2020/11/19 PRIMA2020: Implementation of Real Data for Financial Market Simulation using Clustering, Deep Learning, and Artificial Financial Market
0 likes163 views
The document discusses a new model for financial market simulation that uses actual data, aimed at improving the trustworthiness of simulations following past market instabilities. It focuses on high-frequency trading market-making strategies and compares the performance of a traditional model with a machine learning-based model using data from the Tokyo Stock Exchange. The findings suggest the new model shows improved results, but also highlights the need to address the non-robustness of machine learning approaches.
![Hierarchical Clustering [Uno et al. 18]
• Normalizations for each indices & clustering
• Euclidean distance
• Ward’s method
• 10 clusters
11/19/20 ©M.HIRANO & Izumi Lab. 10
HFT-MM cluster based on indices](https://image.slidesharecdn.com/20201119prima2020implementationofrealdataforfinancialmarketsimulationusingclusteringdeeplearninganda-201120100742/85/2020-11-19-PRIMA2020-Implementation-of-Real-Data-for-Financial-Market-Simulation-using-Clustering-Deep-Learning-and-Artificial-Financial-Market-10-320.jpg)
![Stylized Trader Agents [Chiarella et al. 02]
• Logarithmic return prediction for bid/ask price
𝑟𝑟 =
1
𝑤𝑤𝐹𝐹+𝑤𝑤𝐶𝐶+𝑤𝑤 𝑁𝑁
𝑤𝑤𝐹𝐹 ⋅ 𝐹𝐹 + 𝑤𝑤𝐶𝐶 ⋅ 𝐶𝐶 + 𝑤𝑤𝑁𝑁 ⋅ 𝑁𝑁
• Fundamentals
𝐹𝐹 =
1
mean reversion time
ln
current market price
current fundamental price
• Chartist (trend)
𝐶𝐶 = logarithm averaged return in the past
• Noise 𝑁𝑁 ~ 𝑁𝑁 0, 𝜎𝜎𝑁𝑁
• + margin => decide price
• Every 100 step they make a buy or sell order
11/19/20 ©M.HIRANO & Izumi Lab. 15](https://image.slidesharecdn.com/20201119prima2020implementationofrealdataforfinancialmarketsimulationusingclusteringdeeplearninganda-201120100742/85/2020-11-19-PRIMA2020-Implementation-of-Real-Data-for-Financial-Market-Simulation-using-Clustering-Deep-Learning-and-Artificial-Financial-Market-15-320.jpg)
![Traditional HFT-MM Trader [Avellaneda et al. 02]
• Trader’s price interval:
𝛾𝛾𝑖𝑖�𝜎𝜎𝑖𝑖
2
+
2
𝛾𝛾𝑖𝑖
ln 1 +
𝛾𝛾𝑖𝑖
𝑘𝑘
• Trader’s mid-price
𝑝𝑝𝑡𝑡
∗
− 𝛾𝛾𝑖𝑖�𝜎𝜎𝑖𝑖
2
𝑞𝑞𝑡𝑡
𝑖𝑖
• Note:
𝛾𝛾𝑖𝑖: risk-hedge level
�𝜎𝜎𝑖𝑖: SD in price
𝑘𝑘: a parameter for order arrival
𝑝𝑝𝑡𝑡
∗
: fundamental price
𝑞𝑞𝑡𝑡
𝑖𝑖
:inventory
11/19/20 ©M.HIRANO & Izumi Lab. 16
Price
Sell
Buy
Fundamental Price
Trader’s mid-price
Trader’s price interval](https://image.slidesharecdn.com/20201119prima2020implementationofrealdataforfinancialmarketsimulationusingclusteringdeeplearninganda-201120100742/85/2020-11-19-PRIMA2020-Implementation-of-Real-Data-for-Financial-Market-Simulation-using-Clustering-Deep-Learning-and-Artificial-Financial-Market-16-320.jpg)
2020/11/19 PRIMA2020: Implementation of Real Data for Financial Market Simulation using Clustering, Deep Learning, and Artificial Financial Market
- 1. Implementation of Real Data for Financial Market Simulation using Clustering, Deep Learning, and Artificial Financial Market Masanori HIRANO1, Hiroyasu MATSUSHIMA2, Kiyoshi IZUMI1, and Hiroki SAKAJI1 1 School of Engineering, The University of Tokyo 2 Center for Data Science Education and Research, Shiga University hirano@g.ecc.u-tokyo.ac.jp https://mhirano.jp/
- 2. Motivation • Instability in Financial Markets • 2008 financial crisis • Flush Crush • Price fluctuation by COVID-19 • Regulations are necessary • New regulations like Basel III • Can avoid abovementioned crisis? • Difficulties in Financial markets • nonstationary • Rare phenomena happen frequently • => Simulation is good solution, but not trustable. • Find what’s the matter • Dealing with trustability with actual data ©M.HIRANO & Izumi Lab. DJIA on May 6, 2010 (Flush crush) DJIA in 2020 (COVID-19) 11/19/20 2
- 3. Last year: PRIMA 2019 • We showed the difference between simulation & data • => simulation model can overpass key features. ©M.HIRANO & Izumi Lab. Missing feature in simulation 11/19/20 3
- 4. Our work • Proposed a new model built using actual data • Comparing in Simulation Traditional model Our new model w/ data • Only focus on HFT-MM Specific trader & strategy • Target: Tokyo Stock Exchange • We analyzed a special data provided by JPX 11/19/20 ©M.HIRANO & Izumi Lab. 4 Tokyo Stock Exchange
- 5. What’s the HFT-MM? • High-Frequency-Trader Market-Making strategy • Market-making strategy: • (Basically) order near the best price • Get profit by the spread (1001-999=2) • Do repeatedly • Risk-hedge by high-frequency-trade: • Always have price move risk (Price move >> spread) • Do action faster & hedge risk by setting off their inventory • => These features are easy to recognize in data 11/19/20 ©M.HIRANO & Izumi Lab. 5 Sell Buy
- 6. Data Extraction We need HFT-MM ordering data… ©M.HIRANO & Izumi Lab.11/19/20 6
- 7. Data • “Order-book reproduction data” provided by Japan Exchange Group (JPX) • Containing masked trader information <- Called “Virtual Server (VS)” 11/19/20 ©M.HIRANO & Izumi Lab. 7 Time Ticker Kind Buy/sell VS Price 11:11:50.702813 A Limit Order sell VS1 2570 11:11:50.703600 B Executed buy VS4 Market Order 11:11:50.704001 A Cancel sell VS1 2570 Sample Some columns are not shown such as volume
- 8. Indices for clustering (extracting HFT-MM) • The logarithm of action per ticker ActionsPerTicker = newOrders + changeOrders + (cancelOrders) (numTicker) ActionsPerTickerLOG = ln ActionsPerTicker • Inventory Ratio InventoryRatioABS = 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 𝑛𝑛ticker soldVolume ticker − boughtVolume ticker soldVolume ticker + boughtVolume ticker • Executed order ratio • Cancel order ratio • Market order ration • The logarithm of ticker per VS TickerPerVSLOG = ln (numTicker) (numVS) 11/19/20 ©M.HIRANO & Izumi Lab. 8 Many order Low inventory Many VS usage Low executed ratio High cancel ratio Low market order ratio
- 9. Data outline • Jan. 2015 – mid-Sep. 2015: All 178 business days • All traders: 2654 traders • Only HFT: 181 traders <= based on ActionsPerTicker ≥ 1000 • Hierarchical Clustering for HFT-MM 11/19/20 ©M.HIRANO & Izumi Lab. 9
- 10. Hierarchical Clustering [Uno et al. 18] • Normalizations for each indices & clustering • Euclidean distance • Ward’s method • 10 clusters 11/19/20 ©M.HIRANO & Izumi Lab. 10 HFT-MM cluster based on indices
- 11. Data split • We got ordering data of HFT-MM • 2015/01-07 => model training of HFT-MM • 2015/08 => evaluation of simulation 11/19/20 ©M.HIRANO & Izumi Lab. 11
- 12. Simulation & Models ©M.HIRANO & Izumi Lab.11/19/20 12
- 13. Simulation outline • We used “PlhamJ” as a simulation platform. PlhamJ: Platform for Large-scale and High-frequency Artificial Market (Java version) 11/19/20 ©M.HIRANO & Izumi Lab. 13
- 14. Simulation setting • 1,000 stylized traders + 1 traditional HFT-MM trader vs • 1,000 stylized traders + 1 ML HFT-MM trader (new) • Comparison between behaviors of - • 1 traditional HFT-MM trader in simulation • 1 ML HFT-MM trader (new) in simulation • Real data (out of learning data) ©M.HIRANO & Izumi Lab.11/19/20 14
- 15. Stylized Trader Agents [Chiarella et al. 02] • Logarithmic return prediction for bid/ask price 𝑟𝑟 = 1 𝑤𝑤𝐹𝐹+𝑤𝑤𝐶𝐶+𝑤𝑤 𝑁𝑁 𝑤𝑤𝐹𝐹 ⋅ 𝐹𝐹 + 𝑤𝑤𝐶𝐶 ⋅ 𝐶𝐶 + 𝑤𝑤𝑁𝑁 ⋅ 𝑁𝑁 • Fundamentals 𝐹𝐹 = 1 mean reversion time ln current market price current fundamental price • Chartist (trend) 𝐶𝐶 = logarithm averaged return in the past • Noise 𝑁𝑁 ~ 𝑁𝑁 0, 𝜎𝜎𝑁𝑁 • + margin => decide price • Every 100 step they make a buy or sell order 11/19/20 ©M.HIRANO & Izumi Lab. 15
- 16. Traditional HFT-MM Trader [Avellaneda et al. 02] • Trader’s price interval: 𝛾𝛾𝑖𝑖�𝜎𝜎𝑖𝑖 2 + 2 𝛾𝛾𝑖𝑖 ln 1 + 𝛾𝛾𝑖𝑖 𝑘𝑘 • Trader’s mid-price 𝑝𝑝𝑡𝑡 ∗ − 𝛾𝛾𝑖𝑖�𝜎𝜎𝑖𝑖 2 𝑞𝑞𝑡𝑡 𝑖𝑖 • Note: 𝛾𝛾𝑖𝑖: risk-hedge level �𝜎𝜎𝑖𝑖: SD in price 𝑘𝑘: a parameter for order arrival 𝑝𝑝𝑡𝑡 ∗ : fundamental price 𝑞𝑞𝑡𝑡 𝑖𝑖 :inventory 11/19/20 ©M.HIRANO & Izumi Lab. 16 Price Sell Buy Fundamental Price Trader’s mid-price Trader’s price interval
- 17. HFT-MM Machine Learned Model • Using machine learning for data, we build a model • Model predict the next action of traders ©M.HIRANO & Izumi Lab.11/19/20 17
- 18. Results 11/19/20 ©M.HIRANO & Izumi Lab. 18
- 19. Comparison 11/19/20 ©M.HIRANO & Izumi Lab. 19 Ticks between the best price and ordering of HFT-MM
- 20. Comparison in KL Divergence • Our new ML model outperform traditional model marginally… • Why so big variance? => 11/19/20 ©M.HIRANO & Izumi Lab. 20 Distribution of 𝐷𝐷𝐾𝐾𝐾𝐾 of our new model Q P Mean SD Actual Traditional 0.730009 0.119884 Actual ML 0.648459 0.957854
- 21. Comparison w/ omission • Error case: easy to detect => omit them • => the omission give us strong results 11/19/20 ©M.HIRANO & Izumi Lab. 21 Q P Mean SD Actual Traditional 0.730009 0.119884 Actual ML (w/ omissions) 0.186192 0.085099
- 22. Discussion & Conclusion • Our new model show the strong result w/ omission • Reveal the needs & benefits of real data usage • But, we should deal with non-robustness of ML model 11/19/20 ©M.HIRANO & Izumi Lab. 22 Future work • More robust ML model • Model building with data for all trader