About ML-Based Forecasting Methods
Predictive Cash Forecasting provides machine learning prediction methods.
Note:
Machine Learning will be supported in a future update.Short Term Machine Learning Prediction is designed for invoice-level predictions and is applicable primarily for daily and short-term forecasting needs. (OCF_Short Term Machine Learning Prediction in the Forecast Method dimension.)
Mid Term Machine Learning, for short- and mid-term predictions, is based on historical data. (OCF_Machine Learning Prediction in the Forecast Method dimension.)
This method enables ERP systems to efficiently handle and load short-term machine learning forecasts, enhancing forecasting accuracy and responsiveness at the invoice level.
- Machine Learning as a forecast method is best suited for a periodic forecast tenure aligned to the standard payment term. Example, if the standard payment term for majority of customer transactions is 45 Net, periodic cash forecast up to 45 days in future will have a higher accuracy compared to the 90 th day period. The accuracy of the machine learning model tends to reduce beyond this period due to the reduced availability of transactions which are due 45 days in future in the application as well as due to the reduced availability of the factors contributing to understanding customer payment patterns due to the long lead time to the due date.
- The frequency of training is predetermined as weekly in the solution. Weekly training of the data helps in the model access the most recent historical data and thereby learn the data drifts due to changing trends in the business, its seasonality and other macro-economic factors as well.
- The frequency of prediction by default is predetermined in the solution. However,
the model predictions uptake can be decided based on multiple factors.
- Cash forecast schedule - Organizations performing daily forecasts will benefit in initiating predictions every day or even multiple times within the day to have the cash forecast being most accurate. Organization with a weekly or monthly forecast can have an uptake schedule aligned to the forecast run.
- Volatility of the business - If the organization business may experience high volatility and high volume of transactions like in consumer goods, retail industries, frequent initiating of machine learning predictions is recommended.
- Fusion Accounts Receivable customer’s billing business process can be broadly
categorized into 3 distinct billing models that will capture the industry
variations.
- Model1 is where rating and charges are performed from external applications
and the billing, bill presentment, collections, cash applications are
performed in Cloud Financials. This model is best suited for B2B, B2G2B
customers in Professional Services, Hi Tech, Manufacturing, Wholesale and
Distribution, Logistics industries.
Machine Learning as a prediction method is best suited for customers in model1 since there is an on time availability of transactions along with an accurate coverage of its lifecycle for training and predictions.
- Model2 is where rate and charge generation and presentment is performed in
the source but the same is replicated in Receivables in order to allow
collection, cash application activities to be performed in Cloud Financials.
This model is best suited for B2B2C, B2C customers in Education, Oil and
Gas, Retail industries.
- Customers in model 2 are also well suited for machine learning as a prediction method. It is recommended that customer identify those set of transactions which are immediately paid using credit card, bank transfers, wallet payments be excluded from training. Such transactions when comingled with due dated based transactions may result in the model incorrectly learning the customer’s payment behaviour.
- Care must be taken to ensure that the transaction replicated in Fusion application accurately captures the source information regarding transaction date, payment terms, due dates, details of the underlying sales in the transaction lines for the item or services, quantities, and amount of sale for the model to learn. Incorrect or alias information will negatively impact the model performance
- Model3 is where rate and charge generation, bill presentment, collection,
cash application activities are performed in the source application but only
the accounting information is brought over to Cloud Financials. This model
is best suited for B2B2C, B2C customers in telecommunications, financial
services, Utilities billing etc.
- Customers in model3 may have lesser use of machine learning as a prediction model as both the billing and collections are performing outside of the ERP system and brought into receivables for accounting and reporting purposes only
- Clear identification of such transaction and their exclusion from the training data set is recommended.
- Open transactions migrated or converted from legacy applications to Fusion Receivables and closed in Fusion Receivables immediately for updating customer balance, customer correspondence and reporting should be excluded from training and prediction purposes as they do not reflect the true customer payment behaviour and will negatively influence the model learning.
- There are transactions created to support some business practices which are not best suited for machine learning predictions due to their atypical nature. These include negative invoices, positive credit memos, zero-dollar invoices which are netted against customer payments. Intercompany receivables transactions which do not have an actual settlement or are netted also fall in this category.
- Receivables Transaction sources and Receivables Transaction Types which attribute such exclusion transactions can be configured to be excluded from training and predictions
- The machine learning model will learn the customer payment behaviour unique to each customer by learning from both the transactional data as well as customer reference data relating to credit and customer profile information. Hence it is important for the customer setup and its segmentation for credit and collections profiles be aligned to how billing is performed, payments are received, credit attributes maintained, and collections performed.
- Usage of Fusion Credit Management for maintaining data points, scoring models and case folders in arriving at the customer credit classification, credit recommendation and scoring will help the machine learning model know more about the customer credit profile.
- Usage of Fusion Collections for assigning collectors, customer scoring and segmentation, managing delinquencies, performing dunning and collection strategy tasks will help the machine learning model know more about the customer collection profile.
- The machine learning model accuracy is measured based on the cash forecast accuracy and it can be compared to the average days delinquent as the benchmark comparison.
- The machine learning model is optimized for the entities that is the primary line of business and brings the maximum revenue to the organization.
- Model1 is where rating and charges are performed from external applications
and the billing, bill presentment, collections, cash applications are
performed in Cloud Financials. This model is best suited for B2B, B2G2B
customers in Professional Services, Hi Tech, Manufacturing, Wholesale and
Distribution, Logistics industries.
For Model Training: Below are the recommended practices for machine learning model to maximize its learning:
- Availability of minimum 18 months of historical transactions with complete record of the related transactional activities is required (Example: Receipt Applications, Adjustment, credits to be available in accurate state)
- Historical transactions should have similar business or transaction attributes as the current transaction eligible for prediction. Example: Uptake of a new entity from a legacy application or roll out of new geography to Fusion Receivables, support of a new business process flow for amendments, returns are some practical reasons due to which current transaction may not correlate to the historical transactions