10
Archived Theses
0
DOIs Assigned
0%
DOI Rate
Discipline
Gelişmekte olan piyasa döviz opsiyonlarının zımni volatilite yüzeyinin tahmini: Ampirik bir analiz
This study aims to evaluate the predictive power of time series and machine learning models, namely Autoregressive (AR), Principal Component Analysis-Vector Autoregression (PCA-VAR), Principal Component Regression (PCR), and Feedforward Neural Networks (FNN), in modeling the implied volatility surfaces of five emerging market currencies (TRY, INR, MXN, ZAR and BRL against USD). The research assesses model performance using the Root Mean Square Error (RMSE) metric under both the Expanding Window and Rolling Window frameworks. The findings indicate that Principal Component Regression (PCR) and FNN models deliver similarly high precision, particularly for currencies exhibiting lower volatility levels. The study underscores the critical importance of model selection in financial forecasting and suggests that incorporating country-specific macroeconomic or geopolitical factors may further influence model outcomes.
Karbon emisyonlarının konut gayrimenkul fiyatları üzerindeki etkisi: Birleşik Krallık piyasasından kanıtlar
This paper investigates the impact of CO2 emissions on property prices, moving beyond the conventional notion of a single "green premium" measured solely by energy efficiency. Using a large panel dataset of property transactions and their associated Energy Performance Certificates (EPCs), I employ a fixed effects regression framework to examine how the Energy Efficiency Index (EEI) - a measure of running costs - and the Environmental Impact Index (EII) - a measure of CO2 emissions - independently influence transaction prices. The findings reveal that the market distinguishes between these two attributes: while the EEI is valued for its cost-saving potential, the EII is priced as a proxy for unobserved property characteristics, such as amenities and high-end finishes. The valuation of these attributes exhibits substantial heterogeneity across the market. Houses, with their higher operating costs, place greater value on the EEI, whereas flats exhibit a stronger premium for the EII due to its quality signals. Tenure-based analysis provides support for the split incentive problem, and in the hyper-competitive London market, the premium for environmental impact is significantly muted, likely due to the dominance of fundamental factors such as location. These findings contribute to the literature by demonstrating that cost-saving and environmental performance are distinct value drivers, with important implications for policy-makers, investors, and real estate professionals.
Comparison of different customer segmentation models for the financial sector versus the turkish factoring sector and a clustering model proposal for a sme focused factoring company
Faktoring işlemi, ticari alışverişten doğan vadeli bir alacağın temlik edilmesi yolu ile bir faktoring kuruluşuna devredilmesi işlemidir. Faktoring yapmayı tercih eden firmalar, yapmış oldukları ticaretlerinden doğan alacaklarının tahsilatının vadesini beklemek yerine, ticarete konu bir ödeme araç olan vadeli çek, ticaret faturası ve faktoring sözleşmesi gibi dokümanlar ile aracı faktoring kuruluşuna belirli bir iskonto karşlığında devretmekte ve nakit akışı ihtiyaçlarını düzenlemeyi amaçlamaktadırlar. Türkiye'de kobiler tarafından yaygın bir şekilde kullanılan çek, üzerinde yazan ileri tarihli vade nedeniyle dünyadaki genel kullanımından farklılaşmaktdır. Literatürde bankalar için yapılmış bir çok segmentasyon çalışması mevcuttur ancak kobi sektörüne yönelik bir faktoring müşterisi segmentasyon çalışması bulunmamaktadır. Bu çalışmada faktoring sektöründe faaliyet gösteren bir Türk faktoring firması datası kullanılarak iki aşamalı bir müşteri segmentasyon çalışması yapılmıştır. Çalışmada kullanılan K-Means ve DBSCAN algoritmaları karşılaştırıldığında, K-Means algoritmasının, kümeler içerisindeki müşteri adetlerinin dağılımı anlamında daha dengeli olduğu gözlemlenmiştir. Öte yandan, DBSCAN modeli tarafından üretilen kümeler, özellikle aykırı kümeler için daha homojendir. Her bir gözlemin Mahalanobis mesafesi ile tanımlanan homojenlik, küme içerisinde daha az aykırı gözlem olduğu zaman etkinlik anlamına gelmektedir. DBSCAN modeli, model parametreleri uygun şekilde kalibre edilirse az sayıda müşteri içeren birçok homojen küme oluşturabilir. Aksine az sayıda küme tercih edilirse DBSCAN algoritmasının homojenliği bozulmakta ve K-Means modeli daha dengeli kümeleme sonuçları vermektedir.
BIST 100 endeks getirisini tahmin etmede teknik indikatörlerin önemi
This paper presents a logistic regression-based approach for predicting the direction of the Borsa Istanbul (BIST) index's price movements using technical indicators. Historical price data from 2005 to 2021 is utilized, and five thresholds are determined from -2% to 2% based on the daily or weekly returns of the BIST index. Binary variables which utilized in logistic regression as dependent variable are defined according to the thresholds, and logistic regression is applied using annual or semiannual training data. Daily models perform better than weekly models in logistic regression, and On Balance Volume (OBV) and Average Directional Index (ADX) are the indicators within the best statistical results. 0% threshold models are best in accuracy of prediction (approximately half of prediction is accurate). Weekly models are better than daily models, and annual models are better than semiannual models in accuracy. Some of the models outperformed the BIST index in cumulative return (daily cumulative return is 4,84; weekly is 4,48), with two models having approximately three times the cumulative return of the BIST index namely Daily-1% threshold-6 month's cumulative return is 13,35; Weekly-2% threshold-12 month's is 12,96. In summary, the results show that technical indicators can be successful in predicting stock or index returns depending on the training period, time period, and the combination of technical indicators used.
Boğaz'da sektirmek: Ultra yüksek frekansta sıçrama sonrası getiriler
This study rigorously investigates intraday jumps in highly liquid stocks listed on Borsa Istanbul (BIST) using two common jump test methods and ultra-high frequency data. Additionally, three purely price-based jump methods, referred to as Price Skipping, were developed to improve jump detection accuracy. The findings reveal the presence of reversal post-jump returns, irrespective of the jump direction. A profitable trading strategy was established, exploiting this market behavior by taking long positions after down jumps and short positions after up jumps. The strategy's performance was compared across different jump detection methods and confidence intervals. Logistic regression analyses demonstrate a significant negative correlation between relative tick size and jump occurrence. Additional analyses using Ordinary Least Squares (OLS) and Locally Estimated Scatterplot Smoothing (LOESS) consider factors such as relative tick size, volume, pre-jump returns, sector affiliation, jump timing, index jumps, and alternative jump detection methods. Importantly, a positive correlation was discovered between relative tick size and the magnitude of reversal post-jump returns. Another significant finding suggests a positive correlation between trading volume during jumps and the magnitude of reversal post-jump returns. These results indicate that high volumes, potentially resulting from large market orders, induce stock price jumps involving multiple ticks, and when the relative tick size is large, these movements trigger an overreaction, subsequently leading to a reversal within a defined time period. These findings suggest that the tick size is excessive for the small priced stocks and suboptimal for the middle range. Therefore, it is recommended that Borsa Istanbul's management consider revising their tick size policy to enhance market liquidity, considering the implications of this study.
İleri düzey işlem maliyeti fonksiyonlarını entegre eden gelişmiş dinamik çok dönemli portföy optimizasyonu
This study investigates the impact of transaction costs, often overlooked in multi-period portfolio optimization. Traditional portfolio optimization models tend to underestimate the influence of transaction costs such as bid-ask spreads, market impact costs, and commission fees by assuming these costs to be constant. However, these costs, which can vary over time, were analyzed using daily data from the most liquid ETFs between January 2009 and May 2024 to identify the exogenous variables (e.g. VIX, trading volume) that best predict transaction cost function. These variables were integrated into advanced forecasting models, including ARIMA-X, SARIMA-X, and GARCH-X, to improve transaction cost forecasting accuracy. The performance of these models was compared using metrics such as RMSE and MAE, with the SARIMA-X model achieving the lowest RMSE and MAE values, indicating the most accurate predictions. The transaction costs predicted by the SARIMA-X model were then integrated into the multi-period portfolio optimization framework developed by Boyd [4]. The findings of this study indicate that accurately forecasting these time-varying factors plays an important role in modeling transaction costs and that incorporating these costs into the optimization process enhances multi-period portfolio performance.
Gelişmekte olan piyasalarda yerel para cinsinden tahvil risk primleri: İleri düzey makine öğrenmesi tekniklerinden elde edilen içgörüler
Understanding the determinants of local currency bond risk premia is crucial for emerging market investors and policymakers. This study investigates the determinants of local currency bond risk premia in six emerging markets—Brazil, Hungary, Poland, Thailand, South Africa, and Turkey—through the application of advanced machine learning techniques. The analysis first utilizes yield curve-based variables, including forward rates, forward-spot spreads, and term premia. Subsequently, inflation, implied foreign exchange (FX) volatility, and macroeconomic indicators are incorporated to assess their individual and combined effects on predictive accuracy. The results highlight distinct regional patterns in the drivers of excess bond returns. Findings reveal that in Brazil, Hungary, Poland, and Thailand, yield curve-based variables—especially forward rates and term premia—exhibit strong predictive power, while macroeconomic factors and FX volatility offer limited value. In contrast, Turkey and South Africa display a fundamentally different structure, where inflation, macroeconomic indicators, and implied FX volatility serve as the primary predictors, and yield curve variables fail to explain bond risk premia. A diverse set of machine learning algorithms—including linear regression, principal component analysis (PCA), partial least squares (PLS), neural networks, random forests, XGBoost, and extremely randomized trees—were employed. Country-specific analysis reveals that algorithmic performance varies significantly across emerging markets, reflecting distinct economic structures and dominant predictive factors. In Hungary and Brazil, Neural Networks yield the highest predictive accuracy In Turkey, XGBoost delivers optimal results. Poland stands out with OLS + PCA demonstrating the utility of linear dimensionality reduction in a structured market environment. In Thailand, both Neural Networks and PCA-applied models exhibit nearly equivalent performance. For South Africa, the most effective predictions are achieved using an Extremely Randomized Trees model. This research highlights the importance of regional differences in the drivers of local currency bond risk premia and demonstrates the value of combining diverse data sources with advanced machine learning techniques. These findings provide valuable insights for policymakers, investors, and financial institutions seeking to refine risk assessment frameworks and improve strategic decision-making in emerging bond markets.
Başlıklardan öngörülere: VADER kullanarak duygu temelli bir piyasa endeksi oluşturmak
This thesis introduces a sentiment-based decision-support model that transforms daily financial news headlines into structured quantitative signals for investment analysis. The study focuses on 43 actively traded stocks, primarily listed on the S\&P 500 index, using a two-year dataset of daily headlines collected from publicly available sources. Sentiment scores are generated through a custom framework based on the VADER lexicon and converted into time series. These scores are then evaluated in three main areas: assessing predictive power using Ordinary Least Squares (OLS) regression models, ranking assets for sentiment-driven portfolio strategies, and serving as explanatory variables in asset pricing models such as \ac{CAPM} and Fama-French. Results show that the model provides both statistical and practical value, outperforming market benchmarks under basic portfolio construction methods.
Türkı̇ye getı̇rı̇ eğrı̇sı̇nı̇n derı̇n öğrenme ve ekonometrı̇k modeller kullanılarak tahmı̇n edı̇lmesı̇: Karşılaştırmalı bı̇r analı̇z
This thesis investigates the modeling and forecasting of the Turkish Treasury bond yield curve using both traditional time series models and deep learning approaches. The study focuses on evaluating the performance of models such as Autoregressive (AR), Random Walk (RW), the Dynamic Nelson-Siegel (DLNS) model with xed and time-varying de cay parameters, and Long Short-Term Memory (LSTM) networks. Daily zero-coupon yield data between July 2012 and February 2025 were used, covering maturities rang ing from 3 months to 10 years. Forecasting performance was assessed using an out-of sample expanding window framework for different horizons. Root Mean Squared Error (RMSE) was employed as the main accuracy metric, and the Diebold-Mariano (DM) test was applied to determine whether performance differences across models were statisti cally signicant. The empirical ndings show that while simpler models such as AR and RW perform well in the short term, their forecasting accuracy deteriorates as the horizon increases. In contrast, LSTM models—particularly those trained on latent factors derived from the Nelson-Siegel framework—demonstrate superior generalization performance in medium- and long-term forecasts. Notably, the DLNS model with a time-varying λ also outperforms benchmark models at longer horizons. These results highlight the value of incorporating structured factor dynamics and deep learning techniques for forecasting complex nancial time series such as the yield curve.
BİST Sürdürülebilirlik Endeksinde yer almanın ve kurumsal yönetim ilkelerinin şirketlerin finansal performansı üzerindeki etkisi
This study explores how corporate governance performance and sustainability index participation relate to the financial performance of companies listed on Borsa Istanbul (BIST). Utilizing a firm-level panel dataset covering the 2014–2023 period, the analysis was conducted using the Fama-MacBeth regression approach. This method involves estimating cross-sectional regressions separately for each year and averaging the resulting coefficients over time, thereby accounting for firm-level heterogeneity while capturing time-specific effects. The models include Return on Equity (ROE), Return on Assets (ROA), and EBITDA (Earnings Before Interest, Taxes, Depreciation and Amortization) to Total Assets as accounting measures, while raw and alphas are considered as market-based outcomes. Independent variables encompass corporate governance performance, firm size (measured by market capitalization), systematic risk (beta), book-to-market (B/M) ratio, inclusion in the Sustainability Index Dummy (SIC), and prior-period returns. A total of twelve model specifications were employed, allowing for a comparative analysis of how the same explanatory variables relate to different dimensions of financial performance. This design provided a broader perspective on the ESG (Environmental, Social and Governance) performance nexus and helped assess the robustness of the results. The results show that market capitalization is generally associated with improved financial performance, especially in accounting-based models, although this effect is not consistently significant across all specifications. Corporate governance performance, on the other hand, displays mixed results. In several models, particularly those focused on ROE and EBITDA, higher governance scores are linked to lower short-term performance, which might reflect adjustment costs or lagged impacts. Although the cumulative return analysis suggests that the sustainability index appears to have outperformed the BIST100 over the sample period, the differences in returns are not statistically significant. Moreover, the SIC variable does not exhibit a consistent or statistically significant relationship with financial performance indicators, implying that inclusion in the sustainability index does not inherently result in superior short-term financial outcomes. These results may reflect the short-run costs associated with governance and ESG implementation. Measures such as reporting compliance, policy restructuring, and stakeholder transparency often lead to a rise in operational expenses, which can weigh on profitability and share performance in the near term. Nevertheless, these practices should not be assessed solely through a short-term lens. ESG-oriented strategies are often aimed at reducing long-term risks and improving investor confidence, which may contribute positively to firm value over time. Keywords: Corporate Governance Performance, ESG, Financial Performance, BIST Sustainability Index