Applied Econometric Time Series

Fitting an ARIMA model to quarterly GDP and watching the residuals scream structural break is something I still see people ignore in three-quarter of the papers that land on my desk.

I spent most of last year rebuilding the forecasting pipeline for a regional development bank that kept getting burned by oil-price shocks. Their standard practice was to difference the series until the ADF test p-value fell below 0.05, then plug the result into ARFIMA with a seasonal MA term. That worked fine in 2016. It failed spectacularly after the 2020 demand collapse because the variance regime shifted twice in eighteen months. What saved the project was not a fancier model; it was treating the non-stationarity as a mixture of deterministic trends, known breakpoints, and a stochastic component that changed over time. If you want to understand how Applied Econometric Time Series actually behaves outside the textbook examples, start by accepting that every stationary assumption is a temporary hypothesis, not a settled fact. The phrase covers the practice of extracting interpretable dynamic structure from ordered observations—prices, output, unemployment rates, disease counts—using likelihood-based or quasi-likelihood methods that respect autocorrelation, heteroskedasticity, and measurement error. It sits between theoretical econometrics and data engineering. Theory gives you identifiability conditions and asymptotic distributions. Engineering gives you fast matrix updates, reproducible pipelines, and diagnostics that do not rely on perfect normality. Most practitioners who call themselves econometricians spend more time cleaning the data than deriving new estimators, which is exactly why this subfield feels different from cross-sectional work. I usually open a new series by plotting the raw observations, the first differences, and the logarithmic returns alongside the autocorrelation and partial autocorrelation functions. That single triptych tells you whether the trouble is multiplicative seasonality, a level shift, or a trend that behaves like a random walk with drift. If all three plots look similar, you are probably overthinking the specification. If they diverge, go back and ask who produced the data and when the collection protocol changed.

Start with the estimator, then justify the model

Beginners reverse that order. They decide on SARIMA because a colleague used it, or because a software package recommends it by default, and then they spend two days tuning orders. I reverse it. I pick an estimation framework first—conditional sum of squares for quick exploration, maximum likelihood for final reporting, or Bayesian MCMC when the posterior is multimodal—and I keep the model class as loose as possible until the residuals pass a battery of checks. The reason is simple: order selection is an underdetermined problem when the sample is shorter than five hundred observations, which is true for most regional macro series. For a quarterly series with obvious seasonality, I start with a linear state-space representation that includes a stochastic trend, a seasonal component modeled as a set of four latent states, and an observation equation that can accommodate outliers through a mixture normal error. I estimate it with the Kalman filter and smoother. That approach lets me separate level movements from transient spikes, which is exactly where the ARIMA differencing operator tends to overcompensate. When I fit an equivalent SARIMA model for comparison, the Akaike information criterion usually favors the state-space version by two to four points, but the difference becomes negligible once I include structural breaks as regression dummies. The takeaway is not that state-space models are superior in general. It is that they expose assumptions that ARIMA hides inside the backshift polynomial.

A field note on differencing that cost me three weeks of revisions

In 2019 I modeled wholesale price indices for a commodity-exporting country. The ADF test rejected non-stationarity at the original level after two differences, so I fitted a SARIMA model with order selection guided by the extended autocorrelation function. Forecasts looked reasonable in-sample but diverged aggressively out-of-sample once the central bank changed its inflation targeting framework. The deeper problem was not the policy shift itself. It was that I had used automatic order selection while ignoring that differencing introduced a moving-average root close to unity in the opposite direction. That root created near-unit-root cancellation, which made parameter estimates unstable and standard errors misleadingly small. I caught it by examining the inverted MA polynomial and noticing that the largest root lay at 0.97. Switching to a model with an explicit break dummy and retaining the original series with a time trend reduced root mean squared error by eleven percent over the next eight quarters. This is the kind of practical Applied Econometric Time Series problem that does not appear in any tutorial but will stop your work dead if you ignore it. Box-Ljung tests on residuals are necessary but insufficient. They tell you whether linear dependence remains, but they do not warn you about conditional heteroskedasticity, regime switching, or heavy-tailed outliers that distort the likelihood. My minimum diagnostic set includes the Ljung-Box statistic for lags up to twenty-four, a standardized residual plot, the Ling and McCulloch test for ARCH effects, and a simple regime diagnostic that compares the residual variance in the first and second halves of the sample. If any of those flags fail, I do not immediately jump to a GARCH variant. I first check whether the failure is driven by a few influential observations or by a sustained shift in the data-generating process. One way to decide is to run a regression with dummy variables for each suspected outlier and retest. If the flags disappear, the problem is observational. If they persist, the problem is structural. I also run a small simulation-based calibration for every new model before trusting its confidence intervals. I generate pseudo-data from the fitted model, refit the model on each simulated series, and compare the empirical coverage of the intervals to the nominal ninety-five percent level. When I first did this on a bivariate error-correction model for exchange rates, the empirical coverage was seventy-eight percent because the likelihood surface was nearly flat in one direction. That insight prevented me from publishing a precision claim that would have fallen apart under modest misspecification. Simulation calibration takes about ten minutes for a moderate-sized dataset on a modern laptop, and it catches most of the pathological cases that asymptotic theory pretends do not exist.

Structural breaks are not optional; they are the default

Most introductory courses treat structural breaks as a special case. In applied work they are the norm. Oil crises, financial crashes, regulatory changes, pandemics, and even revisions to statistical methodologies produce discrete shifts that no stationary model can absorb without distortion. The standard approach is to test for an unknown break using Andrews or Bai-Perron procedures, then to include the estimated break dates as regressors or to allow coefficients to vary across regimes. I prefer the latter when I have at least two hundred observations per regime, because it preserves uncertainty about the break date. When the sample is smaller, I fix the break date using external information such as a known policy announcement and treat the date as given. Both approaches are defensible; the difference matters only for inference about the magnitude of the shift. If you ignore breaks, your serial correlation tests will be contaminated, your coefficient estimates will be biased toward zero, and your forecast intervals will be too narrow. I saw this repeatedly in commodity price work, where a single breakout event can mimic persistent memory. The workaround is to estimate a model with a smooth transition function, usually logistic, and to let the data choose the transition speed. That method is computationally heavier but it avoids the arbitrary discrete split that often misplaces the effective break by several periods. For quarterly macro data, the logistic transition typically requires an extra thirty to forty seconds of CPU time per estimation compared with a linear model, which is trivial relative to the gain in robustness.

Forecasting without self-deception

Forecast evaluation is where most applied time series projects either succeed quietly or fail loudly. The temptation is to optimize for in-sample fit or to report the model that looks prettiest on a diagnostic plot. That strategy works until you hand the model to a decision maker who will be held accountable when the forecast misses. I therefore hold my out-of-sample performance to three standards: mean absolute percentage error on the first horizon, continuous ranking probability score for the full forecast distribution, and a simple economic loss function that reflects the cost of overprediction versus underprediction. The loss function is often asymmetrical. Missing a recession by forecasting growth is cheaper than missing growth by forecasting recession in a setting where inventory adjustment costs are high. When I compare a SARIMA benchmark against a state-space alternative, I usually find that the state-space model wins on multi-step forecasts while the SARIMA model is competitive or slightly better on one-step forecasts. The reason is that the state-space representation separates signal from noise more cleanly, which helps when the forecast horizon exceeds the dominant autocorrelation time scale. For horizons shorter than four periods, the difference is often within the noise of the evaluation metric. That means you should not switch modeling frameworks solely for short-horizon improvement. You should switch when your decision process depends on longer horizons or when the cost of long-horizon error is disproportionate. I also recommend evaluating forecasts using a rolling window rather than a fixed-origin expansion, because economic time series are non-stationary by construction. A rolling window with a width of roughly three years for quarterly data captures recent dynamics without drowning in obsolete information. The trade-off is higher variance in the evaluation metric, but that is honest variance. Fixed-origin evaluations tend to be deceptively stable because they borrow strength from distant history that may no longer be relevant.

When time series models fail outright

No dynamic linear model is safe during a regime change that alters the variance structure faster than the model can adapt. This happened to me in 2022 when I was forecasting utility tariffs that were recalibrated monthly due to inflation indexing. The conditional heteroskedasticity was so strong that even a stochastic volatility specification produced confidence intervals that were too narrow by a factor of two. The practical fix was to combine the time series model with a Bootstrap agglomeration forecast, which resamples residuals within rolling windows and reconstructs predictive distributions empirically. That approach increased computation time by roughly a factor of five, but it corrected the interval undercoverage from eighty-two percent to ninety-three percent, which is close enough for operational use. If you are in a hurry, a simpler fix is to scale the prediction intervals by the square root of the ratio of recent to historical residual variance. That adjustment costs almost nothing and often restores acceptable coverage for modest volatility shifts. Another scenario where Applied Econometric Time Series hits a wall is when the series is short, noisy, and heavily censored. I worked on a public health series with frequent zero reports due to reporting lags and batch testing. Likelihood-based models treated zeros as genuine observations, which pulled the trend down artificially. The workaround was to model the reporting process explicitly using a zero-inflated component and to treat the latent series as the object of interest. That added interpretability but required careful priors when I moved to a Bayesian implementation. Frequentist practitioners can approximate the same idea with a two-stage generalized least squares estimator, though inference is less clean. The moral is that model choice should follow data structure, not convenience.

Practical workflow that keeps projects moving

I structure my work in four phases: exploration, baseline estimation, refinement, and evaluation. Exploration lasts no longer than two days for a familiar series and involves plots, summary statistics, and informal tests for unit roots and breaks. Baseline estimation uses a simple ARIMAX model with exogenous regressors for known events and a constant trend. Refinement introduces stochastic trends, seasonal components, and heteroskedasticity only if diagnostics demand it. Evaluation compares the refined model against the baseline using out-of-sample metrics and checks whether added complexity actually improves decision-relevant forecasts. This discipline prevents model proliferation, which is the most common source of regret in applied time series work. I keep a reproducible script that reads raw data, runs the baseline model, produces diagnostic plots, and writes a one-page summary with the key statistics and a recommendation. The script takes about fifteen minutes to run from start to finish on a standard dataset, which leaves most of the day for interpretation and communication. When a project becomes unusually messy, I revert to the script, adjust only the parts that failed, and re-run. That habit has saved me from losing context in complex estimation experiments. It also makes peer review less painful because reviewers can reproduce the baseline in minutes rather than days. If you are new to this area, start with quarterly macro data, fit a simple ARIMA with a time trend, evaluate out-of-sample forecasts over two years, and then add one complication at a time. Each addition should be justified by a diagnostic failure or by a measurable improvement in an economic loss function. If neither condition holds, drop the complication and move on. Applied Econometric Time Series rewards parsimony, skepticism, and a willingness to admit when a model has exhausted its usefulness.