Qu speaker series 14: Synthetic Data Generation in Finance
0 likes177 views
The document discusses a speaker series on synthetic data generation in finance, hosted by Sri Krishnamurthy from Quantuniversity, featuring Stefan Jansen who explains generative adversarial networks (GANs) and their application in generating realistic financial time series data. It outlines the motivations for using synthetic data in finance, introduces the TimeGAN architecture, and presents methods for evaluating the quality of generated data. The conclusion indicates that TimeGAN can create synthetic data that mimics actual stock prices, suggesting potential for further development in the field.
Qu speaker series 14: Synthetic Data Generation in Finance
- 1. Qu Speaker Series Synthetic Data Generation in Finance 2020 Copyright QuantUniversity LLC. Hosted By: Sri Krishnamurthy, CFA, CAP sri@quantuniversity.com www.qu.academy 11/18/2020 Online https://quspeakerseries14.splashthat.com/
- 2. 2 QuantUniversity • Boston-based Data Science, Quant Finance and Machine Learning training and consulting advisory • Trained more than 1000 students in Quantitative methods, Data Science and Big Data Technologies using MATLAB, Python and R • Building a platform for AI and Machine Learning Exploration and Experimentation
- 3. 3 For registration information, go to https://QuFallSchool.splashthat.com
- 5. 5
- 6. 6
- 7. 7 Speakers Stefan is the founder and CEO of Applied AI. He advises Fortune 500 companies, investment firms, and startups across industries on data & AI strategy, building data science teams, and developing end-to-end machine learning solutions. Before his current venture, he was a partner and managing director at an international investment firm, where he built the predictive analytics and investment research practice. He was also a senior executive at a global fintech company with operations in 15 markets, advised Central Banks in emerging markets, and consulted for the World Bank. He holds Master's degrees in Computer Science from Georgia Tech and in Economics from Harvard and Free University Berlin, and is a CFA Charterholder.
- 8. SYNTHETIC DATA GENERATION IN FINANCE WITH TIMEGAN & TENSORFLOW STEFAN JANSEN
- 9. SYNTHETIC DATA FOR FINANCE AGENDA ▸ What are Generative Adversarial Networks (GANs)? ▸ Why could synthetic data for finance be useful? ▸ How can we build a TimeGAN using TensorFlow 2? ▸ How should we evaluate the quality of the results? GANs are the most interesting idea in ML in the last 10 years. Yann LeCun
- 10. SYNTHETIC DATA FOR FINANCE - INTRO WHAT ARE GENERATIVE ADVERSARIAL NETWORKS? ▸ Ian Goodfellow, et al. (NeurIPS 2014): learn a generative model via adversarial process that simultaneously trains two models (~minimax 2-player game)
- 11. SYNTHETIC DATA FOR FINANCE - INTRO WHAT ARE GENERATIVE ADVERSARIAL NETWORKS? ▸ Wave of research, over 500 different GAN architectures by 2018 ▸ From fake celebrities to style transfer, music generation and more ▸ Medical Time Series: Recurrent (Conditional) GAN - Esteban, et al. (2017) ▸ Generate real-valued time-series data as from an Intensive Care Unit ▸ RNN in generator and the discriminator, both conditioned on auxiliary information about the state of the patient ▸ Train early warning system on synthetic data, test on real data => only minor degradation
- 12. SYNTHETIC DATA FOR FINANCE - MOTIVATION WHY SYNTHETIC DATA FOR FINANCE? ▸ Financial data is relatively scarce (compared to web-scale images etc) and does not grow at the same speed as sources in other domains ▸ Limited training data availability increases the risk of model and backtest overfitting ▸ But how can we generate data that reflects the temporal dynamic of financial market time series?
- 13. SYNTHETIC DATA FOR FINANCE - TIMEGAN TIME-SERIES GAN - JINSUNG YOON ET AL. (GOOGLE / CAMBRIDGE), NEURIPS 2019 ▸ Goal: create a good generative model that preserves temporal dynamics so that new sequences respect the relationships ▸ between variables ▸ across time ▸ Generate realistic data by combining “the flexibility of the unsupervised paradigm with the control afforded by supervised training”. ▸ Approach: learn an embedding space (think word2vec), optimized with both supervised and adversarial objectives to encourage the network to adhere to the dynamics of the training data during sampling (see paper for the math…)
- 14. SYNTHETIC DATA FOR FINANCE - TIMEGAN TIMEGAN ARCHITECTURE: 2 NETWORKS, 4 COMPONENTS, 3 LOSS FUNCTIONS 1. Unsupervised adversarial loss on real and synthetic sequences 2. Supervised loss (minimized by training both the embedding and generator networks) captures the stepwise conditional distributions in the data. 3. The embedding network reduces the dimensionality of the adversarial learning space, assuming that temporal dynamics are driven by fewer, lower-dimensional factors.
- 15. SYNTHETIC DATA FOR FINANCE - TIMEGAN TIME-SERIES GAN - LET’S TAKE A LOOK AT THE CODE! ▸ Port original implementation from TF1 to TF2, attempt to simplify ▸ Minor changes: instead of OHLCV, use six different daily closing prices ▸ https://github.com/stefan-jansen/synthetic-data-for-finance
- 16. SYNTHETIC DATA FOR FINANCE - RESULTS EVALUATION EVALUATING THE QUALITY OF SYNTHETIC TIME-SERIES DATA The TimeGAN authors use three practical criteria to assess the generated data: 1. Diversity: the distribution of the synthetic samples should roughly match that of the real data 2. Fidelity: the sample series should be indistinguishable from the real data, and 3. Usefulness: the synthetic data should be as useful as their real counterparts for solving a predictive task
- 17. SYNTHETIC DATA FOR FINANCE - RESULTS EVALUATION ASSESSING DIVERSITY: VISUALIZATION USING PCA AND T-SNE
- 18. SYNTHETIC DATA FOR FINANCE - RESULTS EVALUATION ASSESSING FIDELITY: TIME SERIES CLASSIFICATION PERFORMANCE
- 19. SYNTHETIC DATA FOR FINANCE - RESULTS EVALUATION ASSESSING USEFULNESS: TRAIN ON SYNTHETIC, TEST ON REAL
- 20. SYNTHETIC DATA FOR FINANCE - CONCLUSION KEY TAKEAWAYS ‣ Using a small dataset, TimeGAN creates synthetic data that, to some extent, mimic actual stock price series. ‣ While far from conclusive that we can create artificial data that is in fact useful for model training and backtesting at scale, it’s a promising first step. ‣ There are numerous avenues to build on this architecture by refining the configuration and the training process, besides using more (diverse) data.
- 21. 8 Next Week
- 22. 9 Demos, slides and video available on QuAcademy Go to www.qu.academy 9
- 23. 10 Instructions for the Lab: 1. Go to https://academy.qusandbox.com/#/register and register using the code: "QUFALLSCHOOL"
- 24. Thank you! Sri Krishnamurthy, CFA, CAP Founder and CEO QuantUniversity LLC. srikrishnamurthy www.QuantUniversity.com Contact Information, data and drawings embodied in this presentation are strictly a property of QuantUniversity LLC. and shall not be distributed or used in any other publication without the prior written consent of QuantUniversity LLC. 11