D3M · Language of Data · Time series
Twenty-six years of state home values contain a crash, a boom, and a split. Which one you see depends almost entirely on where you start the line.
Every introductory time-series lesson uses a chart with one line on it. Real business data almost never looks like that. It looks like this: dozens of series that share a clock, move together most of the time, and disagree exactly when it matters.
The dataset here is the Zillow Home Value Index for the fifty states and the District of Columbia, monthly, from January 2000 through June 2026. ZHVI is a smoothed estimate of the typical home value in a region — roughly the middle of the market, not the average sale price, and not an index of what any particular house did. It is the closest thing to a common yardstick across places whose housing stock has nothing else in common.
Three events show up in almost every state's line, and each one lands differently: the run-up and collapse of 2004–2012, the pandemic surge that began in March 2020, and the divergence that opened after mid-2022 and is still open. This piece walks through all three. The controls in each figure are the point — the argument is easier to make by moving something than by reading about it.
The file as downloaded is wide: fifty-one rows, one per state, and 318 columns of dates marching off to the right. That layout is built for reading, and it is the layout a spreadsheet naturally produces. It is also the layout that no charting tool wants.
Wide data hides the variable. There is no column called "month" and no column called "value" — the month is encoded in the column name, and the value is scattered across 318 different fields. To plot value against month, you first have to make both of them into actual variables. That operation is a pivot to long form, and it is the single most common first step in real time-series work.
| RegionName | 2000-01-31 | 2000-02-29 | … | 2026-06-30 |
|---|---|---|---|---|
| California | 190,979 | 192,296 | … | 775,549 |
| Texas | 115,224 | 115,588 | … | 302,999 |
| Florida | 109,595 | 110,131 | … | 378,126 |
| … | … | … | … | … |
| RegionName | Month | ZHVI |
|---|---|---|
| California | 2000-01 | 190,979 |
| California | 2000-02 | 192,296 |
| … | … | … |
| Texas | 2000-01 | 115,224 |
Long form makes the grain impossible to misread: one row is one state in one month. Once the grain is stated out loud, the questions the data can answer become obvious, and so do the ones it cannot. This file can tell you how Nevada compared to Ohio in 2009. It cannot tell you anything about a house, a buyer, or a transaction — those rows do not exist here.
Here is the honest first chart: all fifty-one series, in dollars, unadjusted for anything. Hover to identify a line; click to pin it. Pinned states stay highlighted in every figure below.
The first thing this chart shows is dispersion in levels, and it shows it so loudly that it drowns out everything else. In June 2026 the typical Hawaii home is worth $836,741 and the typical West Virginia home $182,704 — a 4.6× spread. The expensive states are drawn with tall lines and the cheap states with flat ones, and the eye reads amplitude as importance.
But amplitude here is mostly just level. A 10% move in Hawaii is an $84,000 line segment; the same 10% move in West Virginia is an $18,000 segment, visually invisible. On a linear dollar axis you cannot compare rates of change across series with different starting points — you can only compare dollars, which is rarely the business question.
Switching the axis to log is the first real fix. On a log scale, equal percentage changes become equal vertical distances, so parallel lines mean equal growth rates regardless of level. Try it: the flat-looking cheap states stop looking flat.
The other fix is to throw away the levels entirely and plot each series as percent change from a common starting month. That is an index chart, and it is the workhorse of comparative time-series analysis: it puts every series through a shared zero and asks only about growth.
It also has a property most people discover by accident. The base month is a choice, and the choice is not neutral.
Drag anywhere on the chart — or use the preset buttons, or focus the plot and press the arrow keys — to move the black anchor. Every line is re-expressed as percent change from the month you land on. The panel on the right re-ranks the states live, and reports how much that ranking agrees with the ranking you get from a January 2000 base.
Run the presets in order and watch the ranking panel. From a January 2000 base, the leaders are Hawaii (+327%), California (+305%) and Maine (+282%): a coastal, supply-constrained story. From a January 2012 base — the bottom of the bust — the leaders are Nevada (+239%), Idaho (+225%) and Arizona (+204%): a story about how far there was to climb back. From a March 2020 base, the leaders are New Hampshire, Connecticut and Montana, and Nevada has fallen to 37th.
The rank-correlation readout puts a number on this. Against the January 2000 ordering, a January 2012 base gives Spearman's ρ of +0.34; a March 2020 base gives +0.09, indistinguishable from picking the order at random; and a June 2022 base gives −0.17. By that last one the leaderboard has not merely decoupled from the long-run one — it has begun to invert. All four charts describe the same fifty-one states over overlapping windows.
Nevada is the cleanest case. It is first from a 2012 base and 23rd from a 2000 base, because its 2012 value was artificially low: it had just finished falling 62%. Rebasing to a trough converts recovery into growth and rewards whoever fell furthest. Rebasing to a peak does the reverse.
An index chart answers "growth since when?", never "growth". If the base date is not visible in the chart, the chart is not finished. And a base chosen at a local extreme — anyone's peak, anyone's trough — is a claim, not a convention.
One more thing worth noticing in this figure: at a January 2000 base, four series are missing. Zillow's coverage of New Mexico and Wyoming starts in April 2002, Montana in February 2005, and North Dakota not until January 2009. They are dropped from the chart and counted in the readout, because there is no honest way to index a series against a month it does not have.
The tempting move is to backfill them — carry the first value backwards, or interpolate from neighbours — so the chart looks complete. Doing so would put North Dakota on a chart claiming to show change since 2000 using data that begins nine years later. The missing lines are the more truthful chart.
The 2008 collapse is usually narrated as a national event. In the state data it is emphatically regional. The median state's peak-to-trough decline was −15.5%. Nevada's was −62.0%. Alaska's was −1.5%. Ten states fell more than 30%; fifteen fell less than 10%. Averaging those together produces a number that describes nowhere.
Depth is the number everyone quotes. Duration is the one that mattered to the people who lived in it. A useful transformation makes both visible at once: for each month, express the index as a percent below its own running maximum. The series sits at zero whenever it is at a record high and traces out a valley the rest of the time. The width of the valley is how long a homeowner in that state was underwater.
The valleys are wide. The median state took 101.5 months — eight and a half years — to make a new record high after its 2006–08 peak. Nevada took 176 and did not get there until March 2021, meaning its recovery from the housing bust was in the end delivered by the pandemic boom.
The longest wait of all, though, belongs to Connecticut: 177 months, from September 2006 to June 2021. Connecticut fell only −26.1%, less than half as far as Nevada — and still took a month longer to climb out. A shallow hole with gentle sides can hold you longer than a deep one you scramble out of. Depth and duration are separate facts, and only one of them is in the headline number.
It was not alone. Eight states — Arizona, Connecticut, Delaware, Florida, Illinois, Maryland, New Jersey and Nevada — were still below their pre-2008 peak when Covid arrived in March 2020. For a homeowner in those states, "the housing market recovered" was not a statement about 2013 or 2016. It was a statement about 2021.
Note also the small dip at the right edge of many lines: the post-2022 correction registers here as a fresh, shallow drawdown from a peak set in 2022 rather than 2006. The underwater view treats every peak the same way, which is exactly what makes it comparable across eras.
The pandemic surge was faster than the bubble and much more uniform in direction, if not in size. Every one of the fifty-one series rose between March 2020 and its 2022 peak. The median gain was +34.1% in a little over two years; the median state's five-year growth rate immediately before Covid had been 5.5% a year; across the boom it ran at 12.9% annualised.
Seventeen of the fifty-one states posted a twelve-month gain above 20% at some point during the boom — a threshold only thirteen states had ever crossed in the bubble years. And the single fastest year in the entire twenty-six-year record is essentially a tie across sixteen years: Arizona at +40.42% in the year to December 2005, and Idaho at +40.36% in the year to August 2021. The all-time worst year belongs to Nevada, −30.6% in the year to July 2009.
The obvious question is whether these are the same states. Roughly, and only roughly. Plotting each state's 2006–2012 decline against its 2020–2022 gain gives a correlation of −0.41: states that fell harder in the bust did tend to rise harder in the boom.
The relationship is real but loose, and the residuals are the interesting part. Montana and Maine gained as much as the old bust states without having had a bust. The District of Columbia had a middling decline and the weakest boom of all, up just 12% while the median state gained 34%. A single correlation flattens a lot of different stories.
Most states peaked within a few months of mid-2022, when mortgage rates roughly doubled. What happened next is the part of this dataset that is genuinely still unfolding, and it does not look like either previous episode.
Thirty-nine of fifty-one states are above their 2022 peak; twelve are below it. But the split is not random across the map. The gainers cluster in the Northeast and Midwest — Connecticut +25.0%, New Jersey +21.9%, Wisconsin +19.7% — places that had modest booms and, in several cases, had only just clawed back to their 2006 levels.
The decliners are the boom states: the District of Columbia −13.1%, Arizona −9.4%, Colorado −7.2%, Texas −5.8%, Florida −5.1%. Across all fifty-one states the correlation between the size of the Covid boom and the change since is −0.33 (Spearman −0.41): on average, the further a state ran, the more it has given back.
This is the third distinct geography in one dataset. The bubble sorted states by supply elasticity and speculation. The pandemic sorted them by desirability to remote workers. The rate shock is sorting them, so far, by how far they had already gone.
Every figure above is reproducible in Tableau with the same file. The mapping, in order:
In the Data Source pane, select all 318 date columns, right-click, Pivot. Rename Pivot Field Names to Month and change its type to Date; rename Pivot Field Values to ZHVI. Row count goes from 51 to 16,218.
Month on Columns, ZHVI on Rows, RegionName on Detail and Colour. For the shaded recession, use Analytics → Reference Band on the Month axis with constants 2007-12-01 and 2009-06-01; for March 2020, a Reference Line at a constant date. Log axis: right-click the ZHVI axis → Logarithmic.
Create a Date parameter [Base Month], then two calculations:
// Base Value (LOD, one value per state)
{ FIXED [RegionName] :
MIN( IIF([Month] = [Base Month], [ZHVI], NULL) ) }
// Index
( SUM([ZHVI]) - SUM([Base Value]) ) / SUM([Base Value])Format Index as a percentage. Show the parameter control and the chart becomes Figure 3.
SUM([ZHVI]) / RUNNING_MAX(SUM([ZHVI])) - 1Compute Using → Month, partitioned by RegionName. Check the partitioning explicitly; the default direction is wrong about half the time and fails silently.
RegionName carries a geographic role, so double-clicking it produces a filled map. Colour it by Index, not by ZHVI. A choropleth of levels is largely a map of local incomes and land constraints; a choropleth of percent change is a map of what actually happened.