Erken kri̇z uyarici si̇stemi̇: Makroekoni̇mi̇k i̇li̇şki̇ler ve nomi̇nal ve reel deği̇şkenler arasindaki̇ geci̇kme süresi̇

After most recent 2008 Global Recession, the interest in early crisis warning (EWS) system has been rekindled. Majority of existing EWS literature has short-run implications and it mainly focuses on currency crashes and financial crises that represent nominal aspects of the economy. It neglects real aspects of the economy and other types of imbalances such as macroeconomic crises, speculative bubbles and crashes, international financial crises, and wider economic crises. Moreover majority of existing EWS models failed to predict 2008 Global recession which leaves plenty rooms for research studies. In this respect, this thesis attempt to establish a brand-new EWS system that improves existing ones in several patterns. First, it offers more generalized outlook by accounting various types of financial, economic, banking, and currency crises by using a set of real and nominal macroeconomic indicators. Second, it considers the largest sample size for EWS by covering 41 countries from developed, emerging, and frontier markets. While the existing literature mainly studies few developed countries such as the US, the UK, and the EU. Third, the existing literature largely uses static models that ignore long-run implications. This thesis employs more robust methodology of Distributed Lag Logit technique (DLL) under Almon's (1965) approach in multivariate logit framework that wields dynamics aspects as well. This specific methodology helps us to capture cumulative and delayed effects of all indicators accurately. In other words, it integrates often neglected long-run impacts of leading indicators on crisis. Fourth, the study takes into account most prominent leading indicators alongside with few new indicators that have not been used before. The selection of appropriate leading indicators of economic and financial turmoil also plays a crucial role in modelling EWS. However, the selection mainly depends on how the term "crisis" is defined. Unlike the literature that often defines the crisis using self-built indexes, the thesis considers it as shrinks in real output straight-forwardly. In this respect, it attempts to integrate real and nominal variables in the EWS model by utilizing empirically proven macroeconomic relations such as Phillips Curve and Okun's Law. We show that these empirical relations emerge early signals for upcoming imbalances. In other words, notional behaviour of the Phillips curve and Okun's law changes during two states of economy: Tranquil periods, when economic fundamentals are largely sound and sustainable; and Recessionary periods, when economic variables go through an adjustment process before reaching a more sustainable level or growth path. Fifth, the thesis also offers two brand-new techniques for optimal threshold selection, namely, False-to-True ratio (ϕ) and Risk Aversion rate (Ω) due problems in existing threshold techniques. The literature largely uses KLR Noise-to-Signal ratio as a threshold selection technique in order to gauge signals. However, it fails to distinguish consequences of "false alarms" (type II error - cost of taking pre-emptive action in response to a false alarm for non-crisis events) and "missed alarms" (type I error - cost of missed alarm for upcoming actual crisis). Therefore, we regard noise-to-signal ratio as risky threshold technique. Moreover, when probability of type II error is zero, the KLR's noise-to-signal ratio will derive zero notwithstanding to probability of type I error (there will be no difference whether it is 0.01 or 0.99). In other words, the noise-to-signal technique prefers to undertake more missed signal rates for the sake of nulling false signal rate. Thus, this technique does not primarily aim to predict and prevent potential crisis, but aims to minimize false alarms for non-crisis events. This is not in line with objectives of EWS models. A proper threshold selection technique should help EWS model to minimize missed crisis events (or maximize predicted crisis events). In this respect, we evaluate our DLL EWS model under four different threshold techniques comparatively including our two newly proposed techniques. As a result, we show that performance of our model can be optimized with our new threshold selection techniques which derive better prediction rates comparing to existing techniques. Averagely, our model correctly predicts 96.36% (1271 crisis quarters) of 1319 total crisis events (quarters) in advance, while issuing only 12.32% (492 tranquil quarters) false signals during 3995 tranquil periods.