Custom financial predictive analytics
Appfront builds predictive models on financial data: liquidity, revenue, margin and debtor risk. Fed from your accounting and bank transactions, and set up so the outcome can be defended to management and the auditor. A model nobody can explain doesn't get used, however well it scores on paper.
Forecasting on figures that must be accounted for
Financial predictive analytics applies statistics and machine learning to the figures already in your accounts: how much money will be in the bank in eight weeks, which invoices will arrive late, how revenue is tracking against budget. If you're looking for predictive analysis in the broader sense, such as churn or customer attrition, see the page on the predictive analytics dashboard. This is the financial vertical.
A machine maintenance forecast doesn't need defending to an auditor. A cash flow forecast on which management postpones an investment, or on which the going-concern assumption in the annual accounts rests, does. That brings requirements that don't apply elsewhere: reconciliation with the general ledger, every deviation traceable to an assumption, the previous version reproducible. And the controller must be able to say in one sentence why the line drops in October.
Liquidity based on payment behaviour
Calculates with what each debtor actually does, not with the due date on the invoice.
Forecast alongside the budget
The budget remains the agreement, the rolling forecast the expectation. Two separate objects, with the difference in view.
Debtor risk with a next step
For each invoice, an expected payment date and a reason: call, request the order number or discuss the limit.
Where financial forecasts fall apart in practice
The open items list is not a forecast
If you forecast cash flow by placing open items on their due date, you're repeating what's on the invoices. The predictive value lies in payment behaviour per debtor. Client A structurally pays eleven days late. Client B only pays after the first reminder. Client C runs payment batches on the fifteenth and the last working day, so an invoice that arrives on the sixteenth sits idle for two weeks. An aggregated DSO hides this. You need the reconciled payments per debtor; clients with too few invoices go into a segment.
More important is the distinction between two kinds of late. Someone who can't pay is a credit risk. An invoice that sits there because the purchase order number is missing and so never enters the approval flow is a process error on your side. Statistically they look alike, but the action required is completely different.
A single one-off order makes the model structurally too optimistic
Almost every history contains an exceptionally large order, an insurance payout or a sold asset. If you leave it in the training data as ordinary revenue, it lifts the baseline and the model over-forecasts for years. If you strip it out, the series no longer reconciles with the annual accounts. The workable route is to leave the item in and have the controller who knows it flag it separately as one-off.
Nobody needs to teach a model calendar effects: varying working days per month, holiday pay in May, quarterly VAT payments. These belong in as a deterministic component, which makes the learning part smaller and easier to explain. Watch out for breaks too: an acquisition, a new pricing model or a migration to a different accounting package makes the period before it partly unusable.
Forecast and budget don't belong in the same field
The budget is a commitment: approved, frozen, and the basis on which people are held to account. The rolling forecast is an expectation that by definition changes. As soon as a forecast is read as a revised budget, people start managing it instead of reporting it.
We keep three layers separate, each with its own owner: a budget that is fixed for the financial year, a monthly rolling forecast in which every version is retained, and beneath that a thirteen-week cash flow forecast using the direct method, built from actual bank transactions, receivables and payables.
In finance, explainability matters more than accuracy
This is the design choice we make explicit: for financial models we choose the most explainable variant, even if a more opaque model predicts marginally better. A gradient boosting model that hits the cash flow slightly more accurately but cannot be justified to management, the auditor or the bank is not used. A model that actually gets used delivers more than a better model that stays on the shelf.
In practice that means an additive structure: baseline, seasonality, known items, deterministic calendar items and expected payment behaviour per debtor. If the line drops in October, you can point to which component is falling and by how many euros. Where a learning model does earn its place, for example in predicting the likelihood that an invoice is paid late, we show the most heavily weighted factors using SHAP values, with the caveat that this explains the model, not reality.
We also deliver a range with scenarios instead of a single number, and we store every forecast version frozen. Without those versions, you can never measure whether the model is improving.
A prediction is only worth something if it triggers an action
The first design question is not what is technically feasible, but which decision turns out differently if the prediction is right. If you flag everything, you flag nothing: every signal has a threshold and an owner.
Targeted reminders
The work list is sorted by expected loss times amount, not by age, with the reason alongside. Call six customers instead of sending thirty standard emails.
Adjusting a credit limit
Internal payment behaviour often says something sooner than an external score. A credit insurer lowering its limit on a customer is itself a signal.
Postponing or going ahead with an expense
We test an investment against the cautious scenario: not whether it fits on average, but whether it survives the tightest week.
Headroom under bank covenants
If ratios are tested quarterly, the model factors them in, so you can see when headroom is getting tight.
Sources, integrations and one legal pitfall
For the accounting side we work with the APIs of Exact Online, AFAS, Twinfield and Visma: general ledger, open items and above all the reconciliations, where the payment behaviour lives. We read bank transactions via CAMT.053, the ISO 20022 format that replaces MT940 and includes structured payment references. Don't build any new processing on MT940: Rabobank is discontinuing the MT formats on 15 November 2026.
One thing up front, because it regularly goes wrong: retrieving account information live via PSD2 is a licensed service supervised by De Nederlandsche Bank. We do not build our own integration for this, but work through a licence holder. If the forecast ends up in an application your customers use, this touches on our fintech app development.
Test your idea first: a working prototype in 1 day
With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.
View OneDayBuild →GDPR, the AI Act and your auditor
If you only assess legal entities, the landscape is straightforward. If your debtor ledger includes sole traders, freelancers or private individuals, that changes things: the payment behaviour of a natural person is personal data, and an automated decision with legal effect requires human intervention under the GDPR. For that same group, Annex III of the AI Act classifies creditworthiness assessment as high risk, with fraud detection as an exception.
A third framework is often forgotten. If the model estimates a probability of default per debtor, that outcome affects the provision for doubtful debts. The expected credit loss approach from IFRS 9 looks at the entire debtor portfolio and not only at what is already overdue; the Dutch Accounting Standards Board has also permitted it under the Dutch Accounting Standards. Your auditor will then want to know whether that outcome is reproducible.
- Distinguish between legal entities and natural persons in the design
- Human intervention for decisions with legal effect
- Reproducible forecast versions for the audit
When custom isn't the answer
A mature market of off-the-shelf packages exists for financial forecasting, and for many readers a package is the better choice. Visionplanner is widely used by Dutch accountancy firms, integrates with Exact Online and AFAS, and delivers a forecast your accountant recognises immediately. Agicap focuses on cash flow, and for heavier planning there are Jedox, Pigment, Anaplan and Prophix. A package like that doesn't leave you with model maintenance.
Sometimes we give an even more honest answer: don't start with a model. If invoices are booked late or reconciliation lags behind, no algorithm will help. Get your bookkeeping in order first.
Custom becomes sensible when your revenue model doesn't fit the standard mould: subscriptions with variable usage, staged invoicing on long-running projects, consignment, or a platform model in which receipt and payout diverge. The same applies to sources that no package supports, such as claims flows in healthcare and education, or when the forecast needs to land in your own product.
That comes with the burden that custom brings. You become the owner of a model that ages: the customer base shifts, payment terms change, a major customer leaves. Someone has to measure the drift and retrain the model. With a package, the vendor does that.
- Choose a package for a standard bookkeeping setup and a common revenue model
- Choose a package if you have no in-house capacity to maintain models
- Fix your bookkeeping first if postings and reconciliation are behind
- Custom for subscription, staged-invoicing, consignment or platform models
- Assign model maintenance before you start, not afterwards
Frequently asked questions
Technically the difference is small; in terms of accountability it is large. A general dashboard predicts churn, demand or customer attrition, and nobody has to defend that to an auditor. A financial forecast does: as soon as management postpones an investment because of it, every euro must be traceable to an assumption.
For many organisations that is the right choice, and when it is, we say so. Visionplanner integrates with Exact Online and AFAS, Agicap focuses on cash flow, and Jedox, Pigment, Anaplan and Prophix cover heavier planning. Custom development only becomes logical when you have an unusual revenue model, sources for which no integration exists, or when the forecast has to land inside your own product.
By building the model so that the question can be answered. The forecast is the sum of recognisable components: baseline level, seasonality, known items, calendar items and expected payment behaviour per debtor. If the forecast drops in October, you can point to which component is falling. For a learning model, we show the most heavily weighted factors using SHAP values, as an explanation of the model and not of reality.
That depends on who your debtor is. Annex III of the AI Act classes the assessment of creditworthiness of natural persons as high risk, with fraud detection as an exception. If you only assess legal entities, your model falls outside it. For sole traders or private individuals it does apply, with requirements for logging, transparency and human oversight; at the time of writing, the application date has been postponed to 2 December 2027.
You become the owner of the source code and the model specification, with documentation stating which assumption sits in each component. The honest caveat is that owning code is not the same as being able to maintain it; a model you own requires someone who understands the assumptions and reviews them periodically. Assign that role before you start.
That happens, and it is better to measure it than to wait and see. We record every forecast version as it stood at the time and later compare it with the actuals. We report the average error and the systematic bias. For cash flow series with weeks without receipts, we use WAPE or MASE, because MAPE breaks down at low or zero values.
Related services
Predictive analytics dashboard
Predictive analytics outside the financial domain: churn, demand, customer attrition and capacity. See predictive analytics dashboard.
Fintech app development
If the forecast lands in an application your customers work with, additional requirements apply. See fintech app development.
Discuss your forecasting challenge
Tell us which decision you want to be able to make better and which systems your figures come from. We will look at whether your data can support the forecast and whether an off-the-shelf package would help you faster. If it would, we will say so.