Roughly, yes. “May level minus 40 points” predicted the autumn low with a mean error of 9.3 points over 18 out-of-sample years — 43% less than assuming an average year (16.4). Press play to watch 23 years of the city's ten reservoirs.
Drag the strip or use ← → (Shift = a year), Space to play. Click a tank or a row to follow one dam.
Bars start at zero; dashed line = 20%. “vs norm” compares with that dam's average for the same calendar month, 2001–2023.
Each line is one year, from the week of 1 May to the end of February. The median drop is points ( million m³); in the middle half of years it is .
Blue: the median year. Red: the five seasons that fell below 20%. Thick line: the season shown on the map, drawn up to the current month. Hover a line; click it to see that autumn on the map.
May level and the autumn low move together (r = ). “May level minus the median drop” is the rule tested below.
The calculator applies the historical rule; it is not a forecast for any real year (the portal's data end in February 2024). At low May levels the rule ran low: in 2008, 2014 and 2023 the drop was only 15–33 points.
Rolling origin: to predict year Y, only years before Y were used. The May rule missed by points on average; assuming an average year missed by . A fitted regression did no better ().
● actual autumn low · ○ forecast · the line between them is the error. Biggest misses were dry years the rule was too optimistic about.
Share of months below 20% full, 2001–2023. Pabuçdere was there in of months and Kazandere in ; for Terkos the figure is and for Ömerli .
Deeper blue = fuller; red = below 20%. Size alone does not explain it: the smallest, Istrancalar, was below 20% in only of months. The vertical line is the month on the map; click a cell to move there.
| # | What was wrong | Fix |
|---|---|---|
| 1 | 130 rows (1 Apr – 8 Aug 2023) stored as percentages (60.65) while all others are fractions (0.6065) | ÷ 100 |
| 2 | Same block: days 1–12 have day and month swapped (“2023-01-04” is 1 April) | swap back where it restores a continuous daily sequence |
| 3 | 25 Dec 2007: one-day spike in every dam (city 59.6% between 25.6% and 25.7%; ISKI's own series says 25.7%) | mean of neighbouring days |
| 4 | City-wide series: 2 duplicated dates, 2 missing days | keep first; leave gaps |
After cleaning, the 10 dams weighted by capacity reproduce ISKI's published city-wide figure within points on average over months.