Four Columns for the Ferry Forecast

A ferry company forecasts each day's passenger count on the evening before, a horizon of 1 day. Its training table has one row per day, and the row for day $t$ has day $t$'s passengers as its target. The data covers two years. The model trains on the first 21 months and is tested on the last 3, after a cut-off date.

An engineer proposes four new columns for the row whose target is day $t$:

• month_share: day $t-1$'s passengers divided by the average daily passengers of the calendar month that day $t-1$ falls in.
• smooth_5: the average passengers over the 5 days centred on day $t-1$.
• lag1_filled: day $t-1$'s passengers. The ticket gates sometimes fail to report a day, and a missing day is filled in with the average of the days on either side of it.
• days_to_regatta: the number of days from day $t$ until the next harbour regatta. The town publishes the regatta dates a year in advance.

All four are easy to compute from the two-year file. Exactly one of them could be filled in honestly on the evening before day $t$, in training, in testing and after launch. Which one?