Financial Management System for SMEs: Real-World Deployment of Accounts Receivable and Cash Flow Prediction

AI summary

Overview

Research area: Applied machine learning for financial management in Small and Medium Enterprises (SMEs), combining invoice payment-delay classification with cash flow forecasting in a deployed production system.

Technical level: Intermediate. The paper assumes familiarity with standard classifiers (SVM, decision trees, random forests, kNN, Naive Bayes) and time-series baselines (ARIMA, Prophet, SVR), but the emphasis is on system architecture and deployment constraints rather than novel algorithms.

Scope: One sentence: the paper describes a two-module financial prediction system built with the startup Cluee, reports its classification and forecasting accuracy on three invoice datasets plus synthetic work-session data, and discusses the design trade-offs of predicting finances for businesses with very little historical data.

What This Paper Is About

SMEs, freelancers and early-stage businesses operate with small customer bases, short transaction histories and irregular income, so a single late invoice can consume a large share of monthly revenue. Existing financial forecasting research is built for large enterprises with rich historical records, leaving a gap for methods that work under sparse, incomplete and heterogeneous data. The authors build and deploy a system that predicts which invoices will be paid late and uses those predictions to forecast near-term cash flow, then report how it performed on real and public datasets.

Key Contributions

1. A deployed, integrated system combining accounts receivable prediction and cash flow forecasting, designed specifically around SME constraints rather than enterprise-scale data assumptions.
2. Practical engineering solutions for limited, incomplete and heterogeneous financial data, including per-business model training, moving-average trend features, and four separate cash-flow sub-modules so the system still produces output when some data categories are missing.
3. Deployment-based validation using a startup-provided dataset (297 invoices, 60 customers), two filtered public datasets, and a synthetic generator, along with lessons learned from building for resource-constrained businesses.
4. Explicit treatment of transparency and interpretability as design requirements, so predictions appear as inspectable decision-support signals rather than prescriptive recommendations.

Main Findings

• Startup data is the hardest case: On the Cluee dataset the best balanced accuracy was 0.56 (SVM and kNN), versus 0.72 on IBM and 0.68–0.70 on the Kaggle/Payment Date data. The authors describe 0.56 as "modest" but "meaningful given the severe data constraints," noting that 50% of customers had only 1–2 invoices.
• SVM selected as the primary classifier: Across datasets SVM scored 0.56 (Startup), 0.72 (IBM) and 0.68 (Kaggle). For comparison, Random Forest scored 0.52 / 0.69 / 0.67, Decision Tree 0.54 / 0.68 / 0.70, kNN 0.56 / 0.68 / 0.66, and Naive Bayes 0.54 / 0.72 / 0.62. The team chose SVM for consistency across datasets and interpretability, explicitly stepping away from gradient boosting such as XGBoost in this low-history setting.
• Trend features help: Adding moving-average-based features derived from recent payment delays and outcomes improved performance by 3–5 percentage points across datasets, and the system remained applicable with as few as two previous customer interactions.
• The modular cash flow model wins in the cold-start case: Predicting 11 months from only 1 month of history, the proposed approach reached 11.85% MAPE against 159.40% for Prophet and 166.24% for SVR. ARIMA could not be run at that horizon because one month of data was insufficient for AutoARIMA to determine the order.
• Performance across other horizons: For forecasting 3 months with 9 months of training, the method reached 13.06% MAPE versus 14.19% (SVR), 14.29% (Prophet) and 19.61% (ARIMA). For 6 months from 6 months, it reached 9.41% versus 19.11% (SVR), 19.01% (Prophet) and 10.72% (ARIMA). ARIMA was run on weekly aggregated data because it performed much better there than on daily data.
• F1-score checked but not tabulated: The paper reports that additional F1-score evaluation showed trends consistent with balanced accuracy, but the F1 numbers themselves are not given.
• Modularity matters operationally: Monolithic approaches fail when a data category is missing; the four-sub-module design kept producing useful predictions when only a subset of data sources was available.
• Business impact metrics are not reported: The paper describes deployment and integration into the Cluee platform but does not report adoption figures, revenue effects, or measured client outcomes.

Methodology in Plain English

The system has two connected parts.

The accounts receivable part is a binary classifier that decides whether an invoice will be paid within a 7-day grace period after the due date or will be significantly delayed. It combines basic invoice facts (amount, payment terms, customer ID) with historical relationship metrics derived from past interactions: payment ratios of late versus total payments, average delay times, counts of outstanding invoices, and monetary aggregations of past behavior. To catch shifts in a customer's habits, the model also uses moving averages of recent payment delays and outcomes over short and long windows. An SVM is trained separately for each business, so payment patterns specific to an industry or customer relationship are not averaged away.

The cash flow part is deliberately split into four sub-modules, each covering a different kind of money flow: hourly project work (computed from recorded work sessions, hourly rates and project timelines), non-recurring invoices and expenses (predicted conservatively by taking the maximum of mean income over the last 6 months and planned income, or their sum when only one is positive), flat-rate projects (payment tied to project completion dates), and recurring invoices and expenses (subscriptions and regular costs with known schedules). The four outputs are aggregated into a single forecast.

The two parts connect in two ways. The system can predict delays for individual invoices on its own, or the cash flow module can call the accounts receivable module so that expected delays adjust the timing assumptions in the forecast. Because real cash flow records were scarce, the authors also built a synthetic generator producing realistic project-based work patterns for 1,000 simulated users over one year, yielding 422,306 work sessions with variable hours, concurrent projects, and freelance-typical wage dynamics. The whole system runs as a REST API built with Flask and deployed on Google App Engine, returning both compact summaries and component-level details, including which invoices were flagged as risky and which sub-modules contributed most to a forecast.

Why This Matters

Research impact. Most financial prediction literature assumes enterprise-scale data. This paper documents what happens when those assumptions are removed, and shows that a simple, interpretable SVM plus a hand-designed modular forecasting scheme can beat ARIMA, Prophet and SVR in a genuine cold-start scenario. It reframes "deployability under sparse data" as a first-class research problem rather than a preprocessing detail, and it adds a vendor-collaboration case study to a literature dominated by offline benchmarks.

Real-world applications.

• Freelancers and micro-businesses deciding which invoices to chase first and how much cash to expect next month.
• Invoicing and smart-budgeting platforms that want to add risk flags and short-term liquidity projections to their existing tools.
• Early-stage startups with few repeat customers that need conservative planning signals rather than optimistic projections.
• Advisory and bookkeeping services that need explainable outputs they can show clients, listing flagged invoices and the cash-flow components driving a forecast.

Industry relevance. The architecture is intentionally cheap: a Flask REST API on Google App Engine, per-business models, and predictions delivered as lightweight decision-support signals rather than automated decisions. That matches the constraints of the SME software market, where infrastructure budgets are small and users demand to see the reasoning behind a number. The finding that moving-average trend features add 3–5 percentage points is a low-cost improvement any invoicing product could adopt.

Future Directions

• Broader validation across more diverse SME settings, since the current study rests on a small startup dataset of 297 invoices from 60 customers.
• Reducing reliance on synthetic data for cash flow evaluation by collecting real cash flow series from deployments.
• Deeper integration of receivable risk into forecasting, so payment-delay probabilities propagate through the models more richly rather than only adjusting timing assumptions.
• Extending the system as more deployment data accumulates, which the authors position as the route to improving on values such as the 0.56 balanced accuracy observed on the startup dataset.

Target Audience

Applied machine learning engineers and data scientists building production financial products for small businesses; researchers studying forecasting under data scarcity or cold-start conditions; and fintech product and engineering teams at invoicing, accounting or budgeting platforms who need to know which modeling choices survive contact with real SME data. Readers looking for new deep-learning architectures or for measured business outcomes such as revenue impact will not find them here; the value is in the deployment constraints, the architecture decisions, and the honest reporting of accuracy on a very small real dataset.

Authors’ abstract

Small and Medium Enterprises (SMEs), particularly freelancers and early-stage businesses, face unique financial management challenges due to limited resources, small customer bases, and constrained data availability. This paper presents the development and deployment of an integrated financial prediction system that combines accounts receivable prediction and cash flow forecasting specifically designed for SME operational constraints. Our system addresses the gap between enterprise-focused financial tools and the practical needs of freelancers and small businesses. The solution integrates two key components: a binary classification model for predicting invoice payment delays, and a multi-module cash flow forecasting model that handles incomplete and limited historical data. A prototype system has been implemented and deployed as a web application with integration into Cluee's platform, a startup providing financial management tools for freelancers, demonstrating practical feasibility for real-world SME financial management.