Gdp Nowcasting Dynamic Factor

GDP nowcasting with a dynamic factor model

A nowcast is not a forecast of the future — it is an estimate of a number that already happened and has not been published yet. Everything hard about it is in the shape of the data: monthly indicators arriving on different days with different lags, a quarterly target observed once every three months, and a bottom edge of the panel that is jagged rather than flat. The model that handles that is a state-space one, and it is already installed.

The Kalman machinery itself lives in ../../../fin-models/skills/state-space-and-kalman/SKILL.md; univariate baselines and how to score a forecast live in ../../../fin-models/skills/time-series-forecasting-models/SKILL.md. This skill is the mixed-frequency, ragged-edge part and the benchmark discipline.

Everything marked ✅ Measured is printed by scripts/nowcast.py (numpy + pandas + scipy, seed 20260909, 6–25 s depending on machine load, no network). Everything marked ✅ source-verified was read at the URL or in the installed package on 2026-09-09.

1. ✅ DynamicFactorMQ is the NY-Fed-shaped model, already in a BSD dependency

✅ source-verified in statsmodels/tsa/statespace/dynamic_factor_mq.py, statsmodels 0.15.0, read in the installed package. The class docstring opens:

*"Implementation of the dynamic factor model of Bańbura and Modugno (2014) and Bańbura, Giannone, and Reichlin (2011). Uses the EM algorithm for parameter fitting… Can incorporate monthly/quarterly mixed frequency data along the lines of Mariano and Murasawa (2011). A special case of this model is the Nowcasting model of Bok et al. (2017)."*

✅ Measured — the script greps the installed docstring for each claim and all six come back yes: mixed monthly/quarterly, arbitrary missing data, EM algorithm, factors unidentified, standardises by default, news decomposition. Also verified in the Notes:

• "The observed data may contain arbitrary patterns of missing entries." — this is what makes the ragged edge a non-problem rather than a preprocessing step.
• "The estimated factors and the factor loadings in this model are only identified up to an invertible transformation… This model does not impose any normalization."
• Mixed frequency is either endog_quarterly= (a separate quarterly frame) or k_endog_monthly= (one frame, monthly columns first).
• res.news(...) decomposes a forecast revision into the contribution of each new release.

⚠️ A dated inconsistency inside one docstring. The prose says Mariano and Murasawa (2011) and Bok et al. (2017); the References section prints Mariano & Murasawa, Oxford Bulletin of Economics and Statistics 72(1) (2010): 27–46 and Bok, Caratelli, Giannone, Sbordone & Tambalotti, 2018, "Macroeconomic Nowcasting and Forecasting with Big Data", Annual Review of Economics 10(1): 615–43. Cite the References years.

⚠️ The docstring also names what it does not do: "See also Banbura et al. (2013). 'Now-Casting and the Real-Time Data Flow.' for discussion about other types of mixed frequency data that are not supported by this framework."

2. The ragged edge, and what dropna() costs

Quarterly growth is the Mariano-Murasawa weighted average of five unobserved monthly growth rates — weights [1, 2, 3, 2, 1] / 3 — observed on the third month of each quarter and nowhere else. Put that in the design matrix and no monthly GDP has to be imputed.

✅ Measured — the bottom of the synthetic panel at a vintage, with publication lags (0, 0, 1, 1, 2, 2) months across six indicators:

🚨 dropna() over this panel keeps 95 of 293 months and deletes every month you actually want a nowcast for. The whole edge is the object of interest. A state-space filter drops the missing rows of the observation equation for that period instead — five lines — and keeps the months.

3. ✅ Measured — factor recovery, and the revision a nowcast owes itself

At the true parameters, so this measures the filter and not an estimator:

⚠️ The smoothed estimate is better because it uses the future, and it is the one you must not trade. Same distinction, same direction, as ../../../fin-core/skills/regime-detection/SKILL.md measures for regime labels.

✅ Measured at the edge — the filtered value now against the value the same month settles at once the panel fills in:

Mean |revision| 0.0552; mean |error| against the truth 0.5904 now versus 0.5558 settled. 🔑 A nowcast revises itself, and the revision is not a bug — it is the ragged edge closing. Report the revision alongside the level, and never score a nowcast against a back-filled panel that had already closed.

4. ✅ Measured — the benchmarks it has to beat

RMSE of the quarterly growth nowcast by position in the data flow. offset is months from the quarter's last month to the vintage: −4 is before the quarter starts, 0 is its final month, +1 is a month after it ends and still before any official estimate. 79 quarters:

🔑 The value of a nowcast is the slope of the first column, not its level. It is the only method here whose error falls as the quarter fills in — the mean and the AR(1) never move because they do not read the monthly flow, and the bridge reads one series of it. The bridge is blank at −4 because there is no monthly data for a quarter that has not started; that is the honest answer, not a defect.

🚨 Report the ratio, not the RMSE. An RMSE of 0.37 means nothing without the 1.76 it is being compared with, and a nowcast that beats the sample mean has cleared no bar at all.

5. ✅ The published record, for calibration

✅ source-verified, read from the Atlanta Fed's GDPNow FAQ (atlantafed.org/research-and-data/data/gdpnow):

"Since we started tracking GDP growth with versions of this model in 2011, the average absolute error of final GDPNow forecasts is 0.77 percentage points. The root-mean-squared error of the forecasts is 1.17 percentage points. These accuracy measures cover initial estimates for 2011:Q3–2025:Q2."

and, on the same page:

"When back-testing with revised data, the root mean-squared error of the model's out-of sample forecast with the same data coverage that an analyst would have just before the 'advance' estimate is 1.15 percentage points for the 2000:Q1–2013:Q4 period."

🚨 "Overall, these accuracy metrics do not give compelling evidence that the model is more accurate than professional forecasters. The model does appear to fare well compared to other conventional statistical models."

🔑 The best-known nowcast in the world says, on its own page, that it does not beat forecasters. Size your expectations accordingly, and note that "final GDPNow" means the forecast made just before the BEA advance estimate — the easiest one it makes.

🚨 The New York Fed Staff Nowcast has a two-year hole. ✅ source-verified at newyorkfed.org/research/policy/nowcast: "Updates to the New York Fed Staff Nowcast were suspended between September 2021 and September 2023." The cause, in the Fed's words: "The COVID-19 pandemic generated considerable uncertainty and volatility with respect to macroeconomic data, which posed important challenges to the … model." The relaunched model is different — "still based on a dynamic factor model and still employs Kalman-filtering techniques. The current model, however, is estimated with Bayesian techniques", adding "time-varying volatility as well as variance outliers". It is published *"each Friday (except on federal holidays) at or shortly after 12:45 p.m., using data available up to 10 a.m."*

Downloading "the NY Fed Staff Nowcast" as one long feature splices two models across a two-year gap. ❓ No accuracy figure appears on that page — see the unverified list.

6. ✅ Measured — what DynamicFactorMQ recovers

Fitted on 180 months, 6 monthly indicators plus quarterly GDP, one factor, EM, maxiter=60, 2–5 s, 16 parameters:

🚨 Every loading came back negative. That is the identification caveat in the docstring arriving in your results: the factor and the loadings are determined only up to an invertible transformation, so a sign flip and a common rescaling are free. ✅ Measured: divide the estimated loadings by a single common scale (0.7593) and they line up with the truth the model can actually see — λ_i / sd(x_i), because it standardises by default. Compare correlations, ratios and fitted values. Never read a raw loading, and never compare loadings across two fits.

7. Traps

• 🚨 dropna() before fitting. §2. It deletes the edge, which is the whole point.
• 🚨 Imputing the missing edge instead of leaving it missing. Forward-filling a monthly indicator into the vintage month invents an observation with zero uncertainty; the filter's own conditional mean already does this, correctly, and carries the variance.
• 🚨 Reading raw loadings, or comparing them across fits. §6.
• 🚨 Scoring against a back-filled panel. The evaluation must reconstruct the panel as of each nowcast date. The inputs are revised too — ../real-time-macro-backtesting/SKILL.md — and the monthly indicators are seasonally adjusted, which is a second vintage (../seasonal-adjustment-and-x13/SKILL.md).
• 🚨 Comparing to the sample mean and stopping. §4. Beat an AR(1) and a bridge, and report the ratio.
• 🚨 Splicing a published nowcast series. §5: the NY Fed's has a two-year hole and a model change across it; GDPNow's own page links a "Modifications to GDPNow Model" document. A published nowcast is a model's output at a date, not a measurement.
• 🚨 Using the smoothed factor as a signal. §3. Filtered for decisions, smoothed for description.
• ⚠️ The vintage of the target matters as much as the inputs. GDPNow is scored against the BEA advance estimate — the first print — not against today's revised GDP. Pick one, say which, and never mix them inside one error series.
• 🚨 Every specification is a trial. Number of factors, factor order, idiosyncratic_ar1, which indicators, maxiter. Log them in ../../../fin-core/skills/backtest-validation/scripts/trial_ledger.py.
• ⚠️ MIDAS is not pip install midaspy. ⚠️ Secondhand from a prior research pass, not re-verified here: PyPI midaspy is a multiple-imputation package and PyPI midas is a gas-detector driver; neither is mixed-data-sampling regression. If you want MIDAS rather than a factor model, expect to write the Almon/beta weighting yourself.

❓ Not verified

The New York Fed Staff Nowcast's accuracy — no figure appears on its page, and none is quoted here · the exact relaunch date in September 2023 (the page gives only "September 2023") · whether the pre-2021 and post-2023 NY Fed series are published as one downloadable file · GDPNow's performance against Blue Chip quarter by quarter (only the Fed's own summary sentence is quoted) · the MIDAS package names above.

8. Scripts and where this sits

scripts/nowcast.py — the seeded mixed-frequency panel with per-series publication lags, the ragged() vintage view, a Kalman filter/smoother with a time-varying observation dimension carrying the Mariano-Murasawa aggregation in the design matrix, factor recovery and the edge-revision table, the four-way benchmark evaluation by data-flow position, and the DynamicFactorMQ docstring audit and fit. numpy + pandas + scipy; statsmodels imported inside functions and §6 skipped cleanly without it. Seed 20260909, 6–25 s.

• The Kalman filter and smoother themselves, and the filtered/smoothed distinction — ../../../fin-models/skills/state-space-and-kalman/SKILL.md.
• Forecast baselines, scoring rules and why a naive benchmark is mandatory — ../../../fin-models/skills/time-series-forecasting-models/SKILL.md.
• The vintages of every input series — ../real-time-macro-backtesting/SKILL.md.
• The seasonal adjustment inside those inputs — ../seasonal-adjustment-and-x13/SKILL.md.
• When each indicator actually lands, to the minute — ../macro-release-calendar-and-embargo/SKILL.md.
• Recession probabilities as a different question about the same data — ../macro-regime-and-recession-indicators/SKILL.md.
• Where GDP, the monthly indicators and GDPNOW come from — ../../../fin-core/skills/fundamental-and-macro-data/SKILL.md.