Predictive Modeling for Accounts Receivable

Predictive Modeling for Collections of Accounts Receivable

• Accounts receivable (AR) can cause financial difficulty for firms if poorly managed.
• Effective AR management correlates positively with overall firm financial performance.
• The study aims to reduce outstanding receivables through improvements in collection strategy.
• Supervised learning is used to predict payment outcomes of newly-created invoices.
• Models enable customized collection actions tailored for each invoice or customer.
• Models accurately predict if an invoice is paid on time and estimate delay magnitude using transaction data.

Categories and Subject Descriptors

• Probability and Statistics, Simulation and Modeling, and Administrative Data Processing are used.

Key Terms

• Accounts Receivable, Payment Collection, Order to Cash, Invoice to Cash, Predictive Modeling, Knowledge Discovery are used.

Introduction to Order-to-Cash (O2C) Process

• The Order-to-Cash (O2C) process comprises steps to fulfill an order, from order entry to payment receipt.
• The number and nature of steps vary based on firm size and type, but most O2C processes follow a similar workflow as shown in Figure 1.
• The study concentrates on the core of collections activities: account prioritization, customer contact activities, collection calls, escalation, and dispute resolution.
• These essential steps are often slow, expensive, and inaccurate due to manual processing.
• Generic collection actions do not account for customer specifics, causing inefficiencies that delay AR collections.
• Better management of collection steps can significantly improve AR collection effectiveness.
• Prioritizing delinquent invoices based on expected payment time can optimize the use of collection resources.
• The study focuses on predicting payment outcomes of newly-created invoices to enable effective collections management.

Invoice Outcome Prediction

• There are various metrics used to measure collection effectiveness of a firm.
• Average Days Delinquent measures the average time from invoice due date to payment.
• Days Sales Outstanding (DSO) calculates the average time (in days) that receivables are outstanding.
• Predicting invoice outcome enables use of information to improve a desired collection metric proactively.
• Identifying invoices likely to be delinquent allows for reducing time to collection.
• Preemptively contacting potentially delinquent accounts can benefit significantly.
• Knowing which invoices are likely to be paid sooner is beneficial when an invoice is past due.
• Prioritize invoices based on estimating payment lateness due to limited collection activities resources.
• The invoice outcome prediction task is a supervised learning problem.
• The model predicts when a newly-created invoice will be paid, classifying each instance into one of five classes.
• on time
• 1-30 days late
• 31-60 days late
• 61-90 days late
• more than 90 days late
• Data instances correspond to features representing invoices.
• The formulation uses Average Days Delinquent as the collections performance metric.
• An alternative target (class) variable correlated with the desired metric maximizes a different performance metric.
• The Collection Effectiveness Index (CEI) compares amounts collected in a period to what was available to collect.
• Invoices can be labeled based on actual amounts collected in a specified time period.

Data Preprocessing

• The analysis uses invoice records for four firms, including two Fortune 500 companies.
• three firms supply high-tech equipment for telecommunication, networking, and IT services
• the fourth firm specializes in online advertising placement and scheduling
• Data sets cover invoices created from March 2004 to February 2005.
• The study differentiates invoices of first-time customers from returning customers due to different payment behaviors.
• Additional historical information is available for invoices of returning customers.
• Returning invoices are invoices from returning customers.
• Most invoices of three firms are from returning customers, while the majority of invoices for one are billed to first-time customers.
• The input data consists of a set of invoices at the end of the collections cycle.
• Each invoice is described by 54 features that capture information.
• order details
• terms and conditions
• sales representative information
• The study eliminates features specific to an invoice (like invoice IDs) and leakage variables.
• Excluded categorical features have too many distinct values, features with too many missing values, and unique values.
• This leaves invoice-level features (invoice base amount, payment terms, and invoice category).
• Each invoice/instance is labeled with one of the five class labels based on delinquent days computed.
• The three invoice-level features are typically insufficient for effective modeling.
• The study develops additional features that capture the transaction history of a customer and reflect the current status of customer accounts.
• Historical and aggregate features provide significant information for predicting the outcome on a new invoice.
• The study focuses on building predictive models for invoices of returning customers due to information availability.
• The task of modeling on invoices without history is also discussed for completeness.

Summary of Features

• Invoice base amount, payment term, and Category
• Number of total paid invoices and Numbers of invoices that were paid late
• Ratio of paid invoices that were late
• Sum of total paid invoices and Sum of the paid invoices that were late
• Ratio of sum of paid base amount that were late
• Average days late of paid invoices
• Number of total outstanding invoices and Number of the the outstanding invoices that were already late
• Ratio of outstanding invoices that were late
• Sum of the base amount of total outstanding invoices
• Sum of the base amount of outstanding invoices that were late
• Ratio of sum of outstanding base amount that were late
• Average days late of outstanding invoices

Approach

• The task is a supervised classification problem: build a model to classify a new instance into target classes.
• The different settings for studied invoice prediction are on time, 1-30 days late, 31-60 days late, 61-90 days late, and more than 90 (or 90+) days late.
• The study compares C4.5 decision tree induction Naïve Bayes
• PART algorithm performed the best classification

PART Algorithm

• PART builds a partial pruned decision tree similar to C4.5
• The leaf with the largest coverage becomes the rule
• The rest of the tree is discarded
• Overfitting is avoided where rules use a minimum of 100 instances
• In addition to being accurate, PART can handle missing values, nominal values, and produces comprehensible models.
• Experiments were run using 10-fold cross validation and classification accuracy is reported.
• Models were built using just invoice-level features and those using both invoice and historical features.
• Experiments were conducted using data for invoices of returning customers.
• Historical features are only meaningful for these invoices.

Using Historical Data

• Historical and aggregate features provide significant information beyond invoice-level features.
• The payment outcome of an invoice can be predicted accurately more than predicting the majority class.
• Incorporating historical features increases prediction accuracies for all firms
• Using predictions to drive collections workflow can perform better than treating all invoices equally.

Cost-sensitive Learning

• Customizing monitoring and collection activities according to past due behavior is important.
• Focusing modeling on time in payment is a focus due to collection performance.
• These invoices are bad debt requiring special decision making.
• The study builds models to predict the 90+ class accurately.
• By default, algorithms assume that all classes are equally important.
• High-risk invoices are under-represented in the data, causing the task to be compounded.
• To address different misclassifications costs, instance re-weighting is used.
• Instances are re-weighted proportionally to the sum of their misclassification costs.
• In data, instance re-weighting is effective in dealing with class imbalance.

Prediction for New Accounts

• Invoice payment risk depends on customer financial capability and willingness to pay.
• Factors include customer credit worthiness, organizational profile, business market profile, etc.
• Experiments combining feature sets on data resulted in having the customer feature boost prediction accuracy on first-time invoices.
• This proves that customer-level features provide collectability information.

Unified Model vs. Firm-specific Models

• Building models help generalize and learn behaviors common to all firms
• Unified model combining training sets and its use results in performance improvement
• Separate firm models obtain better prediction accuracy
• Collection processes and invoice behaviors for each firm are different.
• There is merit in building individual prediction models for each firm.

Related Work

• Several vendors offer order-to-cash solutions, but these do not incorporate analytics or predictive modeling.
• Credit management and tax collection use predictive modeling approaches.
• O2C domain studies focus on predicting collection amounts on customer accounts or improving cash collection strategy.
• Rule engines are used to prioritize invoices to maximize cash flow, automatically prioritizing based on rule-based system

Conclusions

• The study developed a supervised learning approach for AR collections.
• A set of aggregated features captures historical payment behavior for each customer.
• Having specific features enhances the prediction accuracy and is more valuable in predicting payment delays.
• Using cost-sensitive learning improves prediction accuracy to high-risk invoices.
• Finally the work showed results comparing firm and single generic models.