Projects

A Thousand Years of Weather

This started with a complaint about thermometers. The longest continuous temperature record in the world begins in 1659, in central England. Most places have nothing usable before the twentieth century. If you want to know whether the last few decades are unusual, a couple of hundred years of measurement is a thin thing to lean on, because it is barely long enough to establish what usual looks like.

So I went looking for older observations. Not measurements, since nobody was measuring, but things people wrote down every year for reasons of their own. Which cultures kept continuous written records for longer than anyone has owned a thermometer, and what did they happen to record?

There turned out to be more than I expected. Court officials in Kyoto noted the day the cherry blossom reached full bloom, and someone has done so most years since 812. Town clerks in Burgundy recorded when the grape harvest was declared open. Priests at Lake Suwa logged whether the ice ridge formed across the water each winter, which their tradition holds is a god walking to visit another. Merchants at the mouth of the Torne River noted the spring breakup, because it reopened the river to trade.

None of them cared about the climate. They cared about blossom parties, wine, a god, and shipping. Which is exactly what makes the records worth looking at, since nobody collecting them had a stake in what they would later be used to argue.

Does any of this show anything, in either direction?

What follows is what the arithmetic returns when you ask. That includes a record that gives opposite answers depending on how far back you look, two records that move against each other, and the places where the six simply disagree.

The records

Four of these respond to temperature indirectly, through a plant or a body of water, which makes them proxies rather than measurements. The other two are the thermometer series, included so there is something to check the proxies against over the period where they overlap.

How far back each one reaches

The shaded band marks the years where all six overlap, which is the only period that allows a strictly like-for-like comparison.

What the records show

The same test, run twice on each record. Once over its whole length, and once over the period since 1900. Both are shown because they do not agree, and deciding afterwards which to lead with would be choosing the result first.

How this was measured

Each record is tested on its own, using the median of the slopes between every pair of observations. That estimator ignores outliers and copes with the gaps in the older records, which matters when a series is assembled from diaries.

Significance comes from a rank-based test, then two corrections are applied. The first accounts for warm years tending to follow warm years, which otherwise makes almost any climate series look more certain than it is. The second accounts for testing six records at once. Both make the result harder to claim, not easier.

The choice of period changes the answer, so both are published. An earlier draft of this page led with the figures since 1900, which happened to be the window where all six records agree. That is a bad way to choose a window, and the two now sit side by side with neither marked as the real one.

Record by record

Each series on its own terms, in its own units.

Put them together

The six records measure entirely different things, so to compare them each is expressed as how far it sits from its own 1700 to 1899 average, then flipped where needed so that up always means warmer. Each line is smoothed over 30 years.

Click a record to remove it from the chart.

The average of whatever was being recorded

Drawn faintly where fewer than three records contribute. This line is a visual aid and nothing more. The number of contributing records changes over time, so fitting a trend to it would measure records arriving and departing as much as the climate. The verdict above never uses it.

Do they agree with each other?

    Correlation between each pair over their shared years, smoothed to 30 years. These are different places with different weather, so perfect agreement would be more suspicious than partial agreement.

    When did it start?

    The clearest single shift in each record, found by testing every possible split point and keeping the one that divides the series most sharply.

      Rate of change over a moving 50-year window

      Above the centre line means that window was warming. Because days and degrees cannot share an axis, each record is measured against its own variability, so the vertical scale is how many standard deviations of that record's normal fluctuation it shifted per century. This shows not just whether the records changed, but when the change accelerated.

      Was it ever like this before?

      A trend on its own says very little. The question worth asking is whether any of this sits outside what these records have already done, and that splits into four tests. They do not all agree, and the one that disagrees is the one usually quoted.

      Where each record currently stands

      What doesn't match

      These records disagree in places, and the disagreements are more informative than the agreements.

      • Burgundy gives opposite answers depending on the window. Across its full 634 years the harvest drifts significantly later, which is a cooling signal. Measured from 1900 it moves earlier faster than any other record here. Both figures come from the same data and the same test. The record begins in a warm medieval period and reaches its latest harvests around 1800, so a single line across the whole thing is averaging two opposite regimes. That is an argument for distrusting any one number from this record, not for preferring the one that fits.
      • Burgundy and Central England move against each other. Over their 344 shared years the correlation is negative. Two European records a few hundred miles apart, both responding to growing-season warmth, disagreeing about the direction of travel. No explanation is offered here because none was tested.
      • Kyoto's long record is nearly flat. Over twelve centuries the blossom date barely moves, and the trend across the full series is small enough that it would not be worth remarking on by itself. Almost all of the movement is in the last hundred years.
      • Kyoto and Burgundy barely correlate at all. The two longest proxy records here, both flowering and fruiting dates, share 583 years and show almost no relationship. If both were clean thermometers this would be hard to explain. They are not thermometers.
      • Two records stop before the present. Torne River ends in 2013 and Stockholm's homogenised series in 2017. Recent years exist elsewhere but splicing them on would mix differently processed data, so they are left out.
      • Kyoto and Lake Suwa are both in Japan. They are not fully independent of one another, though they measure opposite seasons through entirely different physical processes.

      What this adds up to

      What this can and cannot show

      • A proxy is not a thermometer. Blossom and harvest dates respond to rainfall, to the choice of cultivar, to pruning and planting practice, and to how the land around them has changed. Temperature is one influence among several.
      • Regional is not global. Kyoto describes Kyoto. Central England describes central England. Nothing here measures the planet. What gives the result its weight is six separate places pointing the same way.
      • Cities warm for more than one reason. Kyoto and Stockholm have grown enormously over these records. Urban warming is part of what the later observations capture, and it is not separated out here.
      • Older observations are rougher. Early entries come from diaries and chronicles, under changing calendars and changing habits of observation. They are treated the same as modern entries, which flatters neither.
      • Lake Suwa is handled differently. Its headline result is not the freeze date but whether the lake froze at all, tested with logistic regression. Using dates alone would quietly discard the warmest winters, the ones with no ice.
      • Everything is computed before the page loads. The analysis runs in a build script with its own test suite, so the figures here are fixed and reproducible rather than recalculated in your browser.

      Sources

      Every dataset here is public. The raw files are archived alongside the code, because one of these sources went offline partway through building this. The Kyoto record survived twelve centuries of hand copying and then nearly did not survive a university closing a website.