Capstones, Data, and Reading Ladder

The common capstone standard

Choose either the finance or economics route. Both must produce an auditable chain:

The capstone is complete only when another learner can reconstruct what was known at each forecast origin and reproduce the main table from raw inputs.

Route A — finance

Choose one bounded question:

• Does a fixed GARCH specification improve one-day volatility or VaR forecasts over a rolling variance benchmark?
• Does a yield or valuation predictor improve an excess-return forecast after persistent-regressor and overlapping-horizon inference?
• Is a theoretically motivated spread stable enough to support an out-of-sample monitoring rule?
• Does a small VAR improve forecasts of returns, volatility, and liquidity without claiming unidentified shocks are causal?

Required evidence

1. adjusted-price, return, calendar, and execution conventions;
2. training, validation, and test endpoints fixed in advance;
3. a simple benchmark;
4. matrix derivation of the estimator or recursion;
5. point, volatility, or tail loss matched to the decision;
6. turnover, costs, liquidity, and stressed-period limitations where relevant;
7. no claim of profit unless an executable portfolio rule is actually evaluated.

Example boundary

Defensible:

For this liquid index and fixed 2018–2025 evaluation design, the GARCH forecast reduced average QLIKE relative to a constant-variance benchmark, but 1% exceptions remained clustered in the stressed subsample.

Not defensible:

GARCH predicts crashes and is the best model for markets.

Route B — economics

Choose one bounded question:

• Do financial conditions improve a real-time inflation or output-growth forecast over an autoregressive benchmark?
• Does a small mixed-frequency state-space model improve a current-quarter nowcast as releases arrive?
• Is a theory-motivated long-run relation stable enough for an error-correction model?
• How sensitive is a reported policy response to recursive ordering, lag length, and a local-projection specification?

Required evidence

1. reference, release, and vintage dates;
2. level, growth, seasonal, and aggregation conventions;
3. a real-time or explicitly retrospective estimand;
4. benchmark and pseudo-out-of-sample design;
5. cointegration rank or structural identification when claimed;
6. revision, regime, and lag sensitivity;
7. a clear separation between predictive and structural conclusions.

Example boundary

Defensible:

Using vintages available before each monthly origin, the mixed-frequency model reduced current-quarter GDP RMSE relative to an AR benchmark; most improvement arrived after labour-market releases, and results weakened before the structural break.

Not defensible:

Employment causes GDP because it improves the nowcast.

Deliverables

Expected effort is 25–35 hours across question design, data audit, derivation, implementation, evaluation, and revision. A text-only table is an acceptable alternative to a plot if it exposes the same values, units, labels, and uncertainty.

Suggested evaluation rubric

This rubric is a course-design recommendation, not an institutional policy.

Authoritative data starting points

Archive the exact input or retrieval recipe where licensing permits. “Downloaded from FRED” is not a complete provenance statement.

Reading ladder

Foundations carried from classical time series

• Brockwell and Davis, Time Series: Theory and Methods: stationary processes, linear prediction, ARMA, and spectra.
• Durbin and Koopman, Time Series Analysis by State Space Methods: matrix filtering, smoothing, likelihood, and missing observations.

Finance

• Engle (1982): ARCH and time-varying conditional variance.
• Bollerslev (1986): parsimonious GARCH persistence.
• Stambaugh (1999): small-sample bias with persistent return predictors.
• Campbell and Yogo (2006): predictive inference near a unit root.
• Fissler, Ziegel, and Gneiting: comparative backtesting of VaR and expected shortfall.

Economics and structural dynamics

• Johansen (1991): reduced-rank cointegration testing.
• Jordà (2005): local-projection impulse responses.
• Giannone, Reichlin, and Small (2008): real-time information flow and nowcasting.
• Plagborg-Møller and Wolf (2021): population equivalence of unrestricted LP and VAR impulse responses.

Forecast evaluation

• Newey and West: positive-semidefinite HAC covariance.
• Diebold and Mariano (1995): predictive-accuracy comparison.
• Giacomini and White (2006): conditional predictive ability.

Recent extensions

• Düker et al. (2024/25): a modern review of VARMA identification and estimation.
• Zheng (2024/25): interpretable high-dimensional infinite-order VAR modelling.
• Fosten and Gutknecht (2026): simultaneous predictive-accuracy testing across multiple horizons.

The recent papers extend rather than replace the foundations. A capstone should use a newer method only when the data size, estimand, and evaluation design require it.

Final defence questions

Be ready to answer:

1. What exactly was observable at the forecast origin?
2. Which mathematical assumption is doing the most work?
3. Why is the benchmark credible?
4. What changes if the horizon or loss changes?
5. Is the main result descriptive, predictive, or structural?
6. Which failure would reverse the decision?
7. Which part belongs to finance or economics rather than generic time-series theory?

Return to the course guide or compare the applied route with Classical Time Series.