September 26, 2026 · About the author · Synthetic learning example

How should you round money in Python?

Read decimal amounts from their original text into Decimal, choose an explicit rounding rule, and round at the stage required by your reporting contract. Rounding each line and then adding can give a different result from adding the unrounded lines and rounding once. Neither rule can be chosen from the number alone.

Consider a US subscription report with three usage charges of USD 0.335 before rounding. Fractional cents can appear in intermediate usage calculations even when the final bill is shown in cents. These amounts are fictional; this is a programming exercise, not a statement about any provider’s billing or tax rules.

Run the rounding comparison

This example uses only Python’s standard library. The assertions check both the stage of rounding and the tie-breaking rule.

from decimal import Decimal, ROUND_HALF_UP, ROUND_HALF_EVEN
cent = Decimal("0.01")
amounts = [Decimal("0.335")] * 3
line_total = sum(
    (x.quantize(cent, rounding=ROUND_HALF_UP) for x in amounts),
    Decimal("0")
)
invoice_total = sum(amounts, Decimal("0")).quantize(
    cent, rounding=ROUND_HALF_UP
)
assert line_total == Decimal("1.02")
assert invoice_total == Decimal("1.01")
assert Decimal("0.345").quantize(cent, rounding=ROUND_HALF_UP) == Decimal("0.35")
assert Decimal("0.345").quantize(cent, rounding=ROUND_HALF_EVEN) == Decimal("0.34")
print(line_total, invoice_total)

Why are the totals $1.02 and $1.01?

With the chosen half-up rule, each 0.335 line becomes 0.34. Three rounded lines add to 1.02. Summing the original values gives 1.005, which becomes 1.01 when rounded once. The one-cent difference is explained by the rounding stage; it does not establish a missing payment.

Do not silently change the rule until your report matches a desired total. Ask whether the source system rounds per item, per usage event, per subtotal or only at the invoice boundary. Store that decision with the metric definition. If reconciliation requires matching a ledger, compare its documented calculation sequence as well as its final display precision.

What does the rounding mode change?

The second test uses 0.345. Half-up produces 0.35; half-even produces 0.34 because the retained hundredths digit is even. Both are defined numerical rules. Python’s Decimal documentation describes these modes and quantize, which applies a chosen exponent. Python Decimal reference.

A label such as “round to two decimals” is incomplete when exact halfway values exist. Name the tie rule and test refunds or other negative amounts too. Do not assume that a rule used by one US, UK or European system applies to another.

Why construct Decimal from text?

A binary floating-point value may already be an approximation of the source decimal. Constructing Decimal from that float preserves the float’s actual value, not necessarily the original decimal text. If the source is a CSV field, keep that field as text until decimal parsing. Converting a previously calculated float to a string is not a reliable way to recover precision that was lost earlier.

Decimal arithmetic still uses a precision context for operations. The defaults are sufficient for this small example, but large amounts, many operations or division require a deliberate precision policy. Decimal does not remove the need to define rounding, validate input or check totals.

Can you store integer cents instead?

Integer cents are useful when all stored amounts are already whole USD cents. They do not directly represent the fractional-cent inputs in this example. One option is an agreed finer integer unit followed by a specified final rounding step. Another is decimal arithmetic. Choose a representation based on the source precision, rather than truncating 0.335 to 33 cents.

Keep the currency code with every amount. This example uses USD only. Adding a GBP or EUR amount to a USD amount because all three look numeric is a separate error; rounding cannot fix it. Currency conversion requires its own rate source, date and conversion policy.

What should a reconciliation check record?

• The original amount and currency, before formatting.
• The aggregation grain and rounding stage.
• The rounding mode and output precision.
• The source-system total and the explained difference.

For this fixture, the expected difference is exactly USD 0.01. That is a test case, not permission to ignore every one-cent discrepancy in production. Investigate differences against the agreed rule before defining any tolerance.

Explain it in an interview

“I would first check whether the system rounds each line or the final total. In this example those approaches give $1.02 and $1.01. I would preserve the original decimal inputs, name the rounding rule, and reconcile at the same grain as the source.”

Continue with orders and payments reconciliation or reading a sales CSV in Python. These exercises help separate numerical representation from missing records and incorrect business filters.