Getting a Handle on State-Level Economic Forecasts
Most people I talk to confuse "states economic outlook" with some proprietary software product. It isn't. It's a general category of forecasting and analysis that every state government, economic development agency, and private firm does differently. The term itself isn't a trademark or a specific tool. It's just what happens when you model employment, GDP, tax revenue, and sector growth at the state level and try to see where things are heading over the next one to three years. If you're looking to build or interpret a state economic outlook, the first thing you need to understand is that none of the standard models work cleanly when applied to a single state. The Bureau of Economic Analysis data is quarterly and revised constantly. The BEA seasonal adjustment procedures were built for national time series, not for a state like Mississippi where agricultural cycles and military base staffing create noise that looks like trend shifts. I learned that the hard way in 2021 when my forecast for Alabama's professional services sector was off by eleven percent because I didn't account for the DoD's delay in announcing base realignment hiring timelines. The BEA numbers for Q2 showed a spike that I treated as structural demand. It was actually front-loaded spending from a single contract that closed in June. That one contract distortion blew out my entire regional outlook model.
Building a Practical States Economic Outlook From Scratch
Start with the data sources that actually matter, not the ones that are convenient. The Bureau of Labor Statistics publishes State Employment and Unemployment monthly, but it has a revision cycle that makes it unreliable for anything shorter than a six-month lookback unless you anchor to the annual benchmark revision. The BEA runs quarterly state GDP at annualized rates, but the initial estimates carry an error margin that shrinks only after two revisions. The Census Bureau releases state-level business formations and closures through the Quarterly Workforce Indicators, which are currently the most granular real-time signal available, updated quarterly with about a ninety-day lag. The typical workflow I use goes like this. I pull the QWI data for the state and its primary metro areas. I overlay BLS state industry employment with BEA industry output. I run a leading indicator composite built from building permits, initial unemployment claims smoothed over four weeks, state tax withholding data when I can get it, and the Philadelphia Fed's State Economic Index if the state is covered. Then I stress-test the composite against historical pivots to calibrate what kind of signal strength actually predicts direction changes versus just following along. I combine those leading signals using a simple weighted regression model rather than anything fancy. The reason is that complex models overfit state-level data too easily. With twelve years of monthly observations and forty-eight industry sectors, a random forest will give you impressive backtest accuracy that collapses in production. A weighted composite of three or four indicators with annual reweighting after each BEA benchmark revision tends to hold up better. I usually reweight once a year when the BEA releases its annual estimate revisions for state GDP by industry.
What Most People Miss About State Economic Data
The first counter-intuitive thing most forecasters get wrong is assuming that state-level manufacturing data reflects domestic demand. It doesn't. Thirty-nine percent of Michigan's manufacturing output is automotive, and automotive shipments are driven by national and global interest rates, not Michigan consumer spending. When you see Michigan manufacturing employment rising, it might just mean Ford and GM are restocking parts yards, not that the state economy is strengthening. You have to separate tradable from non-tradable sectors explicitly. Non-tradable sectors like healthcare, education, and local government are the actual demand-side story for any given state. Tradable sectors are export drivers that will move with national or international conditions regardless of what the state itself is doing. The second thing people miss is that migration is both a lagging and leading indicator depending on which direction you're looking. Out-migration from Illinois over the past decade lagged the tax policy changes by roughly eighteen months because people don't move their entire lives on short notice. But once the trend started, it accelerated faster than any single indicator predicted. The warning sign wasn't any employment metric. It was the decline in individual income tax withholding per capita, which turned negative before the net migration data confirmed it in the IRS statistics. If you want to catch a state trend early, look at per-capita tax withholding and per-capita retail sales before you look at employment. Employment is the confirmation, not the signal.
Get the Full Details
Common Pitfalls That Will Ruin Your Model
Seasonal adjustment is the easiest place to waste weeks of work. Most states have unique seasonal patterns that X-13ARSEA-GLS won't handle without customization. Florida's tourism season overlaps with New England's, but the timing is different enough that using a national seasonal factor for a state like Florida introduces a systematic bias in April and October. I spent three weeks in 2023 debugging a forecast error that turned out to be a seasonal factor mismatch. The fix was running a state-specific seasonal decomposition on the ten-year rolling window before applying any trend extraction. Another pitfall is conflating price and volume. State GDP contains a heavy services component, and services prices rise faster than goods prices. When someone reports that a state's economy grew six percent annually, check whether that's nominal or real. Several states showed apparent double-digit growth in 2022 that was entirely nominal. Inflation in construction materials and professional services drove the number up while actual output either flatlined or contracted.
Where This Approach Breaks Down
State-level economic outlooks have real limitations that no amount of model sophistication removes. The data is fundamentally lagged. The most recent month of BLS establishment survey data comes out three weeks late and gets revised for at least six months. State-level personal income from the BEA is quarterly with a sixty-day lag and is further revised annually. You are always forecasting based on data that is at best two months old and possibly six months old depending on which variable you're using. Small states are especially unreliable. Delaware, Wyoming, and Alaska have population and employment bases so small that a single large employer hiring or laying off five hundred people creates a statistical signal that looks macroeconomic but is really idiosyncratic. The coefficient of variation for monthly employment change in those states is three to four times higher than the national average. Any outlook built purely on time-series methods will oscillate uselessly for those states. If you need a more reliable alternative for small or volatile states, the best workaround I've found is to anchor your state forecast to a peer-state cluster model. Group states with similar industrial composition and economic structure, then let the peer group average smooth out the noise in any single state. A state like Tennessee, for example, tracks closely with North Carolina and Virginia when you control for manufacturing share and population growth. The peer group signal is more stable than the state-specific time series.
Practical Steps to Produce a States Economic Outlook You Can Actually Use
First, define the geographic scope precisely. Metropolitan statistical area, combined statistical area, or state-wide. The data quality and availability differ substantially across these boundaries. MSA-level BLS data exists but only for metros with populations above a certain threshold, and it's less timely than state-level data. If you need metro-level detail, the QWI is your best option and it covers both state and metro levels. Second, pick your primary output variable and commit to it. Forecasting state employment, state GDP, state tax revenue, and state consumer spending requires different models and different data weights. Don't try to produce all four from the same framework. Employment forecasts benefit most from QWI flow data. GDP forecasts need the BEA industry accounts. Tax revenue forecasts require withholding and sales collection data. Consumer spending forecasts work best with per-capita retail sales and credit card transaction aggregates when you can access them. Third, set a revision schedule and stick to it. I revise my outlooks every six weeks, which aligns with the BLS monthly release and the quarterly QWI publication cycle. Between releases, I don't update the model. I track leading indicator deviations and note them, but I don't let a single month of noisy data change the forecast. That discipline alone has kept my error rates down. The biggest source of forecast degradation isn't a bad model. It's overreacting to noise.
If you're just getting started and don't want to build this from scratch, the BEA provides a free StateArea Models dataset that covers county and state-level employment and earnings going back to 1969. It's not a forecasting tool, but it's the foundational dataset most state economic outlooks are built on. The Federal Reserve Bank of Philadelphia also maintains the State Economic Index, which is a free composite of thirteen indicators for forty-two states and is useful as a baseline check against your own model outputs. The Census Bureau's QWI is downloadable through their Research Data Center for anyone with an approved project, and it's the most detailed state labor market dataset available, covering job creation, destruction, hires, separations, and wages by industry and geography at the MSA and county level. The whole process usually takes about four to six hours for a first-pass state outlook if you already know how to clean and merge the datasets. A polished version with uncertainty bands and scenario analysis runs closer to a full workday. The bottleneck is never the modeling. It's the data cleaning, especially when you're reconciling definitions across BLS, BEA, and Census datasets that use different industry classification systems. BEA uses NAICS 2017. BLS used to publish in NAICS 2012 for several years and is still working through the transition. The mismatch causes phantom industry movements that look real if you're not watching closely. A state economic outlook is only as good as its underlying assumptions about demographic trends and federal policy. I've seen solid models fail because they assumed net migration would continue at the pace of the prior decade. Migration patterns shifted sharply after 2020 and the models that baked in pre-2020 migration rates produced outlooks that were wrong in the direction, not just the magnitude. The models that treated migration as a scenario variable rather than a fixed input stayed useful even when the actual trajectory diverged from the base case.
