Cashcast
An open-source accounts receivable forecasting tool that runs locally on the TabPFN tabular foundation model. It predicts invoice overdue probability, expected payment dates, and cash flow for the next eight weeks, while keeping data on the local machine.
Timeline (last 90 days)
Cashcast is an open-source tool that uses the TabPFN tabular foundation model to predict on-device the overdue probability of unpaid invoices, expected payment dates, and cash flow for the next eight weeks, with data never leaving the local machine.
Cashcast reads the user's existing CSV export files, with one invoice per row, containing fields for customer, invoice date, due date, amount, and payment date.
Cashcast outputs four judgments: overdue probability, expected payment date with an 80% interval, recommended next action, and an eight-week weekly cash forecast.
Cashcast uses two TabPFN v2 models on the same set of features: a classifier to determine whether an invoice will be paid after its due date, and a regressor to predict payment days, requesting 49 quantiles.
The month-by-month backtest covers January to October 2013, training each month on the most recent 1,000 invoices that had been paid before that month, producing a total of 1,075 predictions.
In the Reveal validation of the online demo, 78% of the 296 "send reminder now" flags were indeed paid late, while the overdue rate for all not-yet-due invoices in the same period was 43%.
In the online replay validation, the proportion of actual payment dates falling within the 80% interval was 77.1%, with a typical deviation of 5.78 days; the weekly cash forecast fell within the 80% interval only 67.5% of the time, because the simulation treated invoices as independent, making the intervals too narrow.
All Facts (7)
Cashcast is an open-source tool that uses the TabPFN tabular foundation model to predict on-device the overdue probability of unpaid invoices, expected payment dates, and cash flow for the next eight weeks, with data never leaving the local machine.
50%UnverifiedCashcast reads the user's existing CSV export files, with one invoice per row, containing fields for customer, invoice date, due date, amount, and payment date.
50%UnverifiedCashcast outputs four judgments: overdue probability, expected payment date with an 80% interval, recommended next action, and an eight-week weekly cash forecast.
50%UnverifiedCashcast uses two TabPFN v2 models on the same set of features: a classifier to determine whether an invoice will be paid after its due date, and a regressor to predict payment days, requesting 49 quantiles.
50%UnverifiedThe month-by-month backtest covers January to October 2013, training each month on the most recent 1,000 invoices that had been paid before that month, producing a total of 1,075 predictions.
50%UnverifiedIn the Reveal validation of the online demo, 78% of the 296 "send reminder now" flags were indeed paid late, while the overdue rate for all not-yet-due invoices in the same period was 43%.
50%UnverifiedIn the online replay validation, the proportion of actual payment dates falling within the 80% interval was 77.1%, with a typical deviation of 5.78 days; the weekly cash forecast fell within the 80% interval only 67.5% of the time, because the simulation treated invoices as independent, making the intervals too narrow.
50%