Macroeconomic Time-Series Forecasting

Why This Matters

Central banks publish forecasts that move markets and feed directly into rate decisions. The forecast user typically wants two things at once: the best point prediction of GDP, inflation, or unemployment over the next eight quarters, and a story about why a path looks the way it does so that counterfactuals ("what if oil rises 20 percent?") can be priced. These goals pull in opposite directions. Structural models give the story; flexible predictors give the accuracy. Modern forecasting stacks try to keep both.

The empirical record is clear that pure forecast accuracy on macro series favors flexible reduced-form models over fully structural ones (Stock and Watson 2002, JBES 20). Yet DSGE models remain in production at the Fed, ECB, Bank of England, and Bank of Canada because they answer questions ML models cannot: what would inflation be under an alternative policy rule, how does a permanent productivity shock propagate, what is the variance decomposition across structural shocks.

Core Ideas

DSGE baselines. A dynamic stochastic general equilibrium model writes households, firms, and a central bank as optimizing agents subject to budget and resource constraints, then linearizes around steady state and estimates deep parameters by maximum likelihood or Bayesian methods. Smets and Wouters (2007) is the canonical medium-scale model. Forecasts come from the state-space representation. The selling point is not raw accuracy: it is that every shock has an interpretation and the model supports policy counterfactuals.

Large-scale dynamic factor models. Stock and Watson (2002) showed that a small number of latent factors extracted from $N \approx 100$ to $200$ macroeconomic series captures most of the predictable variation in any one series. The forecast equation is $y_{t+h} = \alpha + \beta' \hat{f}_t + \gamma' y_t + \varepsilon_{t+h}$, where $\hat{f}_t$ is recovered by principal components on the panel. These models routinely beat univariate benchmarks and remain a hard baseline for any new method.

Neural forecasters. N-BEATS (Oreshkin et al. 2020, ICLR; arXiv 1905.10437) and NHITS (Challu et al. 2023, AAAI; arXiv 2201.12886) are deep MLP-based architectures designed for univariate time-series forecasting that won and extended results on the M4 competition. Transformers and state-space models have been adapted as well. Central-bank research groups now evaluate these as candidates for nowcasting (current-quarter GDP) and short-horizon inflation forecasting (Athey and Imbens 2019, Annual Review of Economics 11; Chakraborty and Joseph 2017, Bank of England Staff Working Paper 674).

Why structural models still win for policy. A neural forecaster trained on the last 30 years of quarterly data can predict next quarter's CPI well. It cannot tell the governor what inflation would do under an unobserved alternative policy rule, because that counterfactual is outside the support of the training data. DSGE models extrapolate from estimated structural parameters under explicit Lucas-critique assumptions; ML models extrapolate under the implicit assumption that the data-generating process stays put. For policy counterfactuals, the structural assumption is the price of admission, not a bug.

Common Confusions

Watch Out

Better forecast accuracy is not better policy guidance

A model with lower out-of-sample RMSE on inflation can still be the wrong tool for setting rates. The policy question is conditional: what happens under intervention $X$. A flexible model fit on observational data has no mechanism to answer that without strong assumptions equivalent to a structural model. The Lucas critique still binds.

Watch Out

Macroeconomic samples are tiny by ML standards

Quarterly GDP since 1947 is roughly 320 observations. Modern deep nets have millions of parameters. Reported gains over linear baselines should be treated with skepticism unless they hold across multiple subsamples, publication vintages, and recession episodes; many do not.

References

Related Topics

• Time-Series Forecasting Basics
• Recurrent Neural Networks
• Transformer Architecture
• Causal Inference for Policy Evaluation
• Kalman Filter