Gaps in the calendar
The ECB fixes rates on TARGET settlement days only, so most of the calendar is empty. Filling the gaps is a modelling decision, and there are two places where it must refuse.
The 30% of days that have no rate
Of 10,155 calendar days since the series began, 7,121 carry a fixing. The rest are weekends, and 134 weekdays that are not.
Two of the seven rows in each year are structurally empty — the ECB does not fix at a weekend. The interesting part is the handful of pale squares punched out of the weekday block: 1 January, Good Friday, Easter Monday, 1 May, and 25 and 26 December. Seventeen days over the three years.
Read the Easter pair across the years and it moves: 7 and 10 April in 2023, 29 March and 1 April in 2024, 18 and 21 April in 2025. That is why this project carries no holiday calendar. No weekday rule predicts those dates, the only rule that does is the Gregorian computus, so they are observed as absences in the data instead of being asserted from a list somebody would have to maintain forever. (The series has one closure no rule of any kind would give you: 1999-12-31, taken for the millennium changeover.)
So what. A transaction dated on a Sunday still has to be converted, and every option here is a modelling decision, not a lookup. Interpolating between Friday and Monday invents a rate nobody could have dealt at, and it needs the future to compute the past. Leaving the rate null pushes the same decision into every downstream query, to be answered differently each time. So the daily table carries the last fixing forward, which is what a finance system does, and records the date of the fixing it used on every row, so you can always see which rate you are quoting.
Two ways carrying forward goes wrong
One: outside a currency's lifetime there is nothing to carry. The ECB's panel changes over time.
Of the 44 codes the series has ever quoted, 29 are still live and 15 stopped.
Of those, 10 stopped on the last business day before their country adopted the euro: the Greek drachma in 2000, the Croatian kuna in 2022, the Bulgarian lev at the end of 2025. None stopped at a redenomination: the ECB quotes the Turkish lira and the Romanian leu back to 1999 in their post-2005 units, so their 2005 changes of code never reach the series. The remaining 5 simply ceased, and this project does not guess at why.
So the dense series is built per currency between its first and last fixing, and a euro-era drachma never gets invented.
Two: a suspended quote is not a long weekend. The longest closure in the whole series is five days, so the carry-forward is capped at seven. That fills every weekend and holiday while refusing exactly two gaps, both of which are currency crises rather than calendars. The Icelandic króna has no reference rate for 3,347 days between the 2008 banking collapse and February 2018, and the Argentine peso none for 34 days after the January 2002 breaking of the dollar peg.
Those 3,365 rows exist with a null rate and a flag saying so, giving an absence you can count instead of nine years of a rate that had stopped being real.