كنفرانس بين المللی مهندسی صنايع و سيستم ها

صفحه اصلی / دوازدهمین كنفرانس بين المللی مهندسی صنايع و سيستم ها

A Hybrid HDBSCAN-K-Means Framework for Three-Regime De Facto Exchange Rate Classification

نویسندگان :

Hooman Fakhimi¹ Orod Ahmadi² Ashkan Mozdgir³

1- دانشگاه خوارزمی 2- دانشگاه خوارزمی 3- دانشگاه خوارزمی

کلمات کلیدی :

behavioral taxonomy،reserve intervention،density screening،movement index،country-year analysis،unsupervised learning

چکیده :

Classifying exchange-rate regimes from observed behavior is difficult because crisis years, many zero-valued indicators, and sparse country-year observations can distort clustering. This paper proposes a hybrid unsupervised method for identifying three broad de facto regimes. HDBSCAN first identifies the density-supported core and flags low-density observations, while K-means produces an interpretable and complete classification. The method retains the Levy-Yeyati-Sturzenegger (LYS) indicators of exchange-rate movement, within-year dispersion, and foreign-exchange reserve intervention. A zero-aware log-floor transformation and robust scaling limit the influence of extreme values, and the two highly correlated movement measures are combined to avoid duplicate weighting. K-means is fitted to the core in a two-dimensional movement-intervention space. Every observation is then assigned to the nearest centroid while low-density and borderline cases remain separately identified. Applied to 6,963 country-year observations for 186 economies over 1974-2021, the method identifies pegged/low-flexibility, managed/intermediate, and floating/high-flexibility regimes. The full-sample silhouette coefficient is 0.502, the mean country-block bootstrap adjusted Rand index is 0.939, and one-year regime persistence is 0.830. Agreement with the LYS three-way labels is reported only as external concordance (ARI=0.440) and is not used for estimation or tuning. The results indicate a stable and interpretable partition that makes classification uncertainty explicit.