Generative Adversarial Networks for Amplifying and Extending Financial Market Data Michael Wellman Lynn A. Conway Collegiate Professor of Computer Science & Engineering University of Michigan
References • Generating realistic stock market order streams (J. Li, X. Wang, Y. Lin, A. Sinha, and M. P. Wellman). 34 th AAAI Conference on Artificial Intelligence , pages 727–734, Feb 2020. • Market manipulation: An adversarial learning framework for detection and evasion (X. Wang and M. P. Wellman). 29 th International Joint Conference on Artificial Intelligence , pages 4626–4632, July 2020.
Spoofing is the practice of submitting large spurious buy or sell orders with the intent to cancel them before execution to mislead other traders. 3
Price 115.90 - 115.89 - Large spoof sell orders 115.88 - Transacted sell True sell order 115.87 - Profit True buy order 115.86 - Transacted buy 115.85 - 100 200 300 400 500 600 ms 115.84 - Large spoof buy orders 115.83 - 115.82 - 115.81 - Source: Financial Conduct Authority, Animated Example of Mr. Coscia’s Trading 18
Detecting Market Manipulation • The ideal case: adopt supervised learning approaches - Represent an order stream as a variable-length sequence of bidding actions (e.g., price and quantity pairs) Order Streams Manipulation or Detector from Individual Normal Trading Traders 19
Developing Manipulation Signatures for Detection optimization background trading spoofing calibration model strategy market signature surveillance/audit data extractor algorithms machine learning • Given model of spoofing behavior, need calibrated market data source for injection • Approach: Learn to generate realistic financial order streams
Order Book Evolution • Generator outputs next order, conditional on order book state and history
Generative Adversarial Networks (GANs) Interleave training of two deep NNs: Generator Takes noise vector as input, generates sample data item • objective: confuse critic • Critic Takes real or generated sample, classifies as real or not •
Generator NN Architecture Pre-trained CDA network Convolutional layers after a single fully connected layer LSTM layer Noise input Time History x j of length k
Critic NN Architecture Output Convolutional layers after a single fully connected layer LSTM layer History x j of length k Time Real/Gen x i
Experiments and Evaluation • Trained on: 1. simulated financial market 2. thinly traded stock: PN, 20K orders/day 3. thickly traded stock: GOOG, 230K orders/day • Evaluated various statistics comparing real and generated (fake) order streams
Results (Buy Orders)
Results (Bid/Ask Evolution)
Detecting Manipulation: Challenges • Codified manipulation strategies may not be diverse enough; • Adversary may obfuscate actions to evade detection, given a developed classifier. 28
<latexit sha1_base64="Qi6CBnHLXm9RFUR7UBemC2sIbQ=">AB9HicbVDLSgMxFL3js9ZX1aWbYBFclRkp6LKoC5cV7APaoWTStA3NZMbkTqEM/Q43LhRx68e482/MtLPQ1gOBwzn3ck9OEth0HW/nbX1jc2t7cJOcXdv/+CwdHTcNFGiGW+wSEa6HVDpVC8gQIlb8ea0zCQvBWMbzO/NeHaiEg94jTmfkiHSgwEo2glvxtSHDEq07tZz+2Vym7FnYOsEi8nZchR75W+uv2IJSFXyCQ1puO5Mfop1SiY5LNiNzE8pmxMh7xjqaIhN346Dz0j51bpk0Gk7VNI5urvjZSGxkzDwE5mIc2yl4n/eZ0EB9d+KlScIFdscWiQSIRyRogfaE5Qzm1hDItbFbCRlRThranoi3BW/7yKmleVrxqpfpQLdu8joKcApncAEeXEN7qEODWDwBM/wCm/OxHlx3p2Pxeiak+cwB84nz+jI5IE</latexit> An Adversarial Learning Framework to Evade Detection • A case study: modify spoofing to resemble market making . - A market maker provides liquidity by simultaneously submitting buy and sell orders around an estimate of the fundamental value. 101000 101000 MM 100600 100600 100200 100200 3rice 3rice D 0 99800 99800 99400 99400 SP 99000 99000 0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 TiPe TiPe A Manipulation Order Stream (SP) A Market-Making Order Stream (MM) 29
<latexit sha1_base64="xf5t+RLxbFtdzKb5ZJ5kyMdOZmc=">AB/3icbVBNS8NAFNzUr1q/oIXL8EieCqJFOyx4MWDhwq2FZpSNtvXdulmE3ZfxBJ78K948aCIV/+GN/+NmzYHbR1YGbe481OEAu0XW/rcLK6tr6RnGztLW9s7tn7x+0dJQoBk0WiUjdBVSD4BKayFHAXayAhoGAdjC+zPz2PSjNI3mLkxi6IR1KPuCMopF69pEfUhwxKtLrac9HeMBUwXDas8tuxZ3BWSZeTsokR6Nnf/n9iCUhSGSCat3x3Bi7KVXImYBpyU80xJSN6RA6hkoagu6ms/xT59QofWcQKfMkOjP190ZKQ60nYWAms7R60cvE/7xOgoNaN+UyThAkmx8aJMLByMnKcPpcAUMxMYQyxU1Wh42ogxNZSVTgrf45WXSOq941Ur1plqu1/I6iuSYnJAz4pELUidXpEGahJFH8kxeyZv1ZL1Y79bHfLRg5TuH5A+szx/8Zpa0</latexit> <latexit sha1_base64="Yr1pBXtGQ0+Rtn2nqEmST9IduXY=">AB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoMeiBz1WtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq7RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNwI7CQKaRQIbAfjm5nfkKleSwfzSRBP6JDyUPOqLHSw23f65crbtWdg6wSLycVyNHol796g5ilEUrDBNW67mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZIzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwis/4zJDUq2WBSmgpiYzP4mA6QGTGxhDLF7a2EjaizNh0SjYEb/nlVdK6qHq1au2+Vqlf53EU4QRO4Rw8uIQ63EDmsBgCM/wCm+OcF6cd+dj0Vpw8plj+APn8wfDQY12</latexit> <latexit sha1_base64="p3I1zsXs0PIZIi2DkU+U0KLK7TQ=">AB6nicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkoMeiHjxWtB/QhrLZTtqlm03Y3Qgl9Cd48aCIV3+RN/+N2zYHbX0w8Hhvhpl5QSK4Nq7RTW1jc2t4rbpZ3dvf2D8uFRS8epYthksYhVJ6AaBZfYNwI7CQKaRQIbAfjm5nfkKleSwfzSRBP6JDyUPOqLHSw23f7ZcrbtWdg6wSLycVyNHol796g5ilEUrDBNW67mJ8TOqDGcCp6VeqjGhbEyH2LVU0gi1n81PnZIzqwxIGCtb0pC5+nsio5HWkyiwnRE1I73szcT/vG5qwis/4zJDUq2WBSmgpiYzP4mA6QGTGxhDLF7a2EjaizNh0SjYEb/nlVdK6qHq1au2+Vqlf53EU4QRO4Rw8uIQ63EDmsBgCM/wCm+OcF6cd+dj0Vpw8plj+APn8we9K41y</latexit> <latexit sha1_base64="iHEBTNGWZadnMDkzl0OKw5qR8=">AB/3icbVDLSsNAFJ3UV62vqODGTbAIrkoiBbsuHhoJ9QBPCZDJph04ezNwUS8zCX3HjQhG3/oY7/8ZJm4W2Hhg4nHMv98zxEs4kmOa3Vlb39jcqm7Xdnb39g/0w6OejFNBaJfEPBYD0vKWUS7wIDTQSIoDj1O+97kuvD7Uyoki6N7mCXUCfEoYgEjGJTk6id2iGFM9uc9cG+gAZ9qe5q9fNhjmHsUqsktRiY6rf9l+TNKQRkA4lnJomQk4GRbACKd5zU4lTCZ4BEdKhrhkEonm+fPjXOl+EYQC/UiMObq740Mh1LOQk9NFmnlsleI/3nDFIKWk7EoSYFGZHEoSLkBsVGUYfhMUAJ8pgmgqmsBhljgQmoymqBGv5y6ukd9mwmo3mXbPebpV1VNEpOkMXyEJXqI1uUAd1EUGP6Bm9ojftSXvR3rWPxWhFK3eO0R9onz/3tJax</latexit> An Adversarial Learning Framework to Evade Detection • Adapt SP to evade detection while preserving manipulation effects MM • Manipulation intensity the fraction of price deviation; • Transaction risk # transactions / # arrivals D 0 L adv SP SP SP 1 Market Market G 1 Simulator Simulator L reg 32
An Adversarial Learning Framework to Evade Detection • A recursive training procedure 33
Empirical Evaluation • Similarity to market making; • Preservation of manipulation effects. 34
Similarity to Market Making SP SP SP SP 1 SP SP 2 SP SP 3 Ø Quote simultaneously on both sides of the market; Ø Place large orders behind smaller ones. 35
SP1 → SP2 → SP3 → • Closer resemblance to MM • Orders cover a wider range of prices with small quantities • Buy and sell orders are better balanced • Degraded spoofing effect • reduced manipulation intensity • higher transaction risk
Combine Agent-Based Simulation and Adversarial Learning to Detect Market Manipulation Verify the Generate New Manipulation Effects Manipulation Patterns 37
Recap: GANs for Financial Modeling • Train model to generate realistic order streams • An adversarial learning framework for market manipulation - Reason about how a manipulator might mask its behavior - Understand the dynamics of evasion and detection - Generate a diverse set of manipulative patterns to improve detection robustness 38
Recommend
More recommend
