# Sources and methodology

All data accessed 2026-11-05 (America/Denver). Raw files are preserved unmodified in `raw/`. Rebuild everything with `python3 build.py` (needs pandas, openpyxl, xlrd, matplotlib).

## Sources
| Input | Series | URL | Release info at access |
|---|---|---|---|
| Home price | Census/HUD Survey of Construction, median sales price of new houses sold, US (includes land). Sheets "Annual" (1963-2025) and "Monthly" (through Sep 2026) | https://www.census.gov/construction/nrs/xls/price_cust.xlsx (index: https://www.census.gov/construction/nrs/data/series.html) | File last-modified 2026-10-27 14:00 UTC (8:00 AM MDT) |
| Rate, 1971 to 2026 | Freddie Mac PMMS, 30-year fixed, weekly, via FRED MORTGAGE30US | https://fred.stlouisfed.org/graph/fredgraph.csv?id=MORTGAGE30US | First obs 1971-04-02, latest 2026-11-05 (7.18%) |
| Rate proxy, 1965 to 1970 | FHFA (formerly FHLBB) Monthly Interest Rate Survey, Table 10, annual national average contract rate, conventional single-family loans closed, newly built homes | https://www.fhfa.gov/sites/default/files/Table10-2018-by-Year.xls (index: https://www.fhfa.gov/data/mirs) | Historical table 1963-2018; MIRS discontinued |
| Inflation | BLS CPI-U, all items, US city average, NSA, via FRED CPIAUCNS | https://fred.stlouisfed.org/graph/fredgraph.csv?id=CPIAUCNS | Latest obs Sep 2026 (336.94); FRED page updated Oct 14, 2026 |

Also saved for reference: FHFA Table 9 (all homes), which was not used.

## Definitions
- Price: Census annual median for 1965-2025. Census revises these. The current file gives 2021 = $383,500, versus $397,100 in an older 2024 vintage, so the current vintage is used throughout. 2026: no annual figure exists yet, so the value is a PROXY, the median of the Jan to Sep 2026 monthly medians ($410,200).
- Rate (PMMS era): the median of all weekly PMMS readings in the calendar year. 1971 has only 40 weeks (April start). For 2026, weeks are limited to Jan to Sep (38 weeks, median 6.445%) to match the price and CPI window. Sensitivity check: all 44 weeks through Nov 5 give a median of 6.485%, which works out to $2,100.24 real instead of $2,091.50.
- Rate (1965-1970 proxy): FHFA Table 10 is an annual AVERAGE contract rate on loans CLOSED (not a median, not survey quotes, and covering all conventional loans at all terms, with an average term of about 25 years then; fees excluded). It is NOT equivalent to PMMS. The CSV has an `era` column and keeps both raw rate columns, the chart draws it as a separate dashed series with a series-break line, and no value is silently spliced in.
- Overlap check (`proxy_overlap_check.csv`): in 1971-1973 the FHFA rate is within 0.05 points of PMMS. From 1974 on, PMMS runs higher, by 1.1 points in 1980 and 2.6 in 1981, because closed-loan rates lag commitments. In stable years the proxy is close; it is less reliable in fast-moving ones. 1965-1970 was a moderately rising-rate period, so the proxy probably understates those payments slightly.
- CPI: the annual mean of monthly NSA CPI-U (2026 = mean of Jan to Sep). Today's dollars use September 2026 CPI, the latest release. No future month is used. The saved CPI file has no observation for October 2025. The 2025 mean therefore uses the other 11 months. Putting October halfway between September and November would change the 2025 real payment by about $1.50. It does not change the years the essay discusses.

## Formula
L = 0.8 x price. r = annual rate / 1200. n = 360.
M = L x r / (1 - (1 + r)^-n).
Real M = M x 336.94 / CPI(year). The down payment (0.2 x price) is deflated the same way.

## Searches for pre-1971 30-year rates
- Freddie Mac PMMS: begins April 1971, so nothing earlier exists.
- FHFA MIRS (formerly FHLBB): annual contract and effective rates from 1963. Used as the proxy.
- Historical Statistics of the United States, table Dc1210-1254: reproduces the same FHLBB series, for example 1970 = 8.27% for new homes (https://hsus.cambridge.org/HSUSWeb/toc/treeTablePathIdDc1210-1254.html).
- No national 30-year fixed median series for 1965-1970 was found, so none is claimed.

## Limitations
- The price data covers new homes only. Existing-home sales (NAR) differ in mix, and new homes have grown larger over time.
- The median price and the median rate are separate medians. They do not describe any single real transaction.
- The payment is principal and interest only. It excludes property tax, insurance, PMI, HOA, maintenance, closing costs, and the cost of saving the down payment.
- This is NOT income-relative affordability, because no income data is used. The rank columns rank payment size only.
- The lowest-rate year is not necessarily the lowest-payment year. 2021 had the lowest median rate (2.975%), yet it ranks 16th lowest of 56 PMMS years by real payment, or 22nd of 62 including proxy years.
- Context source for the 1981 rate spike (Volcker tightening, interest rates spiking 1979-1981): Federal Reserve History, "The Great Inflation" by Michael Bryan, https://www.federalreservehistory.org/essays/great-inflation (accessed 2026-11-05). Article payment figures come from this dataset only.

## How the essay reads the series

The monthly figure is a modeled payment on the median-priced new home: 20% down, a 30-year amortization, that year's rate, then CPI-U to September 2026. It is not an observed median household payment.

Callout peaks are completed years whose modeled real payment is the highest within five years on either side. 2026 is partial, so it cannot be chosen. A lower year inside a higher year's window drops out, which is why 1980 is not listed next to 1981. On this file the rule selects 1981 ($2,872.28), 2000 ($1,966.07), 2006 ($2,058.43), and 2023 ($2,451.65). These are neighborhood peaks, not a list of the four largest payments. 1982 ($2,706.38) is the second-highest year and part of the 1981 episode.

The sustained rise into the series high is the run of consecutive increases that ends at that high. It starts in 1976. 1975 is slightly below 1974, so 1975 is the pause before the climb, not the start. 1976 to 2026 is fifty years. The chart still plots 1965 onward. The cheapest year in the file is 1965, and it is inside the FHFA proxy. The cheapest Freddie Mac year is 1971. 2012 is the cheapest Freddie Mac year after 1972. 2017 and 2021 are not minima.

## Copies linked from the essay

`python3 build.py` writes the CSV, the charts, and the summary tables in this directory. The essay links to copies under `public/data/mortgage-payments/`. After a rebuild, copy those outputs there. Raw inputs stay in `raw/` and are not published as site downloads.

Charts, all drawn from this dataframe:

- `real_payment_chart`: full 1965-2026 modeled real payment, proxy years dashed, 2026 marked partial, 1976 marked.
- `home_prices_nominal_real`: nominal dollars and September 2026 dollars on separate panels.
- `mortgage_rates`: the rate actually used in the payment, with the same proxy and partial-year marks.
- `real_downpayments`: 20% of the median price, in September 2026 dollars.
- `payment_20_vs_10`: 20% down and 10% down. The second line is 1.125 times the first in every year.
- `loan_structures`: 30-year, 40-year, and interest-only illustrations for 2021 and partial 2026.

## Additional outputs
- `payment_10down_*` columns: same inputs with a 90% loan, so each payment is exactly 1.125 times the 20%-down payment.
- `window_averages.csv`: simple means of annual real payments over fixed windows.
- `rolling_windows_pmms.csv`: every 5-year and 11-year rolling window over the PMMS era only (1971 to 2026), ranked from lowest to highest. The 2026 YTD value is included in windows that end in 2026. Proxy years are excluded so the series doesn't mix.
- `loan_structure_illustrations.csv`: 2021 and 2026 at the same rate and loan size, with 360 and 480 month amortization and interest only (L x r). These are hypothetical. 40-year and interest-only loans are generally not qualified mortgages and are not typical products. The bar chart uses these rows and no others.