Earlier counts are the examples.
The model learns to predict the next month from previous counts: 1, 2, 3, 6 and 12 months earlier, recent averages, and annual seasonal patterns. This is autoregression: a series learns from its own past.
THE LONG VIEW
New Jersey, over time.
Follow the reach of New Jersey’s major family assistance programs.
DEC 2008 — JUN 2026Official DFD program statistics
Explore the published values
Every program, its own scale.
Monthly participation and quarterly child-support collections.
Participating caseload over time
SNAP counts households, TANF counts families, and General Assistance counts cases. Emergency Assistance uses its parent program’s unit. Programs overlap, so these counts cannot be added to estimate unique households. Child Support counts cases with orders from the county case tables; collection dollars appear on Program trends.
Relative change in participation
Each monthly series starts at 100. Units differ; this compares change, not program size.
Who receives SNAP and TANF?
Published adult and child counts; each chart has its own vertical scale.
Emergency Assistance relative to its parent programs
Monthly administrative ratios, with a separate scale for each chart.
Start with the New Jersey state total. Choose a county above to see its counts, household context, history, and Emergency Assistance ratios.
A household reference for every county
2020–2024 American Community Survey · a consistent, fixed reference
SNAP share of estimated households
Participation relative to household size
The same ACS period is used for all 21 counties. The percentage is approximate for SNAP; families and cases in other programs are units per 100 households, rather than the share of unique households served.
Published report snapshots over time.
March, June, September, and December snapshots. Counts describe the report month; they are not quarterly totals. Intervening monthly values are not inferred.
New Jersey total and all 21 counties
Share uses the sum of published county values. Year-over-year change is calculated from downloaded county snapshots.
Emergency Assistance across counties
EXPLORATORY · SIX-MONTH OUTLOOK
Published history and a forecast learned from earlier monthly patterns.
ObservedForecastError range
The shaded range reflects historical forecasting errors. It is an exploratory 80% error range, based on 18 past forecast starting points, and does not guarantee future coverage. Policy changes and reporting revisions can invalidate the outlook.
What kind of machine learning is this?
Supervised learning for a time series · ridge autoregression
Ridge keeps the fit restrained.
Ridge regression penalizes large weights so the model is less likely to chase noise. Inputs are put on comparable scales, and counts are log-transformed to learn proportional changes.
One month leads to the next.
For the six-month outlook, each forecast becomes an input to the following month. This is recursive forecasting. Uncertainty can accumulate as the model looks further ahead.
Test on months it has not seen.
We train on earlier months, predict later months, and compare those predictions with published counts. Repeating this at 18 starting points is rolling time validation; future observations stay outside each fit.
A Census-informed county outlook
Supervised learning across 21 counties · two-quarter outlook
Blue: decreaseNeutral: little changeOrange: increase
Rectangle area shows the latest published caseload; color shows projected change by December 2026. Click a county to explore its history. This pooled county model learns from quarterly changes, seasonal patterns, household counts, median household income, and average household size. It uses no poverty predictor. Its county forecasts are separate from the monthly statewide outlook above.
From forecasts to reviewable planning signals
Historical-error scenarios and explicit review rules.
For each county, we apply eight historical forecast errors to the current projection. A review signal appears when at least six of eight scenarios exceed your increase or decrease threshold. Otherwise the result is “mixed scenarios.” These are scenario counts, not calibrated probabilities. Counties with a zero current count receive “insufficient reference.”
How should I interpret these scenario counts?
The scenarios apply eight past forecasting errors to the current ridge-regression forecast. They show how the planning question changes under errors seen before. Six of eight scenarios passing a threshold is a review rule, not a 75% probability of that outcome.
Calibration means that outcomes assigned a probability of 70% occur about 70% of the time across comparable trials. Eight overlapping historical examples are insufficient to establish that here. Signals support questions about future service capacity; they do not measure unmet need or determine eligibility.
The county change calendar · heatmap
What changed together, and when? Published June counts, compared with the previous June.
Blue cells show decreases; orange cells show increases. Color strength is capped at ±30%, while tooltips show the full change and both counts. Blank cells have no nonzero prior-year reference. This is observed history, rather than a model prediction. The July 2013 reporting source changed for SNAP, TANF, and TANF EA; the calendar cannot separate that effect from other changes.
Unusual monthly changes · anomaly detection
A robust score compares a prediction error with earlier errors using their median and median absolute deviation. Flags do not establish causes, fraud, or data errors.
Look inside the monthly model
Smaller diagnostics explain the fit and its past errors. They do not identify causes.
How close were past predictions?
What patterns did it learn?
Signed weights show how an input changes a prediction while other inputs stay fixed. Related inputs share information; these are not causal effects. Seasonal sine and cosine are paired wave-shaped inputs that represent where a month falls in the annual cycle.
Published numbers. Traceable sources.
Statewide monthly values come from DFD Tables 1 (TANF), 5 (SNAP), 8 (TANF Emergency Assistance), and 10 (General Assistance and GA Emergency Assistance). Each observation records its source PDF, table, and PDF page.
Overlapping histories are reconciled using the most recent downloaded report. Monthly coverage is December 2008–June 2026. Earlier scanned reports are excluded. Child Support uses the actual quarter end printed in Table 12, which can differ from the report month.
These are participation measures, not estimates of unmet need or unique people across programs. Emergency Assistance recipients overlap their parent programs. TANF EA + GA EA is shown only as a blended administrative ratio; it is not a total of unique households or people. SSI Emergency Assistance is excluded from that sum.
County breakdowns use published report-month snapshots from March, June, September, and December reports, starting in March 2010. Tables 2, 7, 9, and 11 supply county participation counts; the Child Support cases-with-orders table supplies county case counts. County Child Support measures differ from statewide collection dollars. County shares use the sum of county rows, and annual changes are calculated from downloaded snapshots. Standalone source-table dashes are treated as published zeros. County/statewide differences caused by source revisions or source inconsistencies are retained and disclosed; intervening months are not inferred.
A model you can inspect.
Ridge autoregression learns the next month’s log-transformed count from lags of 1, 2, 3, 6, and 12 months, trailing averages, and annual seasonal features. It uses up to 84 recent training months and a fixed regularization strength of 1.
Forecasts are recursive across six months. Validation uses 18 rolling time origins, refitting using only earlier data at every origin. The benchmark repeats the same month from the prior year. No random train/test split is used.
Additional Census context comes from official 2020–2024 ACS 5-year Summary Files: B11001 (households), B17017 (household poverty), B19013 (median household income), and B25010 (household size). Household margins of error are shown. A fixed ACS period provides the same denominator for every county; it is not a contemporaneous household count for earlier snapshots or June 2026. SNAP/ACS definitions differ, and other programs measure families or cases.
The pooled county ridge model predicts the next quarter’s log growth from three recent growth features, two seasonal features, cases relative to households, household count, median household income relative to other counties, and average household size. It fits at most 16 quarterly origins across 21 counties, standardizes inputs using training data only, and uses regularization strength 10. Eight rolling origins test two-quarter forecasts. The 2015–2019 ACS is used from March 2021; the 2020–2024 ACS is used from March 2026, after publication. Future Census releases stay outside each fit. Updated historical Census and DFD extracts are not archived real-time vintages. Eight paired historical log-error scenarios supply the county ranges and planning signals; they are not calibrated event probabilities.
Anomaly flags use one-step prediction errors and a median absolute deviation score based only on the preceding 36 residuals, after a 24-residual warm-up. The absolute score threshold is 3.5. Child Support is visualized quarterly and is excluded from the monthly models.
Read the shifts with care.
DFD notes a data-query conversion in July 2013 for TANF, SNAP, and TANF Emergency Assistance. General Assistance is labeled “persons” in older reports and “cases” in newer ones; long-term comparability is limited. Child Support dollars are nominal and not adjusted for inflation.
Backtests use today’s reconciled history, not archived data vintages as known at each forecast origin. Origins overlap, and uncertainty bands have a small, dependent sample. Historical model performance does not establish future accuracy.
The monthly forecast uses no economic, demographic, or policy predictors. It is intended for exploratory analysis, not eligibility decisions or budgeting. General child-care enrollment is not reported in these CPS tables; the narrow post-TANF child-care measure is not a substitute.
Source differences stay visible.
Counts are validated for date continuity, numeric values, and component consistency. Where published GA totals differ from the listed employable and unemployable counts, the published total is preserved.
Download the complete source-linked datasetDownload county snapshots
Download Census estimates and county groups
Download Census-informed county models
Download statewide Census estimates
Download model results and backtests
Report library
Official reports used to assemble this snapshot.