← Back to blog

October 6, 2026 · 4 min read

Funds in Fineract: the three decisions that decide whether the model survives an audit

NAV mechanics are the easy part. What separates an operable fund from an expensive prototype are three design decisions — where value comes from, where the investor position lives, and whether history can be rewritten — that almost always get deferred.

Investment funds · Auditability · Architecture

Modelling a NAV-per-unit fund on Apache Fineract is simpler than it looks: a subscription that converts at the previous close’s price dilutes nobody, and that rule takes an afternoon to implement. It is what everyone argues about, and it is the easy part.

What decides whether you have an operable fund or an expensive prototype are three decisions almost nobody makes up front, because none of the three stops the system from working in the demo. All three collect later.

Decision 1: is fund value keyed in or derived?

It is tempting to expose an endpoint that takes the bank balance, the placed portfolio and other assets, and adds them up. It works on day one, and it is what most implementations do.

The problem is that this is the number every participant’s wealth depends on. Keyed in, a typo distributes or destroys other people’s money with no reconciliation trail, and nobody catches it until the statement goes out.

There is a less obvious and more damaging consequence. The process that backfills days without a manual capture has nowhere to get a per-day value from, so it ends up repeating the same one across the whole window it reconstructs. The unit price then moves only with subscriptions and redemptions, never with the portfolio’s actual performance. Participants see a flat fund for ten days and a jump on the eleventh.

The lending side is derivable: Fineract knows how to compute aggregate outstanding principal as of a date. The treasury side comes out of the general ledger. Starting with manual capture is fair; leaving it as the permanent mechanism is not.

Decision 2: is the investor’s position derived or stored?

If each transaction’s units are stored unsigned, an investor’s holding can only be obtained by summing those rows and inferring the sign from the transaction type: deposits add, withdrawals subtract. It works as a query.

What it means is that no official figure exists for what each participant holds. Every report recomputes it, and it only takes two of them using different criteria — one includes reversed transactions, the other does not — to start an argument nobody can settle, because there is no source of truth to check against.

A unit balance per investor as a first-class entity — fund, account, units, average cost — makes the holding queryable and, above all, reconcilable against the sum of transactions. That contrast between the stored balance and the derived one is the best early warning you will get: when they stop matching, something broke this week rather than six months ago.

Decision 3: is history mutable or sealed?

This is the one that turns a code problem into an audit problem.

Recalculating a day is an operational necessity: a late transaction arrives, a balance is corrected, a failed process is repaired. The question is what happens to the units already allocated. If the recalculation rewrites them at whatever price is current at that moment, the history of holdings is mutable: the answer to “how many units was I allocated in February?” changes depending on when you ask.

A real fund needs sealed closes. Once a date’s price is published, an error is corrected with a dated adjusting entry, not by overwriting the original. It is the same discipline you already apply in accounting, where nobody would think of editing a three-month-old journal entry.

Building it in from the start costs a status column and a rule. Retrofitting it when the regulator asks for the audit trail costs a project.

The case of its own: fees

Storing the fixed fee and the variable fee is not the same as applying them. It is common to find both persisted on the daily close and the unit price computed without them: the result is identical with a zero fee or with a fee of ten thousand.

That is not a programming error, it is a product decision nobody made. Does it come off fund value before dividing? Does it accrue on average net assets over the period? Is the variable component a success fee over a hurdle? Until it is written down, the module does not reflect the contract the participant signed.

Before you take third-party money

Three invariants in automated tests, none of them optional when software distributes other people’s wealth:

  1. At every close, price times units outstanding must equal fund value.
  2. The sum of all investors’ units must equal units outstanding.
  3. A subscription or a redemption cannot move the unit price.

On the first one, a precision warning: Fineract computes with nineteen significant digits, while monetary columns are typically declared with six decimals. Recomputing that control figure from values read back out of the database will not reproduce the stored value exactly, because the price was persisted rounded. For monitoring that is acceptable; to reconcile to the cent you have to decide where rounding happens and write it down before the auditor asks.