← Back to Simulations

Kyoto Cherry Blossoms: Bloom Date Trend

For more than 1,200 years, people in Kyoto, Japan have written down the day the cherry trees reach full bloom — in court diaries, chronicles and poems, and lately in newspapers. It is one of the longest biological records in the world. Is the modern record really drifting earlier, and how sure can we be?

Data: the day each year the wild cherry Prunus jamasakura reached full bloom in Kyoto, Japan, expressed as day-of-year (812–2026; 838 years have a recorded date, with many gaps in the early centuries). Early centuries come from diaries and poems, modern years from newspaper reports of the trees at Arashiyama. Source: Yasuyuki Aono and Genki Katata, Cherry Blossom Full Bloom Dates in Kyoto, Japan (Aono & Kazui 2008; Aono & Saito 2010), as republished by Our World in Data (CC BY 4.0). The 2026 date is Katata's estimate.

Why start at 1900?

Here is the whole record. Over twelve centuries the bloom date wanders for generations at a time — cold spells push it later, warm spells earlier — so a single straight line through everything (the dashed gray line) is almost perfectly flat. The steady slide toward earlier blooms is a modern feature. So everything below uses 1900 onward: 124 years, a similar length to the Nenana record.

The whole record, 812–2026

Context, not a test

Gray dots: years before 1900. Colored dots: 1900 onward. Navy curve: roughly 50-year running average. Dashed gray line: one straight line through all the years.

A caution about the window. We chose 1900 as a round number at the start of the modern-warming era — but we did look at the plot first, and the slope depends on where you start (see the numbers under the chart). Picking a start year because it makes the trend look clearest would make the tests below look more convincing than they deserve, so treat the choice of window as part of the analysis, not a free pass.

Fitting a trend line

Each point below is one year's full-bloom date (1900–2026 is 127 years, but 1919, 1921 and 1945 have no recorded date, leaving 124 points), expressed as a day-of-year number so it can be regressed like any other measurement. The blue line is the ordinary-least-squares (OLS) fit; the red line is the null hypothesis — no trend at all, just the flat average bloom date. The fitted slope is about −0.096 days per year — roughly ten days earlier per century — and this time the points hug the line fairly tightly.

Full-bloom date vs. year

OLS fit vs. null

Testing the trend: shuffle the data

The idea. If year and bloom date were really unrelated, then it shouldn't matter which date we attach to which year — the pairing is arbitrary noise. So: keep the 124 years exactly where they are, and randomly shuffle the bloom dates among them. Refit the line. Do that thousands of times, and you get the distribution of slopes you'd see from pure chance alone.

Shuffle test (permutation NHST)

Years fixed, bloom dates shuffled

Each dot keeps its true value and color, and slides sideways to sit at a randomly chosen year — then the resulting slope joins the null distribution below.

How sure are we of the slope? Resampling

The idea. This time we don't scramble anything — each (year, bloom day) pair stays intact. Instead we ask: if we'd happened to sample a slightly different 124 years (by resampling with replacement from the ones we have), how much would our fitted slope move around? That spread is our uncertainty about the true slope.

Resampling-based confidence interval

Pairs resampled with replacement

Each pick is its own small dot, offset sideways from its year — a value drawn twice shows up as two dots side by side. Years left out of this resample appear as faint dashed outlines. Every dot keeps its true color and its real year–bloom pairing.

Same machinery, different question

Both methods refit a regression line thousands of times and look at how the slope varies. It's tempting to think of them as the same thing — they aren't. The procedure is what tells them apart:

Shuffle test (NHST)
Breaks the year↔bloom-date pairing on purpose.
Asks: "if there were truly no trend, how often would chance alone produce a slope this big?"
Answer is a single number — the p-value.
Resampling CI
Preserves every year↔bloom-date pairing.
Asks: "given the trend we found, how precisely have we pinned down its slope?"
Answer is a range — the confidence interval.
In short: shuffling destroys the relationship to build a null distribution for testing whether it exists at all. Resampling keeps the relationship intact to build a sampling distribution for measuring how precisely we've estimated it. Here, with 124 years and a fairly tight fit, no shuffle comes anywhere near our slope and the confidence interval stays well below zero — but remember the window: this answers "is there a trend since 1900?", not "has the bloom date always trended earlier?" (the top chart says it hasn't).