Machine Learning Applications

Overview

The prebuilt prediction applications based on machine learning model are designed to abstract complexity and provide these benefits:

Improved operations and reduced business risk.
Enhanced visibility and deeper insights that aren't available through exploratory analysis.
Ability to plan for the future by predicting outcomes.

Customer Collections Date Prediction (Preview)

The Customer Collections Date Prediction application predicts the date when an invoice will be paid by customer and calculates the expected delay days to help manage cash flows precisely.

The predicted payment date and predicted delay days related attributes are exposed in the Financials - AR Transaction and Financials - AR Aging subject areas.

Prerequisites

Ensure that the Accounts Receivables and Customer Collections Date Prediction functional areas are activated prior to enabling this application on the Enable Features page. See Activate a Data Pipeline for a Functional Area and Make Preview Features Available.

Configuration Parameters

Configure the Customer Collections Date Prediction application by selecting appropriate values for these parameters:

Historic Periods for Training data – The number of months of training data desired for training a prediction model. Model is expected to get trained and perform better on longer time frames like 60 months. Although there are no preset limits, it is recommended that at least 24 months of training data is provided.
Historic Periods for Scoring data – The number of months of data to consider for generating the prediction results. It must be comprehensive enough to consider all the open transactions for which the prediction output is expected. For example, if today is January 1, 2023 and the Historic Periods of Scoring data and Historic Periods of Training data are set to 6 months and 24 months respectively, scoring data will be from July 1, 2022 to December 31, 2022 and training data will be from June 30,2020 to June 30,2022 (24 months prior to the start of the Scoring data). Training and Scoring data don't overlap.
Future Period for Prediction – The number of future periods in months for which prediction scores will be available based on Scoring data. Some payment schedules extend to many years and this period controls how many months ahead need to be evaluated for risk of late payment. For example, if this is set to 12 months, predictions will be made only for those invoices for which payments are due within the next 12 months.
Invoice Threshold Amount – Predictions will be made for only invoices where due amounts are above the threshold amount. Use this to filter low valued invoices or set it as “0” to consider everything.
Extend Due Date by Days – For prediction model training, the invoice payment is considered late if it's still unpaid after the due date. Setting this parameter allows the prediction model to extend the due date which works as additional grace days. For example, if this parameter is set to 5 days, then this application doesn't consider invoices for which collections are overdue up to 5 days as late.

How to Use the Predictions

The prediction scores and related attributes are available in the Financials - AR Aging and AR Transactions subject areas. View Customer Date Prediction in the Customer folder and Date Prediction Statistics in the Transaction Details folder.

See Subject Areas.

Use these subject areas to create user defined analyses to know the total receivables amount that will get delayed, minimum, and maximum predicted delays for a specific customer.

Column	Definition
Minimum Predicted Delay Days	Minimum number of delay days predicted across invoices for the customer.
Maximum Predicted Delay Days	Maximum number of delay days predicted across invoices for the customer.
Average Predicted Delay Days	Average number of delay days predicted across invoices for the customer.
Customer Delay Decile	Decile ranking for the customer assigned based on Default Score between 1 to 10 in the increasing order of risk.
Customer Delay Percentile	Percentile ranking for the customer assigned based on Default Score between 1 to 100 in the increasing order of risk.
Coefficient of Variation	Calculated as variation between predicted delay days and actual delay days.
Payment Schedule Identifier	Payment schedule identifier for an invoice.
Predicted Payment Date	Predicted date for the payment for each individual invoice
Predicted Delay Days	Calculated as a difference of predicted delay days and schedule due date.

Frequently Asked Questions

Review these questions to understand the application:

How much data do we need for the prediction model to be accurate?

Accuracy of predictions improve if larger amounts of historical data is used for training. The minimum recommended training data is 2 years of invoice payment schedules and payments. The classifier accuracy metric shows the accuracy of model predictions. Minimum data must be 500 records and the recommended data must be tens of thousands.
How frequently does the model create predictions on future data?

The model calculates and generates predictions on future data daily. Previous predictions are overwritten based on the learnings from actual payment data. A snapshot of previous predictions is also maintained for historical reference.
How frequently is the model calibrated or trained?

The model is trained or recalibrated on a weekly basis to improve predictions over time.
What algorithms does the prediction model use?

The algorithms used is a proprietary multi-classification algorithm.

Customer Collections Risk Prediction (Preview)

The Customer Collections Risk Prediction application predicts the risk of a scheduled payment on an invoice being late because it won't be paid by the due date.

This application creates a risk score for each customer by considering the likelihood of late payment over all invoices. The risk score for an invoice or a customer is exposed in the Financials - AR Transaction and Financials - AR Aging subject areas.

Prerequisites

Ensure that the Accounts Receivables and Customer Collections Risk Prediction functional areas are activated prior to enabling this application on the Enable Features page. See Activate a Data Pipeline for a Functional Area and Make Preview Features Available.

Configuration Parameters

Configure the Customer Collections Risk Prediction application by selecting appropriate values for these parameters:

Historic Periods for Training data – The number of months of training data desired for training a prediction model. Model is expected to get trained and perform better on longer time frames like 60 months. Although there are no preset limits, it is recommended that at least 24 months of training data is provided.
Historic Periods for Scoring data – The number of months of data to consider for generating the prediction results. It must be comprehensive enough to consider all the open transactions for which the prediction output is expected. For example, if today is January 1, 2023 and the Historic Periods of Scoring data and Historic Periods of Training data are set to 6 months and 24 months respectively, scoring data will be from July 1, 2022 to December 31, 2022 and training data will be from June 30,2020 to June 30,2022 (24 months prior to the start of the Scoring data). Training and Scoring data don't overlap.
Future Period for Prediction – The number of future periods in months for which prediction scores will be available based on Scoring data. Some payment schedules extend to many years and this period controls how many months ahead need to be evaluated for risk of late payment. For example, if this is set to 12 months, predictions will be made only for those invoices for which payments are due within the next 12 months.
Invoice Threshold Amount – Predictions will be made for only invoices where due amounts are above the threshold amount. Use this to filter low valued invoices or set it as “0” to consider everything.
Extend Due Date by Days – For prediction model training, the invoice payment is considered late if it's still unpaid after the due date. Setting this parameter allows the prediction model to extend the due date which works as additional grace days. For example, if this parameter is set to 5 days, then this application doesn't consider invoices for which collections are overdue up to 5 days as late.

How to Use the Predictions

The prediction scores and related attributes are available in the Financials - AR Aging and AR Transactions subject areas. View Customer Risk Prediction in the Customer folder and Risk Prediction Statistics in the Transaction Details folder.

See Subject Areas.

Use these subject areas to create user defined analyses on risk of late payment for the open invoices or to understand the customers at risk of default.

Column	Definition
Customer Risk Decile	Decile ranking in which the customer falls in based on the customer risk score aggregated across invoices.
Customer Risk Percentile	Percentile ranking in which the customer falls in the customer risk score is aggregate across invoices.
Default Score	Probability of defaulting for the payment schedule.
Default Score Decile	Decile ranking for each record assigned based on Default Score between 1 to 10 in the increasing order of risk.
Default Score Percentile	Percentile ranking for each record assigned based on Default Score between 1 to 100 in the increasing order of risk.
Risk Score	Default score and the unpaid amount that indicates the amount at risk.
Schedule Risk Score Decile	Decile ranking for each record assigned based on Risk Score between 1 to 10 in an increasing order of risk.
Schedule Risk Score Percentile	Percentile ranking for each record assigned based on Default Score between 1 to 100 in an increasing order of risk.
Prediction Date Time	Date and time stamp for the last prediction scoring engine run.
Prediction Processed Indicator	(Y/N) This indicates whether the record was used for processing by the engine or not.
Run Date	Date and time stamp for the last prediction training engine run.
Classifier Score Bucket	Classifier scores from 1 to 100 assigned based on prediction model training data.
Actual Defaults	Number of actual defaults against a particular classifier score.
Classifier Predicted Defaults	Number of predicted defaults against a particular classifier score.
Classifier Accuracy Percentage	Calculated as Actual Defaults/Classifier Predicted Default.

Use these ready-to-use visualization projects to get started with customer collections risk analysis:

Customer Risk Analysis – Look at the overdue amount as a proportion of total outstanding amount for customers. Including risk classification based on deciles as well as risk analysis across payment terms.
Invoice Risk Analysis – Provides visibility for transactions contributing to the risk with default probability and risk score for each individual invoice.

These projects are available in Shared Folder > Oracle > Fusion ERP > Accounts Receivable > Customer Collections Risk Analysis.

Frequently Asked Questions

Review these questions to understand the application:

How much data do we need for the prediction model to be accurate?

Accuracy of predictions improve if larger amounts of historical data is used for training. The minimum recommended training data is 2 years of invoice payment schedules and payments. The classifier accuracy metric shows the accuracy of model predictions. Minimum data must be 500 records and the recommended data must be tens of thousands.
How frequently does the model create predictions on future data?

The model calculates and generates predictions on future data daily. Previous predictions are overwritten based on the learnings from actual payment data. A snapshot of previous predictions is also maintained for historical reference.
How frequently is the model calibrated or trained?

The model is trained or recalibrated on a weekly basis to improve predictions over time.
What algorithms does the prediction model use?

The algorithms used is a proprietary multi-classification algorithm.

Supplier On-time Payments Prediction (Preview)

The Supplier On-Time Payments Prediction application predicts the risk of a scheduled payment on an invoice being late because it won't be paid by the due date.

It also creates a risk score for each supplier by considering the likelihood of late payment for all the invoices. The risk score for an invoice or a supplier is exposed in the Financials - AP Invoices, Financials - AP Aging and Financials – AP Payments subject areas.

Prerequisites

Ensure that the Accounts Payables and Supplier On-Time Payments Prediction functional areas are activated prior to enabling this application on the Enable Features page. See Activate a Data Pipeline for a Functional Area and Make Preview Features Available.

Configuration Parameters

Configure the Supplier On-time Payments Prediction application by selecting appropriate values for these parameters:

Historic Periods for Training data – The number of months of training data desired for training a prediction model. Model is expected to get trained and perform better on longer time frames like 60 months. Although there are no preset limits, it is recommended that at least 24 months of training data is provided.
Historic Periods for Scoring data – The number of months of data to consider for generating the prediction results. It must be comprehensive enough to consider all the open transactions for which the prediction output is expected. For example, if today is January 1, 2023 and the Historic Periods of Scoring data and Historic Periods of Training data are set to 6 months and 24 months respectively, scoring data will be from July 1, 2022 to December 31, 2022 and training data will be from June 30,2020 to June 30,2022 (24 months prior to the start of the Scoring data). Training and Scoring data don't overlap.
Future Period for Prediction – The number of future periods in months for which prediction scores will be available based on Scoring data. Some payment schedules extend to many years and this period controls how many months ahead need to be evaluated for risk of late payment. For example, if this is set to 12 months, predictions will be made only for those invoices for which payments are due within the next 12 months.
Invoice Threshold Amount – Predictions will be made for only invoices where due amounts are above the threshold amount. Use this to filter low valued invoices or set it as “0” to consider everything.
Extend Due Date by Days – For prediction model training, the invoice payment is considered late if it's still unpaid after the due date. Setting this parameter allows the prediction model to extend the due date which works as additional grace days. For example, if this parameter is set to 5 days, then this application doesn't consider invoices for which collections are overdue up to 5 days as late.

How to Use the Predictions

The prediction scores and related attributes are available in the Financials - AP Aging, AP Invoices, and AP Payments subject areas. View Prediction Statistics (Supplier) and Prediction Statistics (AP Installments) in the Supplier folder.

See Subject Areas.

Use these subject areas to create user defined analyses on risk of late payment for the open invoices or to understand the suppliers at risk of default.

Column	Definition
Supplier Risk Decile	Decile ranking in which the Supplier falls in based on the Supplier risk score aggregated across invoices.
Supplier Risk Percentile	Percentile ranking in which the Supplier falls in the Supplier risk score is aggregate across invoices.
Default Score	Probability of defaulting for the invoice instalment.
Default Score Decile	Decile ranking for each record assigned based on Default Score between 1 to 10 in the increasing order of risk.
Default Score Percentile	Percentile ranking for each record assigned based on Default Score between 1 to 100 in the increasing order of risk.
Risk Score	Default score and the unpaid amount that indicates the amount which is at risk.
Schedule Risk Score Decile	Decile ranking for each record assigned based on Risk Score between 1 to 10 in an increasing order of risk.
Schedule Risk Score Percentile	Percentile ranking for each record assigned based on Default Score between 1 to 100 in an increasing order of risk.
Prediction Date Time	Date and time stamp for the last prediction scoring engine run.
Prediction Processed Indicator	Indicates whether the record was used for processing by the engine or not.
Run Date	Date and time stamp for the last prediction training engine run.
Classifier Score Bucket	Classifier scores from 1 to 100 assigned based on prediction model training data.
Actual Defaults	Number of actual defaults against a particular classifier score.
Classifier Predicted Defaults	Number of predicted defaults against a particular classifier score.
Classifier Accuracy Percentage	Calculated as Actual Defaults or Classifier Predicted Default.

Use these ready-to-use visualization projects to get started with supplier payments risk analysis:

Supplier Risk Analysis – Look at the overdue amount as a proportion of total outstanding amount for Suppliers. Including risk classification based on deciles as well as risk analysis across payment terms.
Invoice Risk Analysis – Provides visibility for transactions contributing to the risk with default probability and risk score for each individual invoice.

These projects are available in Shared Folder > Oracle > Fusion ERP > Accounts Payable > Supplier On-Time Payments Risk Analysis.

Frequently Asked Questions

Review these questions to understand the application:

How much data do we need for the prediction model to be accurate?

Accuracy of predictions improve if larger amounts of historical data is used for training. The minimum recommended training data is 2 years of invoice payment schedules and payments. The classifier accuracy metric shows the accuracy of model predictions. Minimum data must be 500 records and the recommended data must be tens of thousands.
How frequently does the model create predictions on future data?

The model calculates and generates predictions on future data daily. Previous predictions are overwritten based on the learnings from actual payment data. A snapshot of previous predictions is also maintained for historical reference.
How frequently is the model calibrated or trained?

The model is trained or recalibrated on a weekly basis to improve predictions over time.
What algorithms does the prediction model use?

The algorithms used is a proprietary multi-classification algorithm.