← Back to the US Economy Barometer
The forecasts at the top of this site — for inflation, unemployment, economic growth, and inflation-adjusted wages — come from a statistical model. This page explains how they are made and their limitations. A more technical description, including the models' out-of-sample accuracy record, is in the methodology note (PDF).
The model uses several decades of official US data: consumer prices, unemployment and other labor-market measures, payrolls, hours and earnings, job openings and quits, industrial production and factory capacity use, retail sales, building permits, consumer sentiment and household inflation expectations from the University of Michigan, mortgage rates, Treasury interest rates, weekly unemployment insurance claims, quarterly GDP, and weekly earnings across the pay distribution. Everything comes directly from the original government and university sources.
Three different forecasting approaches are combined, because decades of forecasting research show that averaging models often performs better than trusting any single one:
1. A dynamic factor model. This is the same type of model the Federal Reserve Bank of New York uses for its GDP “nowcast.” It distills the common movement in all the data series into a few underlying trends and projects them forward. It is especially good at reading the economy's current momentum from data series that arrive at different times.
2. A machine-learning model. A gradient-boosting model (a technique that builds many small decision trees) looks for patterns in how today's data relate to where inflation, unemployment, and growth end up a year later. It can pick up relationships the other models miss, at some cost in transparency.
3. A simple statistical benchmark. Each number is also forecast from nothing but its own history. This very simple approach is surprisingly hard to beat, so it ends up having a substantial weight in the average.
The three forecasts are averaged, with more weight on whichever approach has been more accurate historically for that particular number and time horizon. For wage growth across the pay distribution, only the first and third approaches are used. This is because the quarterly wage data start in 2000, which is too little history to train the machine-learning model well.
Every forecast comes with a “68% range” and, on the charts, a wider “90% range.” These are based on a simple exercise: we started at every month since 2005, gave the model only the data it would have had at the time, let it forecast, and recorded how wrong it was. The ranges show how far off those historical forecasts typically were. If the future is similarly uncertain, the actual number should land inside the 68% range about two-thirds of the time. This also means it will land outside that range about one time in three. Similarly, the forecast should land inside the 90% range nine out of ten times and outside of it one out of ten times. Prediction is hard, especially about the future!
Forecasting models extrapolate patterns; they cannot foresee wars, pandemics, financial crises, or policy surprises. The accuracy test above also uses today's (revised) versions of the data rather than the preliminary numbers available in real time, so the ranges are, if anything, a little too narrow — especially for GDP, which is revised heavily. The forecasts are therefore a disciplined summary of where current data point, not a promise.
The forecasts are refreshed every month when new data come out, and the model's accuracy record (which sets the weights and the ranges) is re-checked once a year.