the first N customers. The curves allow to compare three different criteria and monitor the predictive ranking gain:
Customer number
Figure 10. Ranking performance

Technical

All application logs are stored in the folder "F:\pacm_workspace\logs".
All the input data are archived in the volume "F:\pacm_workspace\archive". If the folder size exceed the current volume size (100Go), the volume size should be increase.
If the size of the data base execeed the SQL database size, the data base should be increase.

Procedure to retrain the model

Attention point: model training is a long process (around 12 hours)
Three models are used:

• Model for prediction of the documents not paid in the current month (prediction_model_training_calibrated_1_PIM.pk)
• Model for prediction of the documents overdue (prediction_model_training_calibrated_1_overdue.pk)
• Regularisation model to correct the probability for long overdue documents (reg_model_All_norm.pk)

When to retrain the model?

The model should be retrained when one of the following happens:

• Major point-in-time degradation of the predictive performance, as shown with Lorentz curve metrics (see above)
• Smaller but progressive drift during two months, assessed through either:
  • Lorentz curve metrics (see above)
  • Expected not PIM amount ratio (comparing the expected not PIM amount with the realized amount not Paid In Month). Point-in-time delta can happen due to very large amounts on some overdue documents, but a continuous drift requires attention.

How to retrain the model (step-by-step)?

Retrain model for prediction of not paid documents in the current month / documents overdue

1. Set the training period in the configuration file (train section)

[train]

1. base model folder where the new trained model folder will be created,
2. named according to current timestamp
   model_base_folder=F:\pacm_workspace\models\
   training_period_in_month = 30
   testing_period_in_month = 6
3. To train the model run: "F:\pacm_workspace\bin\train.bat"
4. A new folder is created in the "model_base_folder"
5. Change in the predict section of the configuration file the new folder path

[predict]

1. model folder to use
   model_folder= F:\pacm_workspace\models\20161111
   This procedure uses lots of memory and is time consuming (~12h). It is advices to retrain the model on a local computer. To re-train localy the model:

• Copy the online DataBase on your local environment
• Install the application on your computer
• Run the train process

Retrain model for regularization model

To retrain the regularization model the data of the past four month are used (in table result_table_with_truth).

1. To train the model run: "F:\pacm_workspace\bin\ train_regularization_model.bat"
2. A new folder is created in the "model_base_folder"
3. Change in the predict section of the configuration file the new folder path

[predict]

1. model folder to use
   model_folder= F:\pacm_workspace\models\20161111
   
   

   What are the checkpoints?

   Check the model validity using the performance assessment file stored in the folder with the trained model. For the overdue model the file is : prediction_model_training_calibrated_1_overdue.csv and for the not paid in month model the file is : prediction_model_training_calibrated_1_PIM.csv.
   Model information stored in the file are the following:

• Total_observation : total number of observation
• Total_observation_Train : number of observations used to train the model
• Total_observation_Test : number of observations user to test the model
• AUC_score : Indicator of the global model performance. This score corresponds to the area under the precision-recall curve. This score should be over 0.85 (see monitoring performance section).
• Confusion matrix : precision, recall and support on the test data. The first line is for the document paid in month, the second line is for document not paid in month.
• Features importances of the variable in the model : List of the feature importance in the variable in the model

You will find below an example of prediction_model_training_calibrated_1_PIM.csv file.
prediction_model_training_calibrated_1_PIM
Total_observation : 8083490
Total_Observations_Train : 7028867
Total_Observations_Test : 1054623
AUC_score : 0.909
Scores_Matrix :
"precision,recall,support"
"0.889,0.939,807373.000"
"0.756,0.616,247250.000"
Features importances :
"variable_name,feature_importance"
"scenario,0.208"
"Rate_Not_PIM_last_12_month,0.087"
"AVG_delay_over_remaining_days_last_12_month,0.072"
"AVG_delay_over_remaining_days_last_3_month,0.052"
"AVG_delay_over_remaining_days_last_6_month,0.045"
"Ratio_remaining_days_in_month,0.041"
"C_COMPCAF,0.036"
"Amount_Rate_Not_PIM_last_12_month,0.033"
"Ratio_AVG_day_of_month,0.031"
"Rate_Not_PIM_last_6_month,0.029"
"Payment Term,0.029"
"0POST_KEY,0.028"
"Nb_docs_last_12_month,0.026"
"Nb_docs_last_3_month,0.025"
"Amount EUR,0.025"
"Nb_docs_last_6_month,0.024"
"Relative_Amount_vs_OD,0.020"
"Amount_ratio_Late_OD,0.019"
"Ratio_AVG_week_number,0.019"
"Rate_Not_PIM_last_3_month,0.017"
"nb_distinct_clear_date,0.016"
"STDEV_week_number,0.016"
"STDEV_day_of_month,0.015"
"Amount_Rate_Not_PIM_last_6_month,0.014"
"C_DOCTYP,0.014"
"LAST_C_PM_MTHD,0.012"
"GBU_group,0.012"
"Amount_Rate_Not_PIM_last_3_month,0.012"
"Create_post_lag,0.010"
"OD_Due_Date_vs_CL_EUR,0.005"
"OB_nb_overdue,0.004"
"C_COMPCDE__K_INTRAT,0.002"

Procedure for Exec builder

If some modification is made to the source code, then the application executable must be re-build. The steps to package the python source code in an application are :

1. Check the version of the application to build in the file "version.py"
2. Run "build.bat"

The compress application named "pacm-<version>.zip" is created in the folder "dist"
Then follow the installation steps of the user guide.

APPENDIX

Off line sql queries

For training/offline, all the function works in cm_view_training schema.
These offlines queries are run to compute the dataset used to train the model. To simulate the offline behavior for each day in the training period, a view of DBFIAR20 is created with only the data availaible at the simulation date. Based on this view, all the customer features are computed (same queries as online on cm_view_training schema). To reduce the size of the training data, only the 1, 5 ,10 ,15, 20, 25 and last day of month are used. Figure 11 shows the training data generation mechanism.

Figure 11 Training data generation

01 – create schema cm_view_training

Folder: app\core\sql\create_view\offline
Files: 000_create_schema_cm_view_training.sql
From: -
Where: -
Group by: -
Description: create the schema cm_view_training

02 – create credit_mgt_view

Folder: app\core\sql\create_view\offline
Files: 001a_modif_input_data_to_keep_doc.sql
From: DBFIAR20
Where: [0PSTNG_DATE] <= cm_view_training.current_date_table."Current Date"
Group by: -
Description: This view is a preliminary filter, to simulate the DBFIAR20 data availaible at a given date. This view is then used to apply all the online queries.

03 – APPLY THE ONLINE QUERIES ON CREDIT_MGT_VIEW in shema cm_view_training

Folder: app\core\sql\create_view\offline
Files: 01_transform_currency.sql, 02a_modif_currency_table_for_JPY_KRW.sql, 02_Currency_preprocessing.sql, 03_Credit_limit_preprocessing.sql, 04_Documents_list.sql, 05_Documents_filtered_perimeter.sql, 06a_Clearing_list_by_payer_for_payer_history.sql, 07_OB_payer_per_current_date.sql, 08_Payer_history_per_month.sql, 10_Open_Documents_at_current_date.sql
Description: Same queries than online apply on cm_view_training schema and based on the credit_mgt_view (modification in 04_Documents_list.sql in the FROM clause)

04– Save the core_perimeter table in core_perimeter_concat

Folder: app\core\sql\create_view\offline
Files: 12_Core_perimeter_concat.sql
From: core_perimeter
Where: -
Group by: -
Description: Save the previous core_perimeter table

05- Create training data with the updated target

Folder: app\core\sql\create_view\offline\final_table
Files: 0_create_result_view.sql
From: core_perimeter_concat
Where: ((C_FCONNUM is not null and 0LOGSYS] = 'PI1_020' and ([0CLEAR_DATE] >= '2017-02-01' or [0PSTNG_DATE] >= '2017-02-01')) or ([0CLEAR_DATE] < '2017-02-01') or (C_FCONNUM is null))
Group by:
Description: Generate the training data keeping only the legal document for the document with a factoring contract number

DB table details

Table 11 describes the input format of the table in the database.
Table 11. Table column definition

Description of the performance monitoring table