Other Research Papers

Adversarial Density Forecast for Macro-financial Risks with Tae-Hwy Lee

An anchored conditional Wasserstein generative model for density forecasting in short macroeconomic samples, with theory and evidence on distributional accuracy and sensitivity to training randomness.

Abstract

Short estimation windows can make adversarial density forecasts sensitive to network size and training randomness. We study a conditional Wasserstein generative adversarial network with a penalty that matches the variance and five quantiles of simulated and observed outcomes within training blocks. We analyze the expected penalty with finitely many simulation draws. Its variance and median components bound average generated second moments and, under weak dependence and approximation conditions, yield prediction risk bounds that also control excess expected continuous ranked probability score (CRPS). After reducing network size, the penalty lowers Wasserstein error in all four simulation settings and average variation across fits by about 25 percent. A relativistic GAN provides competitive accuracy but greater variation across fits. In three macroeconomic applications, the penalty improves CRPS point estimates in five of six same-capacity comparisons; classical benchmarks remain strong and short-window tail coverage is poor.

When Should Forecasts Be Equally Weighted? Signal Recovery and Estimation Costs

Under Review, International Journal of ForecastingPDF

A comparison of equal and estimated forecast weights that separates recovered predictive signal from estimation costs, accounts for tuning and component estimation, and uses past out-of-sample losses to guide the choice.

Abstract

Whether estimated combination weights improve on equal weights depends not only on the signal in forecast differences, but also on how much of that signal the forecasting procedure recovers and at what cost. We express the expected mean squared forecast error (MSFE) gain in terms of signal recovery and estimation costs, giving a boundary for when a specified forecasting procedure improves on equal weighting. We complement this theoretical boundary with a decision rule based on past out-of-sample loss differences to choose between equal weights and the estimated weights. Under stationarity and weak dependence, with estimation windows fixed, the probability of choosing the lower-MSFE method tends to one as past evaluations accumulate, provided the two population MSFEs differ. Simulations and macroeconomic applications illustrate the comparison and the choices of equal and estimated weights.

Additional Working Papers

Generated Instruments with Validation Data with Xinran Liu

Working Paper

Validation data help distinguish robust conclusions about the yield response to FOMC news from conclusions that depend on instrument validity.

Abstract

Which conclusions about the yield response to Federal Open Market Committee (FOMC) news survive when the policy instrument is imperfect? We study nominal and real yields and inflation compensation around 92 FOMC meetings using an instrument constructed from changes in money-market rates around FOMC announcements. Because the instrument may also reflect nonpolicy news, we use historical benchmark data to quantify the resulting instrument error. Treating the instrument as valid, both nominal and real yields rise, with real yields rising more. Allowing the instrument to be imperfect, the evidence that both yields rise remains more robust than their ordering. This validation evidence therefore distinguishes which parts of the yield response continue to support a policy interpretation.

Mixed Bundling from Separate-Sales Data

Working Paper

We characterize when a mixed-bundling menu is optimal under revenue ambiguity using joint purchase tables from historical separate sales.

Abstract

A seller has sold two goods separately at two price levels and knows, for each level, the shares of buyers who bought neither good, one of them, or both. She evaluates a selling mechanism by its worst-case revenue over all value distributions consistent with these purchase tables. For a class of symmetric tables, we show that the mixed-bundling menu charging the higher historical price for either good and the sum of the two prices for both is optimal among all mechanisms if and only if two inequalities in the observed shares hold. The bundle discount must pay for itself, and serving the buyers the menu excludes must not be worth the information rents it creates. With data on each good's sales alone, separate sales would be optimal, so the gain comes from knowing which goods are bought together. The menu can also strictly outperform every randomization over separate sales and pure bundling, and its revenue need not be identified by the data.