Data Lab

Indicators, forecasts, and demos, with sources and dates.

The index of my public-data work: economic indicators on real data from the Federal Reserve, most back to 1976, with recessions shaded; twelve-month forecasts with backtested error bands; a recession probability; and interactive demos. The charts are a static-site port of the R and Quarto econ dashboards I maintain.

Dated FRED snapshots. Each chart is regenerated from FRED whenever I run the data script, and the as-of month is printed on every card, so you can see exactly how current it is.
Recessions shaded. The gray bands are NBER-dated US recessions, the context these indicators are usually read against.
Hover to read any month. A crosshair tracks your pointer and the reading updates live. No backend and no charting library: the data is a static JSON drawn with a hand-rolled, dependency-free SVG chart.

Yield curve, 10-year minus 2-year

0.46 pp

Aug 2026

The 10-year minus 2-year Treasury spread. Below the dashed zero line the curve is inverted, which has preceded every recession since the 1970s.

Source: FRED (Federal Reserve Bank of St. Louis): GS10, GS2. Recession dating: NBER.

Unemployment rate

4.1%

Aug 2026

The share of the labor force actively looking for work. It bottoms late in an expansion and spikes through every recession (shaded).

Source: FRED (Federal Reserve Bank of St. Louis): UNRATE. Recession dating: NBER.

Inflation, CPI year over year

3.35%

Aug 2026

Headline CPI versus the same month a year earlier. The dashed line is the Federal Reserve's stated 2 percent longer-run objective.

Source: FRED (Federal Reserve Bank of St. Louis): CPIAUCSL. Recession dating: NBER.

Federal funds rate

3.63%

Aug 2026

The Federal Reserve's policy rate, monthly average. It is raised to cool the economy and cut, often sharply, into and through recessions.

Source: FRED (Federal Reserve Bank of St. Louis): FEDFUNDS. Recession dating: NBER.

Housing starts

1275k

Aug 2026

New privately owned housing units started, in thousands at a seasonally adjusted annual rate. Residential construction is one of the most rate-sensitive parts of the economy and rolls over early in most cycles.

Source: FRED (Federal Reserve Bank of St. Louis): HOUST. Recession dating: NBER.

Construction spending, year over year

-3.79%

Jul 2026

Total construction spending versus the same month a year earlier, residential and nonresidential combined. It is the demand backdrop for anyone selling into the trades. The series begins in 1993.

Source: FRED (Federal Reserve Bank of St. Louis): TTLCONS. Recession dating: NBER.

Want to poke at one? The yield curve has an interactive version with a time-range toggle, and there is a short write-up on how to read it. Data: FRED (Federal Reserve Bank of St. Louis); recession dating: NBER.

Models

Twelve-month forecasts

A table of models. Exponential smoothing, ARIMA, LightGBM on lagged monthly changes, and their average each forecast twelve months ahead from 24 past starting points. The lowest error wins, scored against carrying the last value forward.
Bands from real misses. The darker band holds 80 percent of the winning model's backtest errors at each horizon, the lighter band 95 percent.
No filled gaps. A month with no observation stays empty; the models skip it rather than invent a value.

Unemployment rate

3.4% to 4.4%

80% band, Aug 2027

Best in backtest: LightGBM on lagged monthly changes, at 0.96 of the naive forecast's error.

Backtest scores

Mean absolute error over 282 forecast months from 24 starting points, Oct 2021 to Aug 2025, as a share of the naive forecast's error. The same backtest picks the model, so the winner's score flatters it slightly.

Data through Aug 2026; models run 2026-10-03. No observation for Oct 2025; the models skip it. Source: FRED (Federal Reserve Bank of St. Louis): UNRATE. Recession dating: NBER.

Inflation, CPI year over year

1.1% to 5.4%

80% band, Aug 2027

Best in backtest: Average of ETS, ARIMA and LightGBM, at 0.94 of the naive forecast's error.

Backtest scores

Data through Aug 2026; models run 2026-10-03. No observation for Oct 2025; the models skip it. Source: FRED (Federal Reserve Bank of St. Louis): CPIAUCSL. Recession dating: NBER.

Housing starts

1,100k to 1,504k

80% band, Aug 2027

Best in backtest: ARIMA, order picked by AIC, at 0.89 of the naive forecast's error.

Backtest scores

Mean absolute error over 288 forecast months from 24 starting points, Oct 2021 to Aug 2025, as a share of the naive forecast's error. The same backtest picks the model, so the winner's score flatters it slightly.

Data through Aug 2026; models run 2026-10-03. Source: FRED (Federal Reserve Bank of St. Louis): HOUST. Recession dating: NBER.

Regular gasoline, monthly average

$3.34 to $4.41

80% band, Aug 2027

Best in backtest: LightGBM on lagged monthly changes, at 0.89 of the naive forecast's error.

Backtest scores

Data through Aug 2026; models run 2026-10-03. Source: FRED (Federal Reserve Bank of St. Louis): GASREGW. Authoring agency: U.S. Energy Information Administration. Monthly average of the weekly price, complete months only.

Recession probability, 12 months ahead

10%

for Aug 2027

A probit model on the 10-year minus 2-year Treasury spread, refit every month on the outcomes known at the time. Gray bands are NBER recessions.

Data through Aug 2026; model run 2026-10-03. NBER dates are used as they stand today, although NBER announces a turning point months after it happens. The New York Fed's published model uses the 10-year minus 3-month spread. Source: FRED (Federal Reserve Bank of St. Louis): GS10, GS2. Recession dating: NBER.

Want to talk data pipelines?

This is the analytics side of what I do: public data pulled, modeled, and charted, with the as-of month printed on every card. Happy to get into the pipeline, the models, or the charts.

Get in touch

Indicators, forecasts, and demos, with sources and dates.

Dated FRED snapshots. Each chart is regenerated from FRED whenever I run the data script, and the as-of month is printed on every card, so you can see exactly how current it is.

Recessions shaded. The gray bands are NBER-dated US recessions, the context these indicators are usually read against.

Hover to read any month. A crosshair tracks your pointer and the reading updates live. No backend and no charting library: the data is a static JSON drawn with a hand-rolled, dependency-free SVG chart.

Twelve-month forecasts

A table of models. Exponential smoothing, ARIMA, LightGBM on lagged monthly changes, and their average each forecast twelve months ahead from 24 past starting points. The lowest error wins, scored against carrying the last value forward.

Bands from real misses. The darker band holds 80 percent of the winning model's backtest errors at each horizon, the lighter band 95 percent.

No filled gaps. A month with no observation stays empty; the models skip it rather than invent a value.