Türkiye Finansal Piyasalarında Gerçek ve Sentetik Verilerle Risk Modelleme Üzerine Karşılaştırmalı Bir Analiz

Öz

Bu çalışma, Türkiye finansal piyasalarındaki risk modellemesini gerçek ve sentetik veriler üzerinden karşılaştırmalı olarak incelemektedir. 2015-2024 dönemine ait BIST 100 endeksi, döviz kurları (USD/TRY, EUR/TRY) ve makroekonomik göstergeler (politika faizi, enflasyon, sanayi üretimi, işsizlik) verileri kullanılarak log-getiriler hesaplanmış; eksik veriler (%5 oranında) lineer interpolasyonla tamamlanmıştır. Sentetik veriler, Denoising Diffusion Probabilistic Models (DDPM) ve TimeGrad modelleriyle (100 epoch, 0,001 öğrenme oranı, 32 batch boyutu) üretilmiş; IBM SPSS Statistics ile tanımlayıcı istatistikler, otokorelasyon, korelasyon, regresyon, Value at Risk (VaR) ve stres testi analizleri yapılmıştır. Gerçek veriler yüksek volatilite (std. sapma=0,0955), pozitif çarpıklık (4,122) ve yüksek basıklık (32,781) sergilerken; sentetik veriler daha yüksek volatilite (std. sapma=0,1067), pozitif çarpıklık (2,336) ve yüksek basıklık (22,234) göstermiştir. Piyasa fiyatları ile döviz kurları ve faiz oranları arasında güçlü pozitif korelasyonlar (r≈0,8-0,96) tespit edilmiş; log-getiriler makroekonomik değişkenlerle zayıf ilişkiler sergilemiştir. Paired-Samples T Testi, gerçek ve sentetik getiriler arasındaki farkın istatistiksel olarak anlamlı olmadığını (p=0,624) doğrulamış; regresyon modeli ise bağımsız değişkenlerin getirileri yeterince açıklamadığını (R²=0,068, p=0,337) göstermiştir. Bulgular, sentetik verilerin gerçek piyasa dinamiklerini kısmen taklit edebildiğini ancak volatilite, çarpıklık, basıklık ve bağımlılık yapılarında yetersiz kaldığını ortaya koymaktadır. Türkiye gibi değişken ekonomilerde sentetik verilerin risk yönetimindeki sınırlılıklarını ve geliştirme gerekliliklerini vurgulamaktadır.

Anahtar Kelimeler

• sentetik veri
• volatilite
• risk modelleme
• BIST 100
• Türkiye

A Comparative Analysis of Risk Modeling Using Real and Synthetic Data in Turkish Financial Markets

Öz

Abstract: This study conducts a comparative analysis of risk modeling in the Turkish financial markets using real and synthetic data. Log-returns were calculated based on data from the 2015–2024 period, including the BIST 100 index, exchange rates (USD/TRY, EUR/TRY), and macroeconomic indicators (policy interest rate, inflation, industrial production, unemployment). Missing data (5%) were completed using linear interpolation. Synthetic data were generated using Denoising Diffusion Probabilistic Models (DDPM) and TimeGrad models (100 epochs, 0,001 learning rate, 32 batch size). Descriptive statistics, autocorrelation, correlation, regression, Value at Risk (VaR), and stress testing analyses were performed using IBM SPSS Statistics. While real data exhibited high volatility (std. dev = 0,0955), positive skewness (4,122), and high kurtosis (32,781), synthetic data demonstrated even higher volatility (std. dev = 0,1067), positive skewness (2,336), and high kurtosis (22,234). Strong positive correlations (r ≈ 0,8–0,96) were found between market prices, exchange rates, and interest rates, whereas log-returns showed weak associations with macroeconomic variables. A Paired-Samples T-Test confirmed that the difference between real and synthetic returns was not statistically significant (p = 0,624), while the regression model indicated that the independent variables did not adequately explain returns (R² = 0,068, p = 0,337). The findings reveal that while synthetic data can partially replicate real market dynamics, they fall short in capturing volatility, skewness, kurtosis, and dependency structures. The study emphasizes the limitations of synthetic data in risk management and the need for improvement, particularly in highly volatile economies like Turkey.While real data exhibited high volatility (std. dev = 0,0955), positive skewness (4,122), and high kurtosis (32,781), synthetic data demonstrated even higher volatility (std. dev = 0,1067), positive skewness (2,336), and high kurtosis (22,234). Strong positive correlations (r ≈ 0,8–0,96) were found between market prices, exchange rates, and interest rates, whereas log-returns showed weak associations with macroeconomic variables. A Paired-Samples T-Test confirmed that the difference between real and synthetic returns was not statistically significant (p = 0,624), while the regression model indicated that the independent variables did not adequately explain returns (R² = 0,068, p = 0,337). The findings reveal that while synthetic data can partially replicate real market dynamics, they fall short in capturing volatility, skewness, kurtosis, and dependency structures. The study emphasizes the limitations of synthetic data in risk management and the need for improvement, particularly in highly volatile economies like Turkey.

Anahtar Kelimeler

• synthetic data
• volatility
• risk modeling
• BIST 100
• Turkey

Kaynakça

1. 1] J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Proc. Adv. Neural Inf. Process. Syst., 2020, vol. 33, pp. 6840–6851, [Online].
2. [2] P. Jorion, Value at Risk: The New Benchmark for Managing Financial Risk, 3rd ed. New York, NY, USA: McGraw-Hill, 2007.
3. [3] J. Danielsson, Financial Risk Forecasting: The Theory and Practice of Forecasting Market Risk with Implementation in R and Matlab. Hoboken, NJ, USA: John Wiley & Sons, 2011, doi: 10.1002/9781119205869.
4. [4] A. Bouveret, “Cyber Risk for the Financial Sector: A Framework for Quantitative Assessment,” IMF Working Paper No. 18/143, 2018, doi: 10.5089/9781484360750.001.
5. [5] N. Patki, R. Wedge, and K. Veeramachaneni, “The Synthetic Data Vault,” in Proc. IEEE Int. Conf. Data Sci. Adv. Analyt. (DSAA), 2016, pp. 399–410, doi: 10.1109/DSAA.2016.49.
6. [6] D. B. Rubin, “Statistical disclosure limitation,” J. Official Statist., vol. 9, no. 2, pp. 461–468, 1993.
7. [7] I. J. Goodfellow et al., “Generative adversarial nets,” in Proc. Adv. Neural Inf. Process. Syst., 2014, vol. 27, [Online]. Available: https://papers.nips.cc/paper/5423-generative-adversarial-nets.
8. [8] P. Glasserman, Monte Carlo Methods in Financial Engineering, vol. 53. New York, NY, USA: Springer, 2004, doi: 10.1007/978-0-387-21617-1.

9. [9] E. Özsöz, “What determines return risks for bank equities in Turkey?,” Center Int. Policy Stud., Fordham Univ., New York, NY, USA, 2011.
10. [10] B. D. Yıldırım, Y. Coskun, O. Çağlar, and K. Yildirak, “How Dangerous is the Counterparty Risk of OTC Derivatives in Turkey?,” Sermaye Piyasasi Dergisi, vol. 10, no. 2, pp. 70–79, 2012.
11. [11] R. Matkovskyy, “To the problem of financial safety estimation: The index of Financial Safety of Turkey,” 2013.
12. [12] I. Dvorski Lacković, V. Kovšca, and Z. Lacković Vincek, “Framework for big data usage in risk management process in banking institutions,” in Proc. Central Eur. Conf. Inf. Intell. Syst., 2016, pp. 49–54.
13. [13] G. Visani, G. Graffi, M. Alfero, E. Bagli, F. Chesani, and D. Capuzzo, “Enabling Synthetic Data adoption in regulated domains,” in Proc. IEEE 9th Int. Conf. Data Sci. Adv. Analyt. (DSAA), 2022, pp. 1–10, doi: 10.1109/DSAA54385.2022.10044960.
14. [14] J. Hu and C. M. Bowen, “Advancing Microdata Privacy Protection: A Review of Synthetic Data,” arXiv preprint arXiv:2308.00872, 2023, doi: 10.48550/arXiv.2308.00872.
15. [15] C. Little, M. Elliot, and R. Allmendinger, “Federated learning for generating synthetic data: A scoping review,” Int. J. Population Data Sci., vol. 8, no. 1, p. 2158, 2023, doi: 10.23889/ijpds.v8i1.2158.
16. [16] V. B. Vallevik et al., “Can I trust my fake data–A comprehensive quality assessment framework for synthetic tabular data in healthcare,” Int. J. Med. Informat., vol. 185, p. 105413, 2024, doi: 10.1016/j.ijmedinf.2024.105413.
17. [17] Y. Xia, C. H. Wang, J. Mabry, and G. Cheng, “Advancing retail data science: Comprehensive evaluation of synthetic data,” arXiv preprint arXiv:2406.13130, 2024, doi: 10.48550/arXiv.2406.13130.
18. [18] Rasul, Kashif, et al. "Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting." International conference on machine learning. PMLR, 2021.